EU law · 24 official languagesArticle-level JSON · typed citation graphMCP server · Claude, Cursor, VS CodeSDKs · Python, TypeScriptFree plan · 500 credits to start
Brussels · CET
Vol. II·Section R — Release Notes
REST API · Dashboard · MCP server

§ Reference — API changes

Changelog

Changes to the LexAPI REST API, the dashboard and the MCP server, newest first; the current API version is 3.1.0.

On this page, now at 3.1.0

User-visible changes to LexAPI — the REST API, the dashboard, and the MCP server. Grouped by month, newest first. For request/response details see the API docs and the OpenAPI spec.

Entries that change the API contract name the contract version they introduced (API x.y.z, the OpenAPI info.version and the X-LexAPI-Version response header). How the numbers work: Versioning & changelog. Deprecations get their own entry headed Deprecated:, with the replacement and the removal date (at least 180 days after it is published) — see Deprecations.

November 2026

LatestNovember 2026

/documents/recent limit; /info on daily-call plans (API 3.1.0)

#
  • Changed: GET /api/v1/documents/recent with a limit that does not start with a number (limit=abc, limit=) or is negative is a refunded 400 INVALID_PARAMS; it answered a charged 200. Every other limit is read as before.
  • Changed: on daily-call plans (accounts from before credit pricing), GET /api/v1/info no longer counts against the daily limit and is answered when the limit is used up; usage.unitsCharged is 0. Credit plans are unchanged.
  • Docs: a document published only moments ago can answer 404 NOT_FOUND: try again in 10 minutes.
November 2026

Every change since 2.18.1, in one release (API 3.0.0)

#

Released 22 November 2026. The API moves from 2.18.1 to 3.0.0, still under /api/v1. Nothing is removed or renamed; credit costs are unchanged. API 2.18.2 to 2.30.0 were not released on their own; they ship together as 3.0.0. What an existing integration or account may notice:

API behaviour

#
  • Document reads (2.21.0) — an unreadable document gave 404 or a text-less 200 → refunded 502/504/500 (batch items: UPSTREAM_ERROR, TIMEOUT, INTERNAL); 404 means it does not exist. Do: retry with backoff, honour Retry-After; keep treating 404 as final.
  • PAYMENT_PAST_DUE (2.27.0, 2.29.2) — after a failed payment, keys were refused at once (403 { "error": "Subscription inactive" }) → they work for up to 7 days, then get 403 PAYMENT_PAST_DUE until the payment succeeds or the subscription ends; accounts already past due at the release stay refused. Do: branch on code; show billingUrl.
  • inForce on /legislation/semantic (2.19.0) — ignored → true keeps only acts known to be in force today; a non-boolean is a 400. Do: send a boolean or omit it.
  • /documents/:celex/at/:date (2.28.0) — some dates return another snapshot; new refunded 404 boundary_unknown; impossible dates and non-EU language codes are a 400. Do: treat boundary_unknown like before_tracking.
  • asOf on /:jurisdiction/legislation/semantic (2.29.0) — ignored → applied; an invalid value is a 400, not charged. Do: send YYYY-MM-DD or omit it.
  • /search, /parse, /documents/recent (2.21.0, 2.28.1, 2.30.0) — a failed search could look like a result → refunded 502; error is an object or a string. Do: retry; accept both shapes.
  • /citations/extract (2.21.0) — success: true with an incomplete count → refunded 502. Do: retry.
  • /resolve (2.21.0) — a failed ECLI lookup was a 404 → refunded 500. Do: retry.
  • asOf on /legislation/semantic (2.19.0, 2.22.0, 2.23.1, 2.25.0) — more hits unresolved (in_force_dates_unknown, not_yet_in_force, new amendment_not_consolidated); some not_in_language hits resolve. Do: accept unknown reasons; use the text only if asOfResolved is true.
  • fetchedAt on document reads (2.20.0) — response time → when the text was fetched. Do: do not use it as the request time.
  • fetchedAt of past versions (2.28.0) — last fetch → when the version was first observed, where known (also observedFrom). Do: do not read it as "last checked".
  • include: ["text"] on /search/semantic (2.30.0) — a hit without text had no snippet → keeps it, with textAvailability. Do: read textAvailability.
  • OAuth scopes (2.28.1) — not required → a token without the route's scope gets 403 TIER_FORBIDDEN; existing connections hold lex; API keys have none. Do: nothing.
  • JSON without its content type (2.24.0) — the 400/422 adds a hint, maybe a message field. Do: send Content-Type: application/json; do not match error bodies exactly.

Billing and credits

#
  • Past due (2.28.1) — set at every failed payment → only when the subscription is past due or unpaid; monthly credits do not renew meanwhile. Do: pay the open invoice.
  • Top-up credits (2.26.1, 2.29.2) — untouched by resets → spent after the monthly credits: each credit reset (monthly, at a plan change, or by LexAPI) deducts those used; the rest stay while the paid plan is active. A refunded purchase removes its credits. Balances at the release are kept. Do: nothing.
  • Automatic top-ups: charging (2.26.1) — not charged → buys a pack when the balance runs low, at most three times per monthly credit cycle, once you confirm the terms and the dashboard shows charging as active. Do: confirm, or switch it off.
  • Automatic top-ups: switched off (2.29.2) — now when a charge is declined, needs your bank's confirmation or cannot be made with the card on file, when a paid subscription ends, or when the account moves to FREE or a plan arranged with LexAPI. Do: switch it on again if wanted.
  • Top-up checkout (2.28.1) — now needs an active subscription, else 409 SUBSCRIPTION_NOT_ACTIVE. Do: pay the open invoice first.

SDKs and MCP server. The releases that match API 3.0.0 are @lexapi/client 0.5.0, lexapi-client 0.5.0 and @lexapi/mcp 0.6.0. Earlier releases (0.4.0, 0.4.0, 0.5.1) keep working without an upgrade; with them a past-due account gets an error with code PAYMENT_PAST_DUE (the SDKs raise LexAPIError). See each package's changelog before upgrading: TypeScript, Python, MCP.

New

#
  • POST /api/v1/legislation/as-of (2.26.0); in-force label and filter (2.19.0); provenance (2.20.0); observedFrom (2.28.0); national asOf labels (2.29.0); /info result caps (2.24.0); dashboard invoices.

Fixed

#
  • Semantic search for FREE accounts marked past due (2.29.3); two 500s, and a credit reset is applied only once (2.30.0); an entry into force dated before the act (2.22.0).

Versioning policy. From 3.0.0, a behaviour change an existing integration may notice ships in a MAJOR version under /api/v1, without a notice period, and is listed in that version's entry. Nothing is removed or renamed without a deprecation of at least 180 days: Versioning & changelog.

Some document reads may be slower for about a day after the release; they are not charged more.

November 2026

Case-law text availability, /documents/recent types, two 500s fixed, versioning policy (API 2.30.0)

#
  • include: ["text"] on POST /api/v1/search/semantic: every hit carries textAvailability; a hit without text keeps its snippet (from the machine-written summary, not court text). Same cost.
  • /documents/recent names an unknown documentType in ignoredDocumentTypes and X-Ignored-DocumentTypes, as /search does; it still filters nothing and is charged as usual.
  • Fixed: two 500s — /search with some documentType values (now answered) and /documents/recent with an invalid days (now a refunded 400); a credit reset is applied only once. Versioning policy: see API 3.0.0.
November 2026

Semantic search on a FREE account marked past due (API 2.29.3)

#
  • Fixed: some FREE accounts got 403 TIER_FORBIDDEN on semantic search.
November 2026

Billing: payments that failed before this release; top-up refunds and automatic top-ups (API 2.29.2)

#
  • The up-to-7-day period applies to payments that fail from API 3.0.0 on; an account already past due stays refused until the payment succeeds or the subscription ends.
  • A refunded top-up purchase removes its credits (a partial refund, its share; the balance never goes below zero).
  • Automatic top-ups switch off when a paid subscription ends.
November 2026

Docs: asOf wording (API 2.29.1)

#
  • Docs: with asOf, POST /api/v1/legislation/semantic returns for each hit the version listed for the date, as POST /api/v1/legislation/as-of does; the docs said "the version in force".
November 2026

National semantic legislation search: asOf labels (API 2.29.0)

#
  • asOf on POST /api/v1/:jurisdiction/legislation/semantic (1800-01-01 to today): a hit is asOfResolved: true only when the publisher's dates show today's wording applied on that date.
  • The search covers today's texts only: for a provision amended since, LexAPI cannot return the earlier wording. An unresolved hit keeps its text. That text is today's wording; it is not established as the text for the date.
  • LexAPI holds such dates for Austria and Spain; for the other countries every hit is asof_not_available. Same cost. Reasons and checkedAt: API docs.
November 2026

Error responses: the credits figures (API 2.28.1)

#
  • credits on an error response may show the charge before an automatic refund; a later response shows the refunded balance. Details: Errors.
November 2026

GET /documents/:celex/at/:date: correct snapshot for the date; clearer answer (API 2.28.0)

#
  • Fixed: /at/:date returns the snapshot LexAPI held at the end of that day (UTC); some dates now return a different one.
  • Added: observedFrom on every version; on /at/:date also document.observedUntil (when the document's next version arrived), dateInFuture and, for EU acts, inForceLookup (a pointer to /legislation/as-of); reason: "before_tracking" on the 404 for a date before tracking.
  • Changed: a date LexAPI cannot place is a refunded 404 boundary_unknown; a date that does not exist (2026-02-31) or a non-EU language is a refunded 400. This is LexAPI's fetch history, not the law in force. 1 credit. A 404 costs 1 credit, except boundary_unknown, which is refunded. API docs.
November 2026

Errors: PAYMENT_PAST_DUE documented; unknown endpoints marked (API 2.27.0)

#
  • PAYMENT_PAST_DUE (403) is now a documented error code, with billingUrl and, when known, accessEndedAt: the answer after the up-to-7-day period that follows a failed payment, until it succeeds.
  • Unknown endpoints: the 404 for a path or method the API does not have carries error.details.reason: "unknown_endpoint"; its status and code are unchanged.
  • Docs correction: truncated exports. Page a truncated /export with consecutive dateFrom/dateTo windows, not changedSince; run one that ends truncatedBy: "recovered" again.
November 2026

Billing: failed payments, invoices in the dashboard

#
  • Up to 7 days to fix a failed payment. When a subscription payment fails, your API keys keep working for up to 7 days. The dashboard shows the exact time.
  • Invoices in the dashboard, each with its PDF; Manage billing for every account with billing history.
November 2026

Top-up credits and automatic top-ups (API 2.26.1)

#
  • Top-up credits are spent after the monthly credits; each monthly reset deducts those used, and the rest stay while your paid plan is active.
  • Automatic top-ups (paid self-serve plans): the dashboard shows the terms (at most three charges per monthly credit cycle), asks you to confirm them, and shows when charging is active.
November 2026

POST /legislation/as-of: the version of an EU act for a date, and one provision of it (API 2.26.0)

#
  • New endpoint: POST /api/v1/legislation/as-of. Send an EU act's CELEX number and a date; asOfVersion names the version listed for that date (the newest consolidated version, or the act as published). Add provision (art_17) for one article, annex or recital. Unresolved: asOfResolved: false with an asOfReason, no text.
  • It does not say whether a provision applied on that date. Consolidated versions are documentation without legal force; only the act as published in the Official Journal is authentic.
  • Costs 1 credit without a text fetch; with one, 4 (5 live), or 1 if it fails or finds no text; up to 5 are reserved meanwhile. API docs.
November 2026

asOf: a hit stays unresolved when an amendment is not consolidated (API 2.25.0)

#
  • New asOfReason: amendment_not_consolidated on POST /api/v1/legislation/semantic with asOf: an amending act may have applied on that date, and no consolidated version listed for the date is known to include it. The hit keeps its indexed text, which is not established as the text for the date.
  • Regulations, directives and decisions. Same cost; clients that ignore unknown asOfReason values need no change.
November 2026

/info lists every endpoint and the semantic result cap; clearer error for JSON sent without its content type (API 2.24.0)

#
  • /info: subscription.maxSemanticResults (case law) and subscription.maxSemanticLegislationResults (legislation, at most 50) are the most results one semantic search returns on your plan; endpoints also lists /citations/transposition/:celexNumber and /citations/counterparts/:id (listed while a country is open to your account).
  • X-LexAPI-Client: an optional header for integrations to name themselves; it never affects authentication, credits or limits. JSON sent without Content-Type: application/json gets a hint in its 400 or 422.
  • Docs: the official SDKs (0.4.0 or later) cover the national routes.
November 2026

asOf: consolidated versions that only add corrigenda in other languages (API 2.23.1)

#
  • Fixed: with asOf, some hits answered not_in_language although the act has a text in force in that language; they now resolve (for example English hits of 32018R1725). Same cost.
November 2026

Provenance: fetch time on national documents, EU licence (API 2.23.0)

#
  • provenance.retrievedAt on national documents is when LexAPI last fetched the text from its source; null until it is next fetched.
  • provenance.licence on EU documents is EUR-Lex legal notice — the notice that sets the reuse conditions for what EUR-Lex publishes. See also Terms §7.
November 2026

In-force label: corrigenda, entry into force, end dates (API 2.22.0)

#
  • Corrigenda carry the in-force label of the act they correct, and inForce: true keeps them when that act is in force.
  • Fixed: an entry into force recorded before the act's own date is ignored (the Digital Services Act now reads not_yet_in_force for an asOf from 19 October to 15 November 2022). endOfValidity is left out when several end dates are recorded.
November 2026

A 404 only when the document does not exist (API 2.21.0)

#
  • A 404 now means the document does not exist. When a document cannot be read, /documents/metadata, /documentContent, batch items and /documents/url answer 502 UPSTREAM_ERROR, 504 TIMEOUT or 500 INTERNAL, refunded in full. A 404 still costs 1 credit.
  • Same rule elsewhere: /citations/extract, /search, /parse and /documents/recent answer a refunded 502 instead of an incomplete result; a failed ECLI lookup on /resolve is a refunded 500.
  • The 502s of document reads and /citations/extract carry Retry-After when the wait is known. Details: Errors.
November 2026

Provenance on every document (API 2.20.0)

#
  • New object provenance on documents and semantic hits: publisher, sourceUrl, sourceIdentifier, retrievedAt, contentHash, versionIdentifier, language, licence — always all eight keys, null where a value is not known. National documents and hits carry it too, with the source and licence that GET /api/v1/national/coverage lists.
  • New object fieldProvenance: which fields are the publisher's text (official_source_text) and which were derived (machine_extracted_structure, machine_generated_interpretation, editorial_enrichment); LexAPI serves no machine translations. Additive; same cost. Details: Provenance.
November 2026

In-force label on EU legislation (API 2.19.0)

#
  • New fields inForce, notInForceReason, endOfValidity, entryIntoForce on POST /api/v1/legislation/semantic hits and on document reads: whether the act is in force today (UTC), as recorded by the Publications Office; left out when not known. They do not cover amendments or dates of application. No extra credits.
  • New filter inForce: true: only hits whose act is known to be in force today; the others are counted in inForceFilter, so a page can be shorter than limit. Same cost.
November 2026

Semantic legislation search: what you get without asOf (API 2.18.2)

#
  • Docs: without asOf, POST /api/v1/legislation/semantic returns the indexed texts as they are (the act as published, or its latest consolidated version where indexed), not "the latest version". For the text in force today, pass today's date (UTC) as asOf and use the hits with asOfResolved: true.
November 2026

Spec clarifications (API 2.18.1)

#
  • Wording only, no behaviour change: the API docs and the OpenAPI descriptions are shorter; every limit, credit cost, header, parameter and error code is unchanged.
November 2026

Connect apps with your LexAPI account (OAuth) (API 2.18.0)

#
  • "Log in with LexAPI" for MCP clients. LexAPI is now an OAuth 2.1 authorization server, so the hosted MCP server (https://mcp.lex-api.com/v1) can be connected to Claude and other MCP clients by signing in and approving, instead of pasting an API key. Every plan can connect, Free included. How to connect, what you approve and what it costs: Connect with OAuth.
  • Dashboard → Connected apps lists each connected app with its last use and its requests and credits for the last 30 days, and disconnects it immediately. Connections don't count toward your plan's API-key limit.
  • Account security. We email you whenever a new app is connected, and resetting your password disconnects every connected app (API keys keep working). We also email you when a connected app's sign-in looks unusual; if you don't recognise it, disconnect the app. Desktop apps that return through an app link (such as Cursor's cursor://…) are supported and always ask for approval.
  • REST API. API keys work exactly as before, and the OAuth tokens are usable only through the hosted MCP server. One visible change: a request to /api/v1 whose Authorization: Bearer value is shaped like a JWT (and carries no x-api-key) and doesn't authenticate now gets 401 with "error": "Invalid access token" and a WWW-Authenticate: Bearer realm="lexapi", error="invalid_token" header, instead of API key required; an API key sent as Bearer lex_… gets the same answer as before. The authorization server's endpoints (/.well-known/oauth-authorization-server, /.well-known/jwks.json, /oauth/*) are new and sit outside /api/v1.
November 2026

National coverage: citation figures no longer missing after a restart (API 2.17.2)

#
  • GET /api/v1/national/coverage: citationGraph could briefly read null; such an answer is now Cache-Control: no-store.
November 2026

Spec clarifications (API 2.17.1)

#
  • Wording only, no behaviour change: GET /api/v1/citations/counterparts/:id follows the EU acts a national document is linked to by a transposition or implementation edge — one its national source records, or a measure notified to the Commission — rather than "the EU acts it transposes or implements"; a notification is not a finding that the measure transposes the act. The via and truncated descriptions say the same. The example X-LexAPI-* headers show the current version.
November 2026

/resolve: ELIs looked up in Cellar, qualified CELEX kept (API 2.17.0)

#
  • POST /api/v1/resolve looks an ELI up in Cellar: http://data.europa.eu/eli/dec/2019/166/oj now answers 32019D0003(01), not the non-existent 32019D0166. Corrigendum ELIs (…/corrigendum/2018-05-23/oj) resolve to the corrigendum (32016R0679R(02)), consolidated-version ELIs (…/2016-05-04) to that version (02016R0679-20160504), an ELI written with the parent type (eli/reg/2019/947/oj for eli/reg_impl/2019/947/oj) to the same act; language or format segments are optional. An ELI Cellar has no work for returns 404.
  • A bare CELEX that names a distinct work is kept as given: 32019D0003(01) no longer resolves to 32019D0003 (a different act), and 32016R0679R(02) and 02016R0679-20160504 no longer return 404. Every CELEX /resolve returns resolves to itself.
  • New response field verified: true when the mapping was confirmed, false when the CELEX was read off the identifier. New resolvedVia value eli-cellar; eli now means an unconfirmed ELI mapping. An ELI that cannot be checked and has no pattern to fall back on returns 502 UPSTREAM_ERROR (refunded). Still 1 credit.
November 2026

National semantic search: incomplete searches are labelled (API 2.16.0)

#
  • POST /api/v1/:jurisdiction/search/semantic and /:jurisdiction/legislation/semantic label a search the national index could not finish. With nothing to return, the answer is 503 NATIONAL_SEMANTIC_UNAVAILABLE with partialReason: "SEARCH_INCOMPLETE" (with Retry-After when a retry may help) or the new 422 NATIONAL_SEMANTIC_FILTER_UNREACHABLE (rephrase the query or change the filters), refunded in full with the national call given back. With some hits left, it is a 200 with partial: true, partialReason: "SEARCH_INCOMPLETE" and a hint, charged as usual. A complete answer never carries partial.
November 2026

Spec clarifications (API 2.15.1)

#
  • Wording only, no behaviour change: for an agent-audited method the re-check is a separate blind review by a person working with a second AI agent; the country pages say "re-checked in a separate blind review" instead of "re-checked by hand".
November 2026

Citation graph: who labelled the checked links (API 2.15.0)

#
  • GET /api/v1/national/coverage: citationGraph.precision gains review — human (people labelled the links) or agent-audited (an AI model labelled them and a separate blind review re-checked a sample); null without labels — and, for agent-audited only, audit: the re-check's size, done, falseAccepts, disagreements and agentReviewed (links that rest on the model's verdict alone). A method whose re-check finds a wrongly accepted link is reported as human, on its human labels alone.
November 2026

Citation graph: small methods checked in full (API 2.14.0)

#
  • GET /api/v1/national/coverage: citationGraph.precision gains census: true when every link of the method was checked, so errorUpperBound is the exact error rate; false for a sample; null without labels.
November 2026

Spec clarifications (API 2.13.1)

#
  • Wording only, no behaviour change: countsSource is counted whenever a citationType, provenance or resolution filter is set, not only citationType (and, with national documents in scope, while withdrawn national documents are filtered out; jurisdictions alone keeps stats); the older error shape now lists the fields the citation routes already send — param and allowed on a 400 INVALID_PARAMS for an unknown filter value, and partial, partialReason and timeoutMs on the /citations/counterparts 503 TIMEOUT; the example X-LexAPI-* headers show the current version.
November 2026

Citation graph: rates that cannot be measured (API 2.13.0)

#
  • GET /api/v1/national/coverage: citationGraph.methods gains unmeasured: the rates a country's sources do not let LexAPI measure, each with a reason code (not-collected, source-empty); those rates are null, with no floor. The Netherlands declares decisionsToActs and Slovenia actsToEu as not-collected.
  • actEuRate is null, with no floor, for a country whose only EU method is text-pattern: its acts carry no recorded EU references to measure, and textPatternRate measures its EU references (the Netherlands).
November 2026

Citation graph: published floors and the notified-measure match rate (API 2.12.0)

#
  • GET /api/v1/national/coverage: citationGraph gains floors: the lowest value each rate (decisionsToActs, actsToEu, textPattern, nimResolved) is expected to stay at; null until set.
  • textPatternRate on the same block: the share of the EU act numbers found in running text (with their EU marker) that produced an edge, from the last full rebuild (measuredAt); null before one ran — the rate the textPattern floor applies to.
  • nimResolvedRate and nimMeasuredAt on the same block: the share of the measures the member state notified to the Commission, among those in scope for LexAPI's matching, that were matched to a document LexAPI holds — from the last matching run that went through all of the country's measures; null without one.
  • precision.provenance (source or resolver): precision always describes a method of the country's own citation extraction, never the matching of notified measures.
November 2026

Citation graph across borders: transposition, counterparts, provenance (API 2.11.0)

#
  • GET /api/v1/citations/transposition/:celexNumber (5 credits, every plan): the measures the member states notified to the Commission for an EU act (CELEX sector 3), grouped by country, from the Publications Office's Cellar records of the notifications — all 27 member states and the United Kingdom, with one measure shape for every country. Every answer carries the note: "Measures are those the member state notified to the Commission for this act. A notification is not a finding that the measure transposes the act, or transposes it completely." On BUSINESS, SCALE and ENTERPRISE, for live jurisdictions, it also says which measures LexAPI holds (id, resolved, resolution, resolutionBasis, counts.resolved, source.lanes) and adds the documents whose national source records that they transpose or implement the act (provenance: "source"); resolvedOnly=true keeps only the held ones. Below BUSINESS, for a jurisdiction that is not live, and while the national-source terms are not accepted (termsRequired: true, not a 403) those fields are null. Measures Cellar no longer lists are left out.
  • GET /api/v1/citations/counterparts/:id (20 credits, BUSINESS and above): a national document's counterparts in the other live jurisdictions, through the EU acts it transposes or implements — ranked 1 (the national source records the relation), 2 (a notified measure LexAPI holds), 3 (a notified Cellar record LexAPI does not hold). An EU act as :id lists its own measures the same way. A lookup that runs out of time is a 503 TIMEOUT (partial: true, partialReason: "TIMEOUT"), refunded.
  • Provenance and resolution on every edge. citations[] on /cited-by and /cites, and the new edgeDetails[] on /network neighbours, carry provenance (notice, cellar, notice+cellar, cellar-nim, source, resolver) and resolution (id, key, abbreviation, text-pattern; null on EU edges). Every node carries id, jurisdiction and documentKind; celexNumber is unchanged.
  • New filters on /cited-by, /cites, /network, /path and /related: jurisdictions (eu,at,de — a national code implies includeNational=true, same plan gate), provenance and resolution (comma-separated), and citationType now takes a comma-separated list too. The new type transposition is in the citationType enum. An unknown value is a 400 INVALID_PARAMS naming the parameter and its allowed values (refunded).
  • unresolved=true on /cites and /network for a national document id adds unresolved — the references the document records that LexAPI could not turn into an edge (for example a code LexAPI does not hold), at most 500, with unresolvedTruncated. A national id that is not visible to you (a jurisdiction not live for you, a withdrawn document) never appears in it. No extra credits.
  • /citations/stats with includeNational=true adds byJurisdiction, byProvenance and transposition (measures, resolved, byCountry — counted with the /citations/transposition rule).
  • GET /api/v1/national/coverage gains citationGraph per live country: the extraction methods, edges, decisionResolutionRate, actEuRate, transposition (measures, resolved), precision (the labelled-sample error bound of the resolution method) and measuredAt — measured values only, null where nothing was measured.
November 2026

Spec clarifications (API 2.10.1)

#
  • Wording only, no behaviour change: the domain parameter's description says what EUR-Lex's own NATIONAL_LAW collection is (member-state measures transposing EU directives and member-state case law, as EUR-Lex indexes it); the national document-id examples name real ids of the launch countries; the national allowance headers say UTC calendar month.
November 2026

Citation graph: where the totals come from (API 2.10.0)

#
  • countsSource on GET /citations/cited-by, /cites and (in paging) /network: stats when totalCitations / totalEdges (and the network's citesTotal / citedByTotal / citesEdgeTotal / citedByEdgeTotal) were read from maintained per-document counts, or counted when they were counted for the request (always with a citationType filter). Both are exact for the request; the totals keep their meaning and stay integers. null on a partial /network response, like the totals.
  • The undocumented edge type cites no longer exists; those edges are reference, the documented value. No documented response changes.
November 2026

National law: texts withheld by the source (API 2.9.1)

#
  • A national document whose source withholds its text now answers content: null with textAvailability: { available: false, reason: "no_machine_readable_rendition" }, instead of serving the source's notice as the text. First case: Lithuanian register acts whose only "text" was "Teisės akto tekstas neskelbiamas …". A text the register publishes later is served. No new field or reason value; the spec's national No text description says so (a clarification, API 2.9.1).
November 2026

National law: where a document's text was read from (API 2.9.0)

#
  • content.textSource and content.ocr on GET /api/v1/{jurisdiction}/documents/{id}, two more keys of the one content shape — always present. textSource says where the text was read from, for countries whose source publishes some documents only as PDF: html (the source's web rendition), pdf-text (the text layer of the source's PDF) or ocr (machine-read from a scanned page); null for countries that do not label their text — none of the countries live today sets it yet. content.ocr is null unless textSource is ocr; then it names the engine, its version and recognition model, the recognition confidence, the pages read and a notice: OCR text may contain recognition errors and is never the official text. A scan recognised below the country's confidence floor is not served as text (content: null, textAvailability.reason no_machine_readable_rendition).
November 2026

National law: one document shape for every country (API 2.8.0)

#
  • Every national response has the same fields, types and formats in every jurisdiction, so one integration works for all of them. Every documented field is always present: a field the source does not provide is null (or []), no longer left out — ecli, eli, dateOfDocument, validTo, inForce and textBasis included.
  • documentKind (act | decision) on the listing, the document and every semantic hit's document: one vocabulary for legislation and case law. documentType stays the source's own label.
  • decision on GET /api/v1/{jurisdiction}/documents/{id}: a decision's court, chamber, caseNumbers and citedNorms ({ text, documentId }), read from whichever fields the source uses (Austria's gericht, Germany's aktenzeichen, Slovenia's case number inside its ordinal…); null for acts. Served also when a decision has no text.
  • content has one shape. fullText, articles, textLanguages, textLanguage, embeddedForeignText, repeal, source and details — always all of them. Every article is { number, title, section, text, inForceFrom, inForceTo, details }; a unit whose text the source does not publish has text: null. What a source adds beyond these fields (its own names for chambers, summaries, unit ids, cited-norm lists) moved under content.details and articles[].details, unchanged.
  • National semantic search: every hit carries document, the matched document as the listing returns it. Legislation hits always carry law_title, article_ref, section, language, in_force, valid_from, valid_to and law_type, null when the index has nothing for the unit.
November 2026

National law: the language of a document's text (API 2.7.0)

#
  • content.textLanguages, content.textLanguage, content.embeddedForeignText on GET /api/v1/{jurisdiction}/documents/{id}, now described in the spec. Dutch treaties (NL-BWBV…) list the languages BWB publishes their text in (["en"] for a treaty published only in its authentic English text, ["en", "nl"] with a Dutch translation). textLanguage is set when the whole text is in one language other than the document's language; embeddedForeignText is true when the text includes text in another language.
  • A court decision whose source publishes no text (rechtspraak.nl: metadata only, or a judgment declared void; RIS: a decision page whose text sections are all empty) is a 200 with content: null and textAvailability.available: false, no longer an empty text.
November 2026

National law: withdrawn documents, a cursor listing (API 2.6.0)

#
  • Withdrawn documents are a 410. A document its source withdrew (for example a court decision taken down to fix its pseudonymisation) is served nowhere. GET /api/v1/{jurisdiction}/documents/{id} answers 410 DOCUMENT_WITHDRAWN with only the id, the jurisdiction and withdrawnAt — no title, no text — and the read is refunded in full. The listing and national semantic search leave it out. On the /citations routes a withdrawn national id is a 410 before anything is charged, and edges to it are left out of includeNational=true results. Delete any copy you hold.
  • The listing pages by cursor. GET /api/v1/{jurisdiction}/documents is ordered newest first by the document's own date (undated last) and pages with cursor / nextCursor (and hasMore) to any depth. page serves the first 1,000 results only. total is counted up to 10,000, with totalExact saying whether it is exact. q takes 3–200 characters. A malformed query is a 400 before anything is charged or counted; a page that takes too long is a refunded 504 TIMEOUT. The cost is unchanged: 3 credits per page.
  • New error code DOCUMENT_WITHDRAWN (410), in both error-code enums of the spec.
  • /citations: a national id of a country that is not live answers its 404 with code: "NOT_FOUND" too (the error string is unchanged).
  • National legislation semantic search validates hyde like the case-law route: anything but a boolean is a 400, and hyde: true is answered with "hyde": false at 5 credits.
November 2026

Launch: national law, country by country — 3 November 2026 (API 2.5.0)

#

This is the national-law launch entry. LexAPI now serves national legislation and case law alongside EU law, under the same API key. Countries open one at a time: which are live, and what each covers, is at lex-api.com/national-law and GET /api/v1/national/coverage.

  • New endpoints. GET /api/v1/{jurisdiction}/documents (list, with q, type, includeRepealed, page, limit), GET /api/v1/{jurisdiction}/documents/{id} (one document with its text, textAvailability and attribution), POST /api/v1/{jurisdiction}/search/semantic (case law) and POST /api/v1/{jurisdiction}/legislation/semantic (legislation). {jurisdiction} is the lower-case country code. A country that is not live answers with the same 404 as an unknown path.
  • Coverage endpoint. GET /api/v1/national/coverage — public, no API key, no credits — lists each jurisdiction with live, launchedAt, legislation and case law separately (scope, document counts, start of coverage, courts), whether semantic search is available, the sources with their licence and required attribution, and lastSyncAt.
  • Every credit plan. FREE, STARTER and PROFESSIONAL get national law within a monthly allowance of national calls — FREE 30, STARTER 500, PROFESSIONAL 2,000 per UTC month — reported in X-National-Limit, X-National-Remaining and X-National-Reset. A used-up allowance is a 402 with the new code NATIONAL_ALLOWANCE_EXHAUSTED, before any credit is charged. A failed call gives its national call back by the credit-refund rule (every 5xx, every 4xx except 404). BUSINESS, SCALE and ENTERPRISE have no separate limit.
  • Costs. A listing page 3 credits; a document read charged 5 and settled at 1 (always our own corpus); national semantic search 5, flat — national indexes have no HyDE, so hyde: true costs 5 and answers hyde: false.
  • Citation graph. includeNational=true on /citations/cited-by, /cites, /network, /path, /related and /stats adds national documents of live jurisdictions, where the national source records the links. BUSINESS, SCALE and ENTERPRISE only (403 TIER_FORBIDDEN below); the allowance does not cover the graph.
  • New error codes. NATIONAL_ALLOWANCE_EXHAUSTED (402), NATIONAL_SEMANTIC_NOT_ENABLED (409: semantic search is not available for that country; refunded) and NATIONAL_SEMANTIC_UNAVAILABLE (503: the country's index is temporarily unavailable; refunded). Both error-code enums in the spec include them.
  • /info lists the national endpoints, with the jurisdictions open to you, once at least one country is live.
  • Licences and permitted use. Each national source has its own licence; every national response carries the attribution to keep when you republish. Consolidated national texts are not the authentic ones. You must not re-identify anyone a court has anonymised or pseudonymised, and analytics on French judges and court clerks (art. 33 of loi n° 2019-222) are prohibited — see Terms §7 and National law in the docs.

October 2026

October 2026

Free plan: 500 credits to start, then 150 a month

#
  • From 30 October 2026, new accounts start with 500 credits for their first 30-day cycle. After that, the Free plan includes 150 credits every 30 days. Unused credits do not carry over, as before.
  • Existing free accounts keep 500 credits until the end of their next credit period after 30 October, then move to 150. The dashboard shows when the current period ends.
  • An account that moves to Free from a paid or trial plan gets 150 a month, with no starting grant. Paid plans, top-ups and the 15 requests a minute on Free are unchanged.
October 2026

MCP server 0.3.0: 26 tools and a hosted endpoint

#
  • @lexapi/mcp 0.3.0 is on npm and in the MCP Registry as io.github.Lex-API/lexapi-mcp. It grows from 10 tools to 26: document batches, observed versions, resolve, bulk export (BUSINESS and above), citation paths, related documents and graph stats, citation extraction, webhook management, and your plan and credit balance. Each tool is billed like the REST endpoint it wraps.
  • Hosted server. No local install needed: POST https://mcp.lex-api.com/v1 (Streamable HTTP), with your API key as Authorization: Bearer or x-api-key. It serves all 26 tools.
  • Read-only mode. LEXAPI_MCP_READ_ONLY=true hides citation extraction and the four webhook writes, leaving 21 tools.

Setup and the full tool list: MCP server.

October 2026

Contract fixes from the SDK and MCP audit (API 2.4.0)

#

Billing and timeouts

#
  • Document reads get a 90-second deadline. /documentContent, /documentContent/batch, /documents/url and /documents/metadata answer 504 TIMEOUT after 90 seconds, refunded in full (the last two answered a 500 before).
  • Failed batch items are refunded, each like a single /documentContent call that failed the same way: NOT_FOUND costs 1 credit, a timeout or upstream error is refunded. A national-law id is reported as UNSUPPORTED_JURISDICTION and not charged.
  • Batch errors[] entries carry a typed code (NOT_FOUND, UNSUPPORTED_JURISDICTION, TIMEOUT, UPSTREAM_ERROR, INTERNAL), plus celexNumber and message. celex and error stay as they were.
  • The batch headers add up. A new X-Batch-Failed header counts the errors[] entries, and X-Batch-Live-Fetches no longer counts failed items. Corpus-Hits + Cellar-Served + OJ-Recovered + Live-Fetches + Failed now equals processed.
  • A timed-out citation network is refunded in full.
  • Documented: what is not refunded. A 5xx LexAPI did not send, and a dropped connection, are not refunded. Every 5xx LexAPI sends is refunded in full.

Errors

#
  • Every 429 carries a stable code. RATE_LIMITED is the per-minute limit, with Retry-After / retryAfter as before. The new QUOTA_EXHAUSTED is a legacy daily-call plan's used-up daily quota or monthly semantic quota. It adds resetsAt (omitted when the reset time is not known to be in the future) and, as before, carries no Retry-After — the official SDKs and the MCP server recognise a quota by that absence and don't retry it. A monthly semantic quota whose billing period on record has already ended has no resetsAt at all: key off the top-level code, as the current SDKs do. Older SDK releases that look only for usage.resetsAt treat that one case as a rate limit and retry it.
  • Legacy-shape errors carry code. The routes whose error is still a string (429s, the semantic endpoints, /search validation and failures, /documents/recent failures, the batch's body validation, the citation routes' failures) now also send success: false and the stable code as a top-level code. Nothing was removed. The docs list exactly which routes use which shape.
  • Semantic queries of 1–2 (or over 1000) characters are a 400, not a 502, and so is a filters.date_from / date_to on /search/semantic that cannot be read.
  • /citations/path: maxDepth of 0, a negative or a non-integer is a 400. It used to be read as 3, and negatives went straight through. Values above 8 are still clamped to 8; an empty value still means 3, and 3.0 counts as 3.

Responses

#
  • usage.unit is sent: credits on credit-pool plans, calls on legacy plans. The docs described it, but it was never there.
  • include: ["metadata"] keeps documentTypeCode, and paragraphs is documented as a valid include token (it always worked).
  • /documents/metadata cleans placeholder authors on corpus reads too. "Provisional data" is dropped, as on /documentContent.
  • GET /webhooks/:id recentDeliveries has the documented delivery shape.
  • Citation paging is deterministic, and counts what it pages. Edges are ordered newest first with a unique tiebreaker, so offset pages no longer repeat or skip rows. A new totalEdges on /cited-by and /cites (and paging.citesEdgeTotal / citedByEdgeTotal on /network) counts the rows that offset pages over. totalCitations counts distinct document pairs, which can be fewer.
  • /info lists every endpoint family: semantic search, versions and export were missing. The author list now includes civil-service-tribunal.
  • /documents/recent says when it narrows a filter. A repeated documentType or subdomain has only ever applied one value, the first valid one. It now says so in an X-Warning header (see the deprecation below).

Spec and docs corrections (no behaviour change)

#
  • The citation endpoints' CELEX path parameter accepts corrigenda (32013R0952R(04)) and consolidated texts (02016R0679-20180524), as the routes always have.
  • SubscriptionTier includes SCALE. Author includes civil-service-tribunal. DocumentByUrlRequest documents bypassCorpus.
  • DELETE /webhooks/{id} is documented as the 200 { success, message } it returns, not 204.
  • 402 CREDITS_EXHAUSTED is listed on every metered route, not only /legislation/semantic.
  • Semantic filters lists the keys the semantic service applies. Case law: date_from, date_to, court, case_number, ecli, language. Legislation reads only language. The case-law example used year_from, which was never applied.
  • The semantic tier caps include FREE (10). parties on semantic case-law hits is documented as the array of strings it is. Also documented: summary_language, the legislation hits' id, snippet, chapter, section and exact_match_law_id (always false on this endpoint), and matchedVersion.id.
  • Also documented: hasResume / hasSummary and titleSource on citation groups and nodes, the language parameter on citation reads, and languageFilter.partialMetadataKept. The headers X-Language-Filter, X-Language-Kept, X-Language-Partial-Kept, X-Cache-Age, X-PostFiltered-Kept (author:3/10;date:2/3) and every X-Source value (corpus, cellar, cellar-oj, live) are documented too.
  • /search: a list of subdomains is matched as any of, except on answers without an X-Search-Source header, which apply only the first.
  • /documents/recent limits, now stated: author=general-court / civil-service-tribunal is not filtered there (you get all case law of the window), and implementing, proposal and guideline match only their first EUR-Lex form. /search handles both.
  • cited-by completeness: its edges are refreshed on read, and relationSync: pending means a recent amendment may not be listed yet.
  • The docs now describe the time limits per fetch and per request.
October 2026

Deprecated: GET /api/v1/documents/recent with more than one documentType or subdomain (API 2.4.0)

#
  • Send one value per request, or use POST /api/v1/search, which accepts lists. /documents/recent has only ever applied the first of several values. The others were dropped without notice. The spec has always declared both parameters single-valued, while these docs said "string / list".
  • Nothing changes yet. A repeated parameter still gets its first valid value, now with an X-Warning naming the ignored ones. Its responses carry the deprecation headers (X-LexAPI-Deprecations: recent-multi-document-type / recent-multi-subdomain). A single value is unaffected.
  • No removal date is set. When one is, it will be at least 180 days after it is published here. After it, the multi-value form gets 410 GONE. We email the accounts that sent a repeated parameter.
October 2026

Changed: bulk export is charged per row, as documented (API 2.3.0)

#
  • GET /api/v1/export now costs 1 credit per row written, at least 1 per export. The docs have always said "1 credit per row", but the server charged 1 credit per export call, whatever the number of rows. Credit plans now pay what the docs say. On legacy daily plans an export still counts as one call against the daily quota. An export now writes the usage row that makes it count, including an export you abort.
  • How the charge works. The request is admitted for 1 credit, which pays for counting the matching rows; an empty pool is a 402 CREDITS_EXHAUSTED before anything is counted. Then it reserves the rest of 1 credit for every row it can return: the matching rows, capped at limit and the 50,000 server maximum. The new X-Export-Credits-Reserved header and _meta.credits show the reservation. When the stream ends, every reserved credit beyond max(rows written, 1) is refunded: rows skipped as metadata-only stubs, rows that disappeared during the export, rows not sent because you disconnected. If the export fails on our side, you pay for the rows written only. The settled charge is on the trailer, in the new _done.credits (reserved, charged, refunded, plus the pool after settlement). refundPending: true marks a refund that is applied automatically within minutes.
  • Low balance. If your pool cannot cover the reservation, the request fails with 402 CREDITS_EXHAUSTED before anything streams, the same as a search whose pages your pool cannot cover. The admission credit is kept, because the count ran. error.details.weight is the reservation and error.details.available your balance; pass a limit at or below it.
  • Counts never cap the export. Rows added after the count are streamed too, each reserved as it comes; an export the pool cannot keep paying for streams as many rows as it can still pay for, then ends with truncated: true, truncatedBy: "credits". A count that takes longer than 15 seconds returns 504 TIMEOUT; the 1-credit admission is kept.
  • A failed export has no trailer. If an export fails after streaming has started, the stream now ends cleanly without a _done line; before, the connection was left open. An export is complete only if it ends with _done. The official SDKs raise an incomplete-export error when it is missing.
  • _done.truncated reports truncation correctly. It counted skipped stubs as rows cut off by the cap, so an export that had reached every matching row could say truncated: true. It now means matching rows were not reached because of limit or the server maximum; the new truncatedBy says which (limit, max_rows or credits). Stub rows beyond the last row written do not count either. skippedStubs is unchanged.
  • truncatedBy: "recovered". If an export is stopped on our side while it is still streaming (for example after a stalled connection), the stream stops at the next page and ends with truncated: true, truncatedBy: "recovered". The recovery's refund of the reservation stands, so _done.credits.charged can be below written; re-run the export from the last row you received.
  • New parameter changedSince. Only rows whose text changed (a new version) at or after this time. It is the cheapest incremental sync under per-row billing. fetchedSince keeps its documented meaning (lastSeenAt: any write to the row, including metadata refreshes and re-checks that left the text unchanged).
  • New parameter documentTypeCode selects a type family. regulation covers Regulation, Implementing regulation, Delegated regulation and the Commission implementing/delegated labels; likewise directive and decision.
  • documentTypeCode values for implementing and delegated acts. Documents labelled Implementing regulation, Delegated regulation, Implementing directive, Delegated directive, Implementing decision or Delegated decision now carry documentTypeCode regulation, directive or decision in every response, like the Commission implementing/delegated labels already did. They used to carry implementing_regulation, delegated_regulation and so on.
  • Fixed: documentType and author were case-sensitive. The docs' own example, documentType=regulation, matched nothing because the stored label is Regulation. Both now match whole stored values, ignoring case. documentType=regulation returns Regulation documents only; use documentTypeCode for the family. A joint act's author is one value: author=European Parliament, Council of the European Union selects it as a whole.
  • Fixed: dateFrom / dateTo now select by document date. Rows without a document date are left out of a dated export.
  • Malformed parameters are now a 400 INVALID_PARAMS, not charged: a dateFrom/dateTo that is not YYYY-MM-DD, a fetchedSince/changedSince that is not ISO 8601 with a timezone, a limit that is not a positive integer, and a repeated language.
October 2026

Changed: webhook URLs must be public HTTPS endpoints (API 2.2.0)

#
  • Webhook urls must be public https:// endpoints on port 443 or 8443, without credentials in the URL, on a public hostname or IP address. POST /webhooks, PUT /webhooks/:id and POST /webhooks/:id/test answer 400 WEBHOOK_URL_NOT_ALLOWED with details.reason when a URL doesn't qualify, and a FAILED or PAUSED webhook can only be set back to ACTIVE once its URL does.
  • A delivery to a URL that no longer qualifies fails (errorMessage starting Webhook URL not allowed) and is retried like any failed attempt. Webhook reads carry a new urlError field naming the problem with a stored URL.
  • Redirects are not followed. A 3xx answer is a failed delivery; point the webhook at the final URL.
  • responseBody in test results and delivery history now holds the first 500 characters of your endpoint's reply only when its content type is text, JSON or XML (text/*, application/json, +json, application/xml, +xml; null otherwise), stored as received rather than re-serialized. Test results gain durationMs.
October 2026

API versioning and deprecation policy (API 2.1.0)

#
  • Every response now names the API version. X-LexAPI-Version carries the contract version the server implements (2.1.0), and X-LexAPI-Release the deployed build — that version plus the commit it was built from, e.g. 2.1.0+633eb67. Both are sent on errors and on /health as well. Quote the release when you report a problem.
  • The contract version follows semantic versioning from here on. A MINOR bump adds something (endpoint, optional parameter, response field, enum value, header), a PATCH bump corrects the spec without changing behaviour, a MAJOR bump marks a breaking change. Every bump gets an entry here. 2.1.0 is the first version under this policy; earlier changes are listed under the months below.
  • GET /api/v1/info returns the contract version as version (2.1.0), the same value as the header. /health now reports the release as version, plus apiVersion, commit and builtAt.
  • A deprecation policy. Additive changes ship at any time; breaking changes only under a new major path (/api/v2). Anything under /api/v1 that is retired is deprecated first and keeps working for at least 180 days after its removal date is published. The 180 days are enforced in code: a removal date is refused unless it is at least 180 days after the day it is published, and CI refuses a publication date in the past. Responses that use a deprecated endpoint, parameter or field carry Deprecation (RFC 9745), Sunset (RFC 8594, once a date is set), Link: <…>; rel="deprecation" (and rel="successor-version") and X-LexAPI-Deprecations. After the removal date the request gets 410 Gone with the new error code GONE, naming the replacement. We email the accounts that used it in the previous 90 days when it is deprecated, and again when a removal date is set. Details: Deprecations.
  • Browser clients can read these headers. CORS now exposes X-LexAPI-Version, X-LexAPI-Release, Deprecation, Sunset, Link and X-LexAPI-Deprecations.
  • changelog.md is built from this file on every deploy.
October 2026

Deprecated: POST /api/v1/parse (API 2.1.0)

#
  • Use POST /api/v1/search instead. /parse is the original name of the search endpoint, kept as an alias: same handler, same request body, same response, same price. It is now listed in the OpenAPI spec as deprecated: true, and its responses carry the deprecation headers above (X-LexAPI-Deprecations: parse-endpoint).
  • No removal date is set. It keeps working. When a date is set, it will be at least 180 days after it is published here. We email the accounts that called /parse in the previous 90 days now, at deprecation, and again when the date is set.
October 2026

Semantic legislation search: consolidated texts and the text in force on a date

#
  • One hit per article when the act and its consolidated version both match. The legislation index now holds the latest consolidated versions alongside the acts as published in the Official Journal. When both match the same article, /legislation/semantic returns one hit: the consolidated article, labelled consolidated, consolidatedAsOf, baseCelex and consolidatedNote. The act as published is under baseAct, at the higher of the two scores. Consolidated texts are not authentic; the Official Journal text is.
  • New asOf parameter (YYYY-MM-DD). Each hit answers with the version in force on that date. That is the act as published before its first consolidation, or the consolidated version in force. Hits that cannot be answered for that date keep their latest text, with asOfResolved: false and an asOfReason. The response adds asOf and asOfSummary.
  • What asOf costs. The search costs 5 credits, as before. A version our corpus does not hold is fetched live, by default at most 3 per request. Each live fetch is billed like a document read for the same outcome: 4 or 5 credits, 1 when there is no text, 0 when it fails. A fetch your pool cannot cover is skipped, and that hit is insufficient_credits.
October 2026

Title language label (titleLanguage)

#
  • New titleLanguage field on document, batch and metadata responses. Some documents have no title in the language you asked for, for example a corrigendum that corrects only the Dutch version. We then show the title from a language that has one, and this field now names that language (en for the English fallback of a non-English request, or e.g. nl). It is absent when the title is in the requested language, and include: ["metadata"] keeps it.
October 2026

Fixed: article titles, section headings and author on consolidated texts

#
  • Article titles on consolidated texts. Some consolidated versions (e.g. 02025R0202-20260101) returned every article with title: null and the title at the start of content; title is now set.
  • Section headings no longer stick to the previous article; they go to paragraphs[].
  • author is no longer "Provisional data" or "Not available". Both are placeholders for consolidated texts, not institutions. /documentContent, /documents/metadata, /documents/recent, /search and webhook payloads now return author: null for these documents. If a real author is listed alongside it, only the real author is kept. The base act (CELEX 3…) still carries its real authors. This applies to documents already in our corpus as well.
  • The article title and section heading fixes apply to consolidated texts fetched from now on.
October 2026
#
  • Amendment, corrigendum and repeal edges for legislation now come from the Publications Office's Cellar repository, and a new relationSync field on /citations/cited-by and /citations/network says how current they are (fresh, refreshed, pending, …). Same price.
  • New citationType: "corrigendum". Corrigenda (32008R1272R(01), …R(02), …) used to be counted as amendments. They now have their own type, and each corrigendum is a separate citing document. Their titles are in the language(s) each corrigendum was published in.
  • English titles on these edges wherever Cellar has one; other languages are used only when no English title exists.
  • Fixed: wrong document dates on /documents/recent. Since late September, many rows carried another document's dateOfDocument, and the list order was off with them; some /search answers were affected the same way. Dates and ordering are correct again. If you stored dates from these endpoints since 22 September, re-fetch them.
October 2026

Pointer from a text-unavailable base act to its consolidated version

#
  • New textAvailability.latestConsolidated (and upcomingConsolidated). When a legislation base act (CELEX sector 3 — e.g. REACH 32006R1907, CLP 32008R1272, or a qualified id such as 32019D0003(01)) answers available: false on /documentContent, /documentContent/batch or /documents/url, the block may now name the act's newest consolidated version (CONSLEG, e.g. 02008R1272-20260701) — the act with its amendments integrated — plus the next one when the Publications Office has published it ahead of its date. Pass the celex back to /documentContent; that is a normal, separately charged read.
  • The hint is free and best-effort. No extra credits; omitted when the lookup doesn't answer, so its absence proves nothing. source: "corpus" marks an answer from our own rows while the Publications Office was unreachable (may not be the newest version).
  • It does not guarantee text. A consolidated version is its own document and can itself be too_large or not_in_language. Consolidated texts are documentation without legal value of their own — the authentic text is the one in the Official Journal.
  • /documents/metadata carries the same fields only when a recent document read already looked them up.

September 2026

September 2026

Search & recent documents

#
  • Unfiltered /documents/recent requests settle at a single page: 3 credits whatever the limit, at most 200 rows per call, marked X-Source: cellar. Requests filtering on documentType, author, domain or subdomain are still priced per page (X-Source: live).
  • /search names the source that answered in X-Search-Source (API docs); a cellar answer bills one page. A search answered from our corpus comes back with partial: true, source: "corpus" and X-Search-Source: corpus: the list may be incomplete, and text is matched against titles only. If nothing can answer, the request fails with a full refund (504 TIMEOUT or 500), never an empty list presented as "no results".
  • Identical searches are cached for an hour. The same filters and maxPages within 60 minutes are served from a result cache (X-Search-Cache: hit|miss|bypass; fetchedAt stays the real fetch time; same price). Pass fresh: true to force a re-fetch. query: "*" is now treated as "no text constraint".
  • Fixed: since 2 August, /search answers from the EUR-Lex web service (X-Search-Source: soap) returned too few results for year/month and none for domain: "EU_LAW". year/month now filter on the document date, and EU_LAW adds no restriction. Re-run searches made since 2 August.
  • Fixed: day-level date ranges. Some searches returned a whole month for a one-day range; those results are now trimmed to your dateFrom–dateTo.
September 2026

Plans & billing

#
  • Annual plans reset their monthly credits every month, as the pricing page says.
  • Plans we grant by hand (managed plans) no longer show self-serve billing screens (billing portal, top-ups, automatic top-ups). Plan changes made by our team send you an email naming the new plan, monthly credits and next refresh date.

August 2026

August 2026

Credit pricing: Cellar-served answers now settle at 4 credits

#
  • Three-tier settlement on document reads. Every document fetch, by-URL fetch, metadata call and batch item is charged 5 upfront and settles by who answered it: 1 when served from the LexAPI corpus (unchanged — still the typical case), 4 when the EU Publications Office's Cellar API answered (X-Source: cellar / cellar-oj), and 5 only for a browser fetch of EUR-Lex itself. Previously Cellar answers settled at 1. Batch responses expose the split via X-Batch-Corpus-Hits / X-Batch-Cellar-Served / X-Batch-OJ-Recovered / X-Batch-Live-Fetches (these four counts now partition the batch exactly).
  • OJ recovery after a failed live attempt bills the full 5. When a live fetch comes back textless and the text is then recovered from an Official Journal issue, the answer keeps X-Source: live.
  • Fixed: batch double-refund on text-unavailable items. Refunds are now exactly one per item.
August 2026

Documents whose text EUR-Lex doesn't publish

#
  • Correction to a June entry: until 19 August some documents without text returned 200 with a EUR-Lex error page as content.fullText (/documentContent, the batch, /documents/url and /export rows) instead of the textAvailability answer.
  • New textAvailability block on every document response. Usually { "available": true }. When EUR-Lex publishes a document's notice but not its text, you now get 200 with full metadata, content: null, an X-Text-Available: false header, and a machine-readable reason — not_in_language (with the languages that do have text), not_published_yet, no_machine_readable_rendition, or unknown. Deliberately not a 404: the document exists, and /resolve, /citations and /versions all answer for it. content is now nullable — check textAvailability.available before reading it.
  • These answers cost 1 credit, not 5. Same price as /documents/metadata, which is all you received. On batch requests they stay in documents[] with a textUnavailable count and an X-Batch-Text-Unavailable header, rather than being moved to errors[] where their metadata would be lost.
  • Parliamentary questions now return their real text. Written and priority questions have no per-document rendition on EUR-Lex — their text is published inside an Official Journal issue. We now extract it (X-Source: cellar-oj). Questions from before 2012 have no text anywhere upstream and say so.
  • A request for a language EUR-Lex can't serve now falls back to English properly. Previously the fallback only fired when the whole document was missing, so language=fr on a document with no French text returned the error page instead of the English text.
  • bypassCorpus does what the docs said: a forced re-read is now genuinely fresh.
  • /export skips metadata-only rows by default; includeStubs=true restores the old behaviour, and the export trailer reports skippedStubs.
  • Published CELEX pattern corrected. The OpenAPI spec advertised ^[0-9A-Z]{2,20}$, which rejects corrigenda (32013R0952R(04)) and consolidated texts (02016R0679-20180524) that the API has always accepted — generated clients were refusing requests the server would have answered.
  • acceptLanguages: the authentic text in one call. Most General Court decisions have no English text — the authentic text is the language of the case, usually alongside French. Pass acceptLanguages (an ordered list of ISO codes, or "any") on /documentContent or the batch endpoint and, when the requested language has no text, the first accepted language that does is served instead — clearly labeled with languageRequested, languageFallback and X-Language-Fallback, the same semantics the automatic English fallback has always used. Off by default: without it you keep the textAvailability answer with its availableLanguages list.
  • /corpus publishes a third number. Documents held as metadata only — because EUR-Lex publishes no text for them — are now counted separately instead of being folded into "with full parsed text."

July 2026

  • Faster, indexable docs & blog. /docs and every blog post are now served as pre-rendered HTML — instant first paint, working link previews, and reliable search indexing.

June 2026

June 2026

Pricing v2 — the monthly credit pool

#
  • New pricing model launched. Plans now come with a monthly credit pool instead of a daily call cap: FREE 500 · STARTER 30,000 · PROFESSIONAL 150,000 · BUSINESS 500,000 credits/month. Operations draw credits by weight — 1 for fetch/search/citations lookups, 2 for citation extraction, 5 for semantic search, 20 for citation network/graph calls.
  • Existing customers grandfathered. Accounts created before the cutover keep their daily-call plan; their responses report subscription.pricingMode: "daily" and no credits object.
  • Top-up packs. Buy +10k/+50k/+200k credits mid-cycle; top-ups are spent after the monthly allowance and don't expire at the monthly reset; they end with the paid plan. FREE-tier accounts cannot buy top-ups. (How used top-up credits are deducted at each reset changed in API 3.0.0.)
  • Semantic search on FREE. On the new pricing, the credit pool gates semantic search, so FREE keys can use it too (legacy FREE keys still get 403).
  • Typed 402. When the pool is empty, metered calls return 402 CREDITS_EXHAUSTED with operation, weight, available, used, and resetsAt in error.details.
  • Mode-aware stats + docs. Usage/stats endpoints and the docs now speak credits for v2 accounts and daily calls for legacy accounts.
June 2026

MCP server

#
  • @lexapi/mcp published to npm. Plug LexAPI into Claude and other MCP-capable agents: npx -y @lexapi/mcp with your API key.
  • New resolver & graph tools. lex_resolve (any identifier → canonical document), lex_citation_path (shortest citation path between two documents), and lex_related (neighborhood exploration).
June 2026

Account security

#
  • Email verification is now required. New accounts must verify their email before API keys can be issued.
  • Two-factor authentication. Opt-in TOTP 2FA (authenticator app) for dashboard login.
  • Self-serve password reset from the Settings page.
June 2026

API behavior fixes

#
  • Typed error envelope. Failures on document, search, resolve, and citation routes return { success: false, error: { code, message } } with a stable machine-readable code — a missing document is now always a 404 NOT_FOUND, never a placeholder document. See the error catalog. (Partly overstated — see the August 2026 correction: documents whose notice existed but whose text did not still returned a placeholder until 19 August 2026.)
  • language honored end-to-end. /documentContent localizes both content.articles and content.fullText, with explicit English fallback (languageRequested, languageFallback, X-Language-Fallback header). Unsupported codes are a typed 400.
  • Truncation is explicit. Search responses carry truncated / truncatedAt / truncatedReason when tier caps cut a result set; /export sets X-Export-Truncated. Requested document types the engine can't map are surfaced via ignoredDocumentTypes + X-Ignored-DocumentTypes instead of being silently dropped.
  • documentTypeCode. Documents carry a stable lowercase type enum alongside the display documentType.
  • Quota visibility. usage.resetsAt is exposed on successful responses (not just 429s), and every metered envelope reports unitsCharged.
  • Semantic hint field. Low-confidence or empty semantic results carry a best-effort hint so callers can adjust queries.
  • Case-law citation graph. Case→case citation edges added — judgments now appear in cites/cited-by/network results.
  • Citation network hardening. Long traversals return 200 with partial: true, partialReason: "TIMEOUT" instead of hanging.
  • Upstream timeouts are typed. Slow EUR-Lex fetches return 504 TIMEOUT (retry-safe) rather than a generic 500.
  • Search date-range fix. Year-aligned date ranges combined with documentType no longer return 0 results.
  • Consolidated CELEX accepted. Point-in-time identifiers like 02016R0679-20180524 validate correctly.
  • Case-law text structure. paragraphs[] for judgment grounds; availableLanguages surfaced on search hits.
  • /export parity. Bulk export rows match /documentContent field-for-field, including the subdomain filter.
  • Field selection. include/fields trims document payloads; articleId fetches a single article.
June 2026

Docs & dashboard

#
  • AI-readable docs. The full reference ships as Markdown at /docs.md plus llms.txt, with a copy-for-LLM button on /docs.
  • Credit-aware dashboard. Usage widgets show the credit pool for v2 accounts; usage chart gained a 24h/7d/1m/3m range switch; live /status page.

May 2026

  • Document corpus. LexAPI keeps its own copy of the EU corpus with version history; X-Source says what answered, and bypassCorpus: true forces a fresh fetch.
  • Bulk export. GET /api/v1/export streams the corpus as NDJSON with filters and incremental fetchedSince sync (BUSINESS tier).
  • OpenAPI 3.1 spec published.
  • Password reset emails and the corpus-coverage docs page; site-wide editorial redesign.

Earlier

  • February 2026 — Semantic search launched: AI-powered case-law search by meaning, with results enriched with full EUR-Lex text. Promotion codes enabled at checkout.
  • November 2025 — Webhooks (signed deltas for saved searches) and citation extraction launched.
  • October 2025 — FREE tier introduced; daily limits on paid tiers raised significantly; search API expanded with comprehensive EUR-Lex filters (dates, types, authors, procedures, domains).