Skip to content

Connect your AI agent (MCP)

Tarno ships a neutral MCP server so an AI agent can query the INAPI trademark corpus through the Model Context Protocol (everything is a read, except subscribing and unsubscribing the applications your own org watches, with watch:write). It exposes the same read contract as the REST API — backed by the same query layer and the same schemas — so the two transports return identical results.

Endpoint and transport

Property Value
Endpoint https://mcp.tarno.cl
Transport Streamable HTTP (MCP)
Auth X-API-Key header, sent per request
Server identity tarno-read-mcp

The server answers MCP JSON-RPC over the Streamable HTTP transport. A GET /health probe is the only unauthenticated route; every tool call is authenticated.

One JSON-RPC message per HTTP request. Batches (an array of messages in the body) are rejected with HTTP 400 and JSON-RPC error -32600. The MCP specification removed them in 2025-06-18 and the official clients do not send them. Every tool call and every read of a resource with a REST twin (resources/read) spends one unit of quota and of burst limit, exactly like its twin REST request. A call whose arguments do not match the tool's schema spends nothing (like a REST 400) and returns the error in the same Problem Details body as every other error (status: 400, code: "validation_error" and per-field errors[]); the protocol (initialize, tools/list, resources/list…) spends nothing either.

Authentication

Authentication is per tool, not per connection. Each tool handler verifies the presented key before it reads anything, so a missing or invalid key never returns corpus data. Send your operator-issued key in the X-API-Key header on the transport's requests (placeholder <API_KEY> in these docs):

X-API-Key: <API_KEY>

Each tool requires the same scope as its REST twin: brands:read for most; insights:read for the system, aggregate and holiday ones; watch:read/watch:write for the four watch tools; legal:read for the documentary base; and each report, its own permission. An unrestricted (empty-scope) key has the base reads — see the auth guide. A key that is missing/invalid is rejected as an unauthorized tool error; a valid key lacking the scope is rejected as forbidden.

⚠ The eight reports (get_daily_bulletin, get_clearance_report, get_portfolio_report, get_opposition_brief, get_opposition_response_brief, get_conflict_landscape, get_registration_signals and get_filing_context) carry their own permission: reports:read (all) or the report's own (reports:daily-bulletin:read, reports:clearance:read, reports:portfolio:read, reports:opposition-brief:read, reports:opposition-response-brief:read, reports:conflict-landscape:read, reports:registration-signals:read, reports:filing-context:read). They are granted explicitly — an unrestricted key (empty scopes) does not get them, nor does the trial key — and each call counts one quota unit.

⚠ analyze_logo also carries its own permission, logo:analyze, granted explicitly under the same rule (neither an unrestricted key, nor the trial key, nor reports:read gets it): every call costs money at a third party and has its own daily quota as well (logo analysis). And find_similar_logos, the future operation, carries its own, logo:similar, under the same rule (logo:analyze does not open it either); it is granted to no key until it is announced as available.

The transport-level initialize handshake may succeed without a key — the hard invariant is that a tool call never returns corpus data unless the key is valid and scoped. Always send the header.

Quota

The same monthly quota that governs the REST API applies to the MCP. When a key is over quota the server responds with HTTP 429 and a Retry-After header (seconds until reset), byte-for-byte matching the REST quota response:

HTTP/1.1 429 Too Many Requests
Retry-After: 1209600

{ "type": "https://api.tarno.cl/problems/quota_exceeded", "title": "Rate limit exceeded", "status": 429,
  "detail": "Monthly quota exceeded", "code": "quota_exceeded", "requestId": "..." }

If the key cannot be verified because the key-validation service is temporarily unavailable, the server answers HTTP 503 with code: "service_unavailable", the same body as REST, and the tool does not run. Retry with backoff.

Resources, prompts and rules

Beyond the tools, the server exposes (the manifest contract/mcp-manifest-38.json announces exactly these names: a test builds the server and compares what is served with what is announced):

  • Resources — the corpus catalogs, enumerable by URI: iris://catalogs/states, iris://catalogs/sections, iris://catalogs/nice-classes, and iris://catalogs/holidays/{year}. They come from the SAME source as the REST routes and are read with the SAME key and scope as their route (brands:read; insights:read for holidays), and each read spends one unit of quota like its route (states, with no REST route, spend nothing); listing them needs no key and spends nothing. The Nice classes (list_nice_classes ≡ GET /nice-classes and this resource) carry the headings of the edition in force in Chile TODAY: 13-2026 until 2026-12-31 and 13-2027 from 2027-01-01 (classes 6, 9, 21, 29 and 31 change), with nobody doing anything; pt-BR is Tarno's own translation. The catalog tools (list_nice_classes, list_legal_norms, list_sections, list_holidays) give in meta.dataAsOf the date of their source —WIPO, BCN, the calendar—, not the trademark sync; for list_sections, the day the INAPI code legend was last read. The family of a section (family, and sectionFamily on annotations) is an open string: today presentacion, forma, fondo, publicacion, pago, oposicion, resolucion, renovacion and, since 2026-10-10, caducidad (the acts of a cancellation-for-non-use proceeding; not the status caducado) and, since 2026-10-11, cancelacion_y_limitacion_territorial (cancellation and territorial-limitation proceedings). A value you do not know is another family, not an error. Every annotation carries annotationNumber, INAPI's annotation number (the «N° de anotación» of the file, the one cited); a resolution carries the number of the annotation it resolves. It is null when unknown (the Estado Diario does not print it): a gap is null, never a guessed number. meta.coverageThrough is null on catalogs (they have no "through"), except on list_holidays: the last day with a declared calendar. Details in Coverage.
  • Prompts — three, by their real names: availability_check (is a name available? — it uses search_trademarks with similarTo for similar marks and q for the literal match), case_history (what happened to an application?) and weekly_watch (which matches do the applications I watch have? — its only argument, applicationNumber, is optional: with it, it reads list_similarity_hits of that application; without it, list_watched_applications and list_similarity_hits of all of them). They give signals, not verdicts.
  • Version — the contract date (2026-10-04), the same as the OpenAPI.
  • Instructions — the four reading rules of the corpus, each with the failure it prevents: state the data's date; omit what is unknown; give signals, not verdicts; deadlines only with their article.

Tool catalog

The server registers 54 tools —one of them, find_similar_logos, a future operation that always answers 503 today—. All of them read except two that write, the watch tools under watch:write (watch_application, unwatch_application), which only touch what your own org watches. 53 are the exact twin of a REST route: same validation, same data, same error codes; the one without a route is explain_section_code, which reads the INAPI knowledge base, not the corpus. Parity reads "every tool is a route" (except that one), not the other way round: four REST routes exist only because they are REST and have no tool of their own — GET /trademarks (linkable/cacheable) is the REST twin of search_trademarks, GET /trademarks/application-numbers that of list_application_numbers, GET /trademarks/renewal-calendar that of get_renewal_calendar (the three without the portfolio object, which does not travel in a URL), and POST /insights/trademarks/search (for filters that do not fit in a URL) that of get_trademark_insights. search_trademarks and get_trademark are the entry points (detailed below); the full catalogue is in the table that follows (your MCP client enumerates it with tools/list) and includes, among others, get_trademark_insights (aggregates by groupBy), list_holidays (the holiday calendar and the range it claims to cover), get_system_freshness, list_sync_runs, list_parties and the four similarity-watch tools —watch_application, unwatch_application, list_watched_applications, list_similarity_hits— (guide). watch_application returns the same body as REST and does not distinguish a new subscription from a repeated one (REST: 201 / 200); unwatch_application returns { "applicationNumber": "N", "watched": false } (REST: 204 with no body). list_watched_applications announces the end of a watch with terminalSince, watchEndsOn (the last day of the watch, not a legal deadline) and watchEndBasis; the day after, the application is no longer listed. list_similarity_hits returns only the matches in force: one that becomes terminal or no longer resembles disappears with no notice and, if it comes back, does so with a new id (how a watch changes). It is a feed: next is always returned, you page while hasMore is true, and the last next, sent later as cursor, brings only what the organisation came to see since (how to read what is new).

The 54 tools, with their REST twin and the scope they require (reports:read also grants every report):

Group Tool REST twin Scope What it does
Trademarks search_trademarks POST /trademarks/search brands:read Collection, batch, similarity and opposition window in one call
Trademarks get_trademark GET /trademarks/{applicationNumber} brands:read One trademark, with optional expansion
Trademarks compare_trademarks POST /trademarks/compare brands:read Compares two marks channel by channel: signals, no verdict
Trademarks get_deadlines GET /trademarks/{applicationNumber}/deadlines brands:read The deadlines of a trademark with their rule: to appeal a rejection or a partial acceptance, to pay the grant fees and to answer the transfer of an opposition
Trademarks get_renewal_calendar POST /trademarks/renewal-calendar brands:read The renewal milestones (art. 24) of a set of marks
Trademarks list_application_numbers POST /trademarks/application-numbers brands:read Every application number that matches a filter, in pages of up to 10,000 (next is the cursor; exact total; 409 cursor_stale if a sync ran between pages); takes the portfolio object. Without it, the same answer as GET /trademarks/application-numbers
Trademarks get_trademark_insights GET /insights/trademarks insights:read Aggregates by groupBy over the matching marks
Annotations search_annotations POST /annotations/search brands:read Cross-cutting annotation search and the feed of detected acts
Annotations list_trademark_annotations GET /trademarks/{applicationNumber}/annotations brands:read A trademark's annotations
Annotations get_annotation GET /annotations/{id} brands:read One annotation by id
Parties list_parties GET /parties brands:read Parties, with their trademark count
Parties get_party GET /parties/{id} brands:read One party
Parties list_party_trademarks GET /parties/{id}/trademarks brands:read A party's trademarks, by role
Parties list_party_relations GET /parties/{id}/relations brands:read A party's relations with others
Parties get_party_outcomes GET /parties/{id}/outcomes brands:read The outcomes of a party's trademarks
Catalogues and calculation list_nice_classes GET /nice-classes brands:read The 45 Nice classes of the edition in force
Catalogues and calculation get_class_stats GET /nice-classes/{niceClass}/stats brands:read Statistics of a Nice class, with its coverage
Catalogues and calculation recommend_nice_classes POST /nice-classes/recommend brands:read Nice classes for a free description
Catalogues and calculation list_sections GET /sections brands:read INAPI's sections
Catalogues and calculation get_section GET /sections/{code} brands:read One section, with its equivalences
Catalogues and calculation explain_section_code — brands:read What a section code means (no REST route)
Catalogues and calculation list_legal_norms GET /legal-norms brands:read Catalogue of cited statutes
Catalogues and calculation get_legal_norm GET /legal-norms/{id} brands:read One statute, with its articles
Catalogues and calculation list_holidays GET /holidays insights:read Holiday calendar and the range it claims to cover
Catalogues and calculation compute_business_days GET /business-days brands:read The Nth business day after a date, with its rule
Catalogues and calculation calculate_fees POST /fee-schedule/calculate brands:read The fees of an application or renewal
Reports get_daily_bulletin GET /reports/daily-bulletin reports:daily-bulletin:read A day's Estado Diario bulletin
Reports get_clearance_report GET /reports/clearance reports:clearance:read Availability signals for a name
Reports get_conflict_landscape GET /reports/conflict-landscape reports:conflict-landscape:read Who represents and holds the live similar marks
Reports get_registration_signals GET /reports/registration-signals reports:registration-signals:read A name's indicative index, by signals
Reports get_filing_context GET /reports/filing-context reports:filing-context:read The facts of a filing, each with its source
Reports get_portfolio_report GET /reports/portfolio reports:portfolio:read A party's portfolio
Reports get_opposition_brief GET /reports/opposition-brief reports:opposition-brief:read Facts to oppose a mark
Reports get_opposition_response_brief GET /reports/opposition-response-brief reports:opposition-response-brief:read Facts to answer an opposition
Documentary base list_guidance_documents GET /guidance/documents legal:read INAPI's documents with their version in force
Documentary base get_guidance_document GET /guidance/documents/{code} legal:read A document with its Summary and versions
Documentary base get_guidance_section GET /guidance/documents/{code}/section legal:read A literal section, with page, version and citations
Documentary base search_guidance POST /guidance/search legal:read Search the documents in force
Documentary base list_guidance_changes GET /guidance/changes legal:read Which sections changed and when
Documentary base get_legal_norm_article GET /legal-norms/{id}/articles/{article} legal:read An article's text on a date
Documentary base search_statutes POST /legal-norms/search legal:read Search statute articles
Documentary base get_fee_schedule GET /fee-schedule legal:read The fee table in UTM and pesos
Documentary base list_economic_indicators GET /economic-indicators legal:read SII indicators (UTM, UF, CPI…)
Logos analyze_logo POST /logos/analyze logo:analyze Describes a logo and proposes Vienna codes (model)
Logos find_similar_logos POST /logos/similar logo:similar FUTURE: always 503 today
Watch watch_application POST /watch/applications watch:write Watch an application by number
Watch unwatch_application DELETE /watch/applications/{applicationNumber} watch:write Stop watching it
Watch list_watched_applications GET /watch/applications watch:read What your organisation watches
Watch list_similarity_hits GET /watch/hits watch:read The matches of what you watch
System get_system_me GET /system/me insights:read Your key: scopes, quota and 12-month usage
System get_system_freshness GET /system/freshness insights:read Corpus freshness per source
System get_system_coverage GET /system/coverage insights:read Ingestion coverage by bulletin date
System get_vigilance_horizon GET /system/vigilance-horizon insights:read The window retention keeps
System list_sync_runs GET /system/sync-runs insights:read Sync-run history

⚠ [BREAK] 2026-10-04: the ten watchlist tools (create_watchlist, list_watchlists, get_watchlist, update_watchlist, delete_watchlist, add_watchlist_item, remove_watchlist_item, list_watchlist_items, list_watchlist_hits, list_watchlist_deliveries) were retired at once, with no coexistence. Those four replace them: you watch an INAPI application by its number, with no lists, filters or webhook. The before → after map is in the watch guide.

get_renewal_calendar ≡ POST /trademarks/renewal-calendar — the renewal calendar (2026-09-25). With the required from/to (at most 366 days, both ends included) and the trademark filter —the batch of up to 100 applicationNumber, renewalStatus, holder, niceClass…—, one row per (mark, renewal milestone) in the range: renewal_window_open and grace_end (Ley 19.039 art. 24), each identical to the milestone in the mark's milestones[], plus its applicationNumber, by date and number. "The renewal dates of my list this year" is ONE call. Pages with limit + cursor, exact total. q, similarity and the opposition window are validation_error. Takes the portfolio object (a mark added by hand is in even when it fails portfolio.refine); without it, the same answer as GET /trademarks/renewal-calendar. Details: reports.

list_parties — both orders are fast. Without sort, parties come by trademark volume (largest first), served by a count stored at the end of every sync; sort: "name" chooses the page first. While a sync is running the volume order is computed on the fly (same response, 1–3 s; up to 8 s with kind/role). Parties with no marks are never hidden (trademarkCount: 0), and meta carries dataAsOf and coverageThrough.

list_party_relations — the relations of one party (2026-09-24). Twin of GET /parties/{id}/relations. Typed edges with other parties: co_holder, represents/represented_by, assigned_to/assigned_from (A07T assignments) and opposed/opposed_by (identified opponents). Each with the other party, trademarkCount, firstOn/lastOn, up to 5 evidence marks (+ evidenceTruncated) and coverageFrom: oppositions only since 2024-04-19 — "never opposed" reads "not since then". Filters type (a list) and relation. ⛔ Signals with evidence, never verdicts: no "same group", no name similarity. Details: people search.

The one-call reports (2026-09-24). Their own permission (reports:read or the report's own), denied by default. Details: reports. - get_daily_bulletin ≡ GET /reports/daily-bulletin: what the Estado Diario published on date, row by row with parties, paginated, with an exact total and the day's index in meta.sections ({code, label, count}). A day with no ingested bulletin is out_of_coverage (422) with the coverage, never an empty list. - get_clearance_report ≡ GET /reports/clearance: given text and niceClass, the confusable marks by channel (phonetic, fuzzy, semantic) with inOppositionWindow (membership; the close travels in milestones as opposition_close), the classes' dominant holders and the norms cited in the confusables' annotations. ⛔ Signals, never "registrable" or "infringes". - get_portfolio_report ≡ GET /reports/portfolio: the portfolio of ONE party by its partyId (never by its name) — counts by status and by class, the marks with an expiry (expiry) within horizonDays (default 90, max 366), those inside their opposition window today (membership; the close travels in milestones as opposition_close), who opposed, its representatives and the summary of its relations. Every mark carries its renewalStatus (art. 24), and byRenewalStatus counts them by status, with the undated and the not-applicable ones apart. Details: reports. - get_opposition_brief ≡ GET /reports/opposition-brief: to FILE an opposition against applicationNumber — its card with milestones, whether it is in its window today (membership; the close travels in milestones as opposition_close), with ownApplicationNumber where it ranks among the 100 marks most similar to yours in its classes (similarity; without your mark the key is absent), its holders' opposition history and the norms cited in the oppositions its holders' marks received. Details: reports. - get_opposition_response_brief ≡ GET /reports/opposition-response-brief: to RESPOND to the oppositions received by applicationNumber — its opposition acts (section in three languages, norms, resolved opponents), and per opponent its oppositions and its current marks in your mark's classes, found by its id and never by its name. Unresolved opponent mentions are counted, never named. Details: reports.

⛔ All five give signals, never verdicts: none says whether a mark is registrable or whether an opposition will succeed, and none publishes an opposition closing date outside milestones.

The documentary base (2026-09-28). Its own scope, legal:read, denied by default: an unrestricted key does not have it (it is granted explicitly). Details: documentary base. - get_legal_norm_article ≡ GET /legal-norms/{id}/articles/{article}: the LITERAL text of an article of a catalog norm (id from list_legal_norms) in the version that applied on asOf (default: today in Chile), with its BCN letters, subsections and notes. letter/subsection narrow it; transitional: true reads the transitional article. An article that could not be split safely is served whole (split: false, article_not_split), never with an invented cut. A norm with no loaded text or a date outside its windows is out_of_coverage (422); a duplicated key, ambiguous_selector (409). dataAsOf is the statutes source date, not the trademark corpus date. - list_guidance_documents ≡ GET /guidance/documents: INAPI's documents (chapters of the Guidelines, circulars, Madrid page and manual, fees), each with its current version and its extractionStatus; filters kind and q (title, accent-insensitive). - get_guidance_document ≡ GET /guidance/documents/{code}: one document with its outline (every section of the version, in order, WITHOUT bodies) and its version history; version reads an earlier one. An image_only version answers outline: [] and says so; an ocr one (images machine-read) has one section per page and its text is NOT literal. - get_guidance_section ≡ GET /guidance/documents/{code}/section: one LITERAL section with its page, its version, its children and the norms it cites, by path (III.D.▪1.§7) or by paragraph (the § a lawyer cites). The stable citation is code + version.seq + path. A repeated paragraph is ambiguous_selector (409) with the paths; a version without text (image_only), out_of_coverage (422). dataAsOf is the INAPI source date. - search_guidance ≡ POST /guidance/search: searches INAPI's CURRENT sections by words (Spanish, accents ignored) and by meaning (semantic search), fused; kind/document narrow it. - search_statutes ≡ POST /legal-norms/search: searches statute ARTICLES in the text in force today (the same rule as get_legal_norm_article), in the one of asOf, or with allVersions: true in EVERY validity — the same text repeated across validities comes once, with all of them in validity. norm narrows to one catalog norm (no text: out_of_coverage).

Both: mode hybrid (default), text or semantic; matchedBy says which channel found each hit; if the meaning channel cannot run, hybrid says so in meta.channels.semantic: "unavailable" and semantic is service_unavailable (503) — never an empty list. LITERAL excerpts, no scores. Details: search. - get_fee_schedule ≡ GET /fee-schedule: the fee table of chapter 13 of the Guidelines in UTM, LITERAL, with amountClp DERIVED on read (amountUtm × utm.value, exact, not rounded) with the UTM of utmMonth (default: the current month in Chile); utm.basis says whether it is that month's (requested_month) or the latest published (latest_published); no UTM, null. surchargeDoubt is the informative note about the late-renewal surcharge: the rule is the Directrices 2026 table, and the literal notes of the INAPI publications that differ block nothing. Details: fees. - list_economic_indicators ≡ GET /economic-indicators: UTM, UTA, IPC, UF or observed dollar by code and from/to range, with meta.coverage; a range entirely outside is out_of_coverage (422). ⚠ Every IPC row carries its baseNote: points of different bases are not comparable. - list_guidance_changes ≡ GET /guidance/changes: which sections changed between versions of an INAPI document or a BCN text and when we detected it (BCN's literal effectiveDate); at most 100 sections per change (sectionsTruncated). meta.dataAsOf: the source of the publisher asked for, or the older of the two. - get_annotation with expand: ["citations"] ≡ GET /annotations/{id}?expand=citations: the explained citation — for each norm the annotation cites, the LITERAL text of the article/letter that applied on its date (text.asOf; the same rule as get_legal_norm_article) and the guidelines' signals: the documents that cite that article/letter with citingCount/unnamedCount and up to three sections, never «the» section that interprets it. textStatus says why a citation has no text. It requires legal:read in addition to brands:read (without it, forbidden, consuming no quota); only on the detail — the lists reject expand. Without expand, the usual annotation. Details: explained citation.

Analysis and calculation. Neutral tools that calculate or gather signals with their source, never verdicts. Details: analysis and calculation. - compute_business_days ≡ GET /business-days: the N-th business day STRICTLY after from (Monday to Friday without the calendar's holidays — the same count as the opposition window), with the weekend days and holidays skipped, the calendar's declared coverage and the legal basis of the computation (Ley 19.039 art. 11, Ley 19.880 art. 25). Exactly one of businessDays (you set N: term.source: caller, term.legalBasis: null — a count, not a legal term) or preset; the only preset is opposition (art. 5, 30 business days from publication), with the same date as the trademark's opposition_close. A non-business-day anchor or a span outside the calendar give basis: estimated with its reason. Scope brands:read. - calculate_fees ≡ POST /fee-schedule/calculate: INAPI's OFFICIAL fees for application_and_registration or renewal, by niceClassCount or applicationNumber, in literal UTM and pesos at read time. In the renewal grace period the month row of the Directrices 2026 table applies (renewal.surchargeRuleSource) and the TOTAL is computed, with the other publications' discrepancy as a note (renewal.surchargeDoubt); after it, lapsed. Never includes the Diario Oficial nor professional fees (excluded). Scope brands:read. - get_party_outcomes ≡ GET /parties/{id}/outcomes: how a party's trademarks ended up, by published status group, with n and grantShare over the decided ones (pending excluded; null with none decided) and the full definition. Not a «success rate». Scope brands:read. - compare_trademarks ≡ POST /trademarks/compare: two trademarks (text or applicationNumber) compared channel by channel —trigrams, edit distance, Spanish phonetics, semantic similarity— with each side's generic terms DECLARED and distinctiveSignals over the cores; no global score, no level (compare). Scope brands:read. - get_class_stats ≡ GET /nice-classes/{niceClass}/stats: the statistics of a class by whole filing years (by default, the last five complete ones): population, outcomes with the groups of get_party_outcomes, APPROXIMATE percentiles (±15 days) of the time to registration with their n, share with substantive, formal and opposition acts, and the coverage stated in coverage. 503 class_stats_pending before the first computation (statistics). Scope brands:read. - get_conflict_landscape ≡ GET /reports/conflict-landscape: the LIVE similar trademarks of clearance aggregated by representative and by holder, with evidence and their total in the class (landscape). A REPORT: scope reports:conflict-landscape:read or reports:read. - get_registration_signals ≡ GET /reports/registration-signals: the orientative index as a BREAKDOWN of six signals per class (1 to 3), each with its value, n, evidence and source, and a 0–100 signalIndex with its literal formula and the applied (provisional) weights; "not a verdict" (index). A REPORT: scope reports:registration-signals:read or reports:read. - recommend_nice_classes ≡ POST /nice-classes/recommend: the Nice classes of the edition in force (13-2026; 13-2027 from 2027-01-01) for a free-text description, by four declared routes —official WIPO terms by meaning and by text, headings and corpus evidence—, with the WIPO attribution (recommend). Scope brands:read. - get_filing_context ≡ GET /reports/filing-context: the FACTS of a filing (power of attorney, opposition, answer to an opposition, answer to observations, withdrawal), each with its source; the terms with an article in milestones (the opposition close and, since the change of 2026-10-10, the answer to the transfer, opposition_response_close), and the omitted ones with their reason (milestonesOmitted). No drafting: the consumer writes it (facts). A REPORT: scope reports:filing-context:read or reports:read; with expand=legalTexts, legal:read as well. - analyze_logo ≡ POST /logos/analyze: a language model (Anthropic) describes a logo in INAPI's terms and proposes Vienna VCL10 codes, checked against WIPO's catalogue; the response says a model generated it (generatedBy, disclaimer). The image goes in the arguments (base64, ≤ 700 KB, ≤ 2,000 px), is sent to Anthropic and is not stored (logo). Permission logo:analyze; own daily quota (20 per key, 200 in total); with no provider key on the server, 503 logo_analysis_not_configured. - find_similar_logos ≡ POST /logos/similar — ⛔ FUTURE OPERATION, not available yet: the trademarks whose logo looks like an image. It is declared (input: the same image as analyze_logo, niceClasses, limit; future output: results[] with score and signals.visual of visual resemblance) so that you can program against it, but today it always answers 503 image_similarity_not_available, never results. Do not retry it nor present its answer as "no similar trademarks" (resemblance by image). Permission logo:similar, granted to no key yet. It is the last one in tools/list. - search_annotations also takes partyId/taxId (+ partyRole, relation) and trademarkStatus; search_trademarks, partyRole next to taxId (annotations by party, RUT). - Every annotation (search_annotations, get_annotation, the bulletin) carries inapiTermLabel: the term INAPI prints in the act's name, quoted (basis: "inapi_label", e.g. "30 días") for a closed list of 14 observation codes; null otherwise. ⛔ Not a date nor a milestone (the INAPI label). - search_annotations by list and by detection: applicationNumbers (up to 100 trademarks by cursor; more ⇒ 422 batch_too_large), detectedAfter/detectedUntil (when the corpus learned of it), denomination and detectedAt on every annotation and the meta.detectedThrough horizon to chain «changes since X» without skipping rows (by list and detection). detectedAfter on its own (no trademark, list or party) is the corpus-wide feed: the acts of all trademarks in detection order, with no content filters; maxActAgeDays leaves out acts that were already more than N days old when detected (a page can come short or empty with next: follow it) (the feed).

get_trademark_insights — the groupBy dimensions. filedMonth, status, niceClass, holder, representative, since 2026-09-24 expiresMonth and transferYear, and since 2026-09-25 renewalStatus. All of them return { key, label, count } with label in es-CL/en/pt-BR, and apply the full search_trademarks filter (or 400). - expiresMonth: marks per month of expiry (key = 2027-03-01, label = "marzo de 2027" / "March 2027" / "março de 2027"). ⚠ Marks without an expiry date fall in no month and are DECLARED in meta.undatedCount (Σ count + undatedCount = filtered marks). There are many — pending applications do not expire yet —: do not treat them as nonexistent. - transferYear: per year of the act, the marks with a transfer (INAPI code A07T) that retired a known holder, since 2011-08-01. A mark counts in every year it had one; name changes are not transfers. - renewalStatus: marks by their renewalStatus (the same value every mark carries; art. 24, today in Chile time): not_due "Renewal not yet open", window_open "Renewal window open", grace "In force, in renewal grace period", lapsed "Renewal term passed". ⚠ Marks without an expiry date go to meta.undatedCount and those with a date the rule does not apply to (Madrid, or neither registrada nor esperando_renovacion) to meta.notApplicableCount: Σ count + undatedCount + notApplicableCount = filtered marks. Never a guessed status. - niceClass labels the number in the three languages: "Clase 9" / "Class 9" / "Classe 9".

Withdrawn on 2026-09-24 — the trademark listing tool and the aggregate body-search tool were exact twins of search_trademarks and get_trademark_insights (same schema, same envelope, same path). Calling either is now an unknown tool; the SAME question, with the same arguments, is asked with the one that remains. The REST routes do not change.

⚠ The 2026-09-23 migration guide is historical and keeps the names of that time: where it says list_trademarks use search_trademarks, and get_trademark_insights instead of search_trademark_insights (same arguments).

People say what KIND of party they are. Every person in get_trademark (parties, holders, representatives, formerParties) carries kind (natural/legal/unknown), kindSource (tax_id/suffix/manual/name_pattern, omitted, never null, when there was no signal), observation (the jurisdiction clause INAPI publishes glued to the name, verbatim), publishedName (the name exactly as INAPI published it) and revision. In list_parties each person carries the same kind, kindSource (there null when there was no signal), publishedName, the clause as jurisdictionClause and the review as review. They are read under the same brands:read scope. REST≡MCP parity is verified by a deep-equal test against the real database, including the omission of kindSource. What each one means: people search.

⚠ Former names (the contract cut, [BREAK] of 2026-09-23): these fields used to be called tipo, tipoOrigen, observacion and nombrePublicado, with the values juridica/desconocido and rut/sufijo; none of them is emitted any more.

search_trademarks

The trademark tool: the collection with its filters, the batch, similarity and the opposition window, in one call. Identical to REST POST /trademarks/search; without the three fields that choose an engine (similarTo, similarToTrademark, oppositionWindow) it answers exactly what GET /trademarks answers for the same question — same filters, same envelope, same JSON — batch included. (Until 2026-09-24 that GET route had its own twin tool; it was withdrawn because it was the same question.)

Input (all fields optional; ⛔ an argument the tool does not declare is an input error, never silently ignored):

Field Type Notes
q string Free text, up to 256 characters. With it, results are ranked by relevance.
applicationNumber string[] The batch, up to 100, each number up to 17 characters. The other filters apply; see below.
niceClass integer[] Nice classes 1–45 (OR within the list), up to 45.
status string[] The published (derived) status, controlled catalog, up to 14.
signType string[] Sign type (controlled catalog), up to 14 codes of up to 20 characters.
holder · representative · taxId string[] Parties, OR within the list: one item per name, commas included ("PEPSICO, INC." is one; a comma never separates). Up to 20 values of 2 to 256 characters with at least one letter or digit; empty or [] ⇒ validation error; relation (current/former/all, default current) says whether the current or a former party counts.
denomination · registrationNumber string The denomination (text, up to 512 characters; q, 256) · the registration number (exact, up to 17).
goodsAndServices string ⭐ The goods-and-services text of the mark's classes, by words, accent- and case-insensitive (calzado finds «Calzados»); 2–200 characters. -word excludes, but at least one word that must appear is required: only stop words (de la) or only exclusions (-calzado) is a 400 validation_error.
filedFrom/filedTo · expiresFrom/expiresTo · registeredFrom/registeredTo · publishedFrom/publishedTo YYYY-MM-DD Ranges over the four dates.
milestoneFrom/milestoneTo · milestoneKind YYYY-MM-DD · string[] ⭐ Marks with a milestone in the range: expiry, opposition_close, renewal_window_open and/or grace_end (milestoneKind narrows; without it, any of the four). Each kind matches exactly the date the mark publishes (the close, derived or estimated; the renewal ones, never on old-regime or Madrid marks, nor outside registrada/esperando_renovacion). Inverted range = invalid_range; milestoneKind without a range or with an unknown kind = validation_error.
renewalStatus string[] ⭐ Only the marks whose renewalStatus (today, Chile time) is one of not_due · window_open · grace · lapsed · renewed_expiry_pending (OR). With the batch, "which marks of this list are in their grace period?" is ONE call, and every returned mark carries that same renewalStatus. null marks never match; any other value = validation_error.
liveOnly boolean Excludes marks in a terminal status: rechazada, denegada, desistida, abandonada, anulada, caducado, vencida.
partyRole string Limits taxId to one role: holder or representative. Without it, both. Requires taxId.
party string[] ⭐ Parties by id (from list_parties) and role: "<id>:holder" or "<id>:representative", OR within the list (up to 20). Same role meaning as list_party_trademarks; relation applies. Works for parties without a RUT.
includeApplicationNumber · excludeApplicationNumber string[] ⭐ The portfolio by hand (up to 500 each): ((party filters ∪ included) − excluded) ∩ every other filter. The party filters are party/holder/representative/taxId; an included mark still has to match the other filters; exclusion wins; without a party filter, the included marks are the portfolio. Numbers as INAPI writes them (991652209-2).
portfolio object ⭐ The portfolio as an object {party, refine, includeApplicationNumber, excludeApplicationNumber}: (((parties ∩ refine) ∪ included) − excluded) ∩ top-level filters. refine (the predicate filters of list_application_numbers) narrows only the parties' trademarks: an included mark is always in; exclusion wins; the top-level filters apply to everything. Up to 10,000 numbers per list. With it, a top-level party/includeApplicationNumber/excludeApplicationNumber is a validation_error. Also on list_application_numbers and get_renewal_calendar.
text object ⭐ Configurable text search (Algolia's vocabulary): {query, queryType, typoTolerance, minWordSizefor1Typo, minWordSizefor2Typos, exactOnSingleWordQuery, removeWordsIfNoResults, phonetic, semantic, getRankingInfo}. With every default it answers what q answers. Configured: the keyword block first (deterministic, complete, exact size in total or, with approximate hits, in meta.keywordTotal; exact → prefix → words → typos → sort), then the approximate hits of phonetic/semantic (only in meta.estimatedTotal). With q/denomination it is a validation_error. In list_application_numbers and portfolio.refine.text, only as a set (phonetic/semantic false).
similarTo · similarToTrademark · allClasses · oppositionWindow string · string · boolean · string Choose the engine instead of filtering; see "The three fields that choose an engine" below. allClasses only applies with similarToTrademark.
citedNorm · article · letter string Declared, but they do not filter trademarks: any of them makes the call a 400 validation_error that names the field and points to search_annotations (they filter annotations by the statute they cite).
facets string[] niceClass, status: counts in meta.facets.
sort string relevance, filedOn:desc… status/status:asc/status:desc orders by the published status.
limit · cursor · minScore Page (default 20, max 100) · opaque cursor · relevance floor.

Output: the contract's common envelope — { results: [...], limit, next, hasMore, total?, meta } —, identical to REST. meta always carries dataAsOf and coverageThrough; meta.estimatedTotal is a lower bound; meta.facets (niceClass, status) only when requested; meta.mode: "advanced" when a competing field displaced q. matchedFields[].field names the filter that surfaced each mark (denomination, applicationNumber, registrationNumber, holder, representative, taxId). The payload is the result's first text block (parse it as JSON).

The batch. With applicationNumber the response carries the marks in request order and meta.notFound with the numbers that do not exist; the other filters APPLY (a mark that exists but does not match simply does not appear), total is exact and there is no next page. limit, cursor, sort, minScore, q, denomination and facets with the batch are a validation_error. Over 100 is batch_too_large, never a silent truncation.

Milestones. Every mark in the result carries milestones[] ([] when it has none). With the batch, "which marks of this list have a milestone in the next 30 days, and when?" is ONE call: { "applicationNumber": [ …up to 100… ], "milestoneFrom": "2026-09-24", "milestoneTo": "2026-10-24" }. And "which marks close their opposition window next week?", another: { "milestoneKind": ["opposition_close"], "milestoneFrom": "2026-09-28", "milestoneTo": "2026-10-04" }. The core publishes the dates; who to alert and when is your application's decision.

⛔ An empty batch returns an empty batch, not the collection: building the list from a portfolio that happens to have no marks would otherwise return the first page of the whole corpus.

The three fields that choose an engine (mutually exclusive: asking for two is a validation_error): similarTo (free text, ranked by similarity over four channels: phonetic, lexical, semantic and distinctive element; matchedBy says which found it), similarToTrademark (an existing mark: excludes the base mark and narrows to its classes; allClasses: true widens) and oppositionWindow: "open" (marks with a publication date whose 30th business day is today or later and whose status is neither terminal nor registrada/concedida — the same term the trademark object publishes, which does not look at the status: a registrada mark with expired: false is NOT in the window; requires niceClass and one of q or applicationNumber). With similarity, applicationNumber is a validation_error; with the window it is ONE application number — the reference mark to rank against — not a batch: more than one is a validation_error. With either engine, sort only accepts relevance.

get_trademark

Fetch ONE trademark by its application number, with optional expansion. Identical to REST GET /trademarks/{applicationNumber}.

Input:

Field Type Notes
applicationNumber string The application number (required).
expand string[] classes, parties, annotations, encumbrances, norms. What is not asked for travels null (≠ []). norms (the norms cited in the annotations, with counts, dates and ids; no text) also requires legal:read.
dedupe boolean Default true: rows describing the SAME fact collapse into one, the others in duplicates[]. false returns every version. renewals is identical either way.

Output: the flat trademark object. An unknown number returns a structured not_found tool error, mirroring REST's 404: status: 404, title: "Not found" and the same detail (trademark 123 not found).

milestones[] ALWAYS travels: the mark's dated milestones, [] when there are none. Each says where it comes from (basis): source (an INAPI datum, with its verbatim field) or derived (a computed deadline, which always carries anchor, rule and legalBasis — no article, no derived milestone). Or estimated (the same shape plus estimateReasons: a deadline the law or the calendar does not let us affirm). legalBasis is a list of {norm, article, role}. Published: expiry (source, from expiresOn) and opposition_close (the close of the opposition window: Ley 19.039 arts. 5 and 11, Ley 19.880 art. 25; estimated when published on a non-business day or outside the holiday calendar's coverage; absent without a publication date), and, on national registrada or esperando_renovacion marks with an expiry from 2022-05-09, renewal_window_open and grace_end (expiry − 6 and + 6 months, Ley 19.039 art. 24, Decreto 82/2022 art. 27, Código Civil art. 48 and Ley 19.880 art. 25; grace_end moved to the next business day when it is a non-business day). All four filter with milestoneKind. ⭐ (44-14) And appeal_close: the last day to appeal a rejection (024) or a partial acceptance (444), the 15th business day after its publication in the Estado Diario (Ley 19.039 art. 17 bis B and 11, Ley 19.880 art. 25), published but not filterable; its auditable anchor and what is omitted and why come from get_deadlines. ⭐ (2026-10-10) And two more, not filterable either: opposition_response_close (to answer the transfer of an opposition: the 30th business day after it was notified, Ley 19.039 art. 9; estimated when the anchor is inferred) and grant_fee_close (to prove payment of the grant fees: the 60th business day after the acceptance was notified, Ley 19.039 art. 18 bis E; always estimated and with reservation in es/en/pt — the law counts from the day the resolution becomes final, INAPI from its notification: quote the date with that reservation). ⭐ (2026-10-26) And resolution: the date INAPI accepted or rejected the application (not a deadline; anchor.kind acceptance/rejection with the act, Ley 19.039 art. 13 with the role notification), absent when the application was never decided. ⭐ (2026-10-28) And designation_ceased: the date INAPI declared a Madrid designation ceased after a WIPO DEATH (cessation 1238/1240/1241/1242, not 1237; anchor = the cessation act; Madrid Protocol and its Common Regulations with the role extinction); present when the mark publishes status: caducado because of it; ⭐ (2026-11-04) and also, anchored on the DEATH (anchor.kind: "wipo_death"), when WIPO's cancellation reached INAPI during the procedure and INAPI granted afterwards. ⭐ (2026-10-28) Apart from the milestones, a few marks carry notices[] in get_trademark: facts checked by hand in the expediente and with WIPO; wipo_designation_ended says WIPO recorded that Chile is no longer designated, with the recording as the source (source.reference, recordedOn, madridMonitorUrl) and the expediente document as the backing (corroboration.documentIndex); wipo_refused_recording, that WIPO did not record INAPI's statement of grant because Chile is not designated. A notice does not decide status or milestones: the annotations do (Notices). The list of kinds, with an example, is in Milestones.

renewalStatus ALWAYS travels (on get_trademark and on every mark of search_trademarks): not_due | window_open | grace | lapsed | renewed_expiry_pending, evaluated on the date of the response (Chile time), or null when the rule does not apply — no expiry date, neither registrada nor esperando_renovacion, or filed via Madrid. It describes the term to request the renewal; in grace the registration is still in force and can still be renewed (it lapses for want of renewal only if the grace period ends without a renewal request: Ley 19.039 art. 24, 2nd paragraph; Decreto 82/2022 art. 27, 2nd paragraph). renewed_expiry_pending, before or after the expiry: the renewal is already granted and the published expiry is not yet updated; the mark then publishes neither renewal_window_open nor grace_end. The rule is in Milestones. It is also a filter of search_trademarks (renewalStatus), with the same rule and the same date as the field.

renewals[] (on get_trademark ≡ GET /trademarks/{applicationNumber}, always present, empty when there is none) are the renewal acts of the mark in chronological order, and each says what it is in kind: request (the renewal was requested: filed, or observed and pending; not yet granted on that date), grant (INAPI grants, accepts or records it) or outcome (INAPI resolves the request without granting it: abandonment, withdrawal, rejection…). lastRenewalOn is the date of the latest grant, and is omitted when there is none: a renewal requested and later abandoned or rejected does not count as a renewal.

procedureHistory[] (on get_trademark ≡ GET /trademarks/{applicationNumber}, always present, empty when there is none, identical with or without dedupe and with any expand) is INAPI's e-file index: the steps of the procedure table on the application's public record page and, mixed with them by date, the numbered annotations (filings such as a transfer or a renewal, and the resolutions on them). It comes without the Estado Diario rows and without the per-fact grouping of annotations[]. Each entry carries annotationId (the same id as in annotations[]), date, code (INAPI's code), inapiText (INAPI's wording of that code, the same as sectionName; null when INAPI never named it), reference (the resolution or writ number INAPI prints on the row, such as 2026/82718 or C/2026/123550; null when it prints none or the row has not been read since it is kept: today, most historical rows) and annotationNumber (INAPI's annotation number of the row; null on a step of the procedure table).

A filing INAPI lists both as a step and as a numbered annotation shows once, as the numbered annotation, paired one to one by date and code: two filings of the same code on the same day are two entries. When that entry has no reference, it takes the step's. A step INAPI returned once with its observation and once without it shows once, with the annotationId of the version with observation. Ordered by date and, within a day, by the place of each code in the flow of INAPI's procedure (the acceptance for processing 009 before the publication payment 439, the grant 158 before the certificate 1230), learnt from the e-file indices; codes without a known place come after, and the remaining ties follow the order the rows were recorded. INAPI's position is not stored, so some day may still come in a different order than at INAPI. An annotation stored without its number (part of the history lacks it) cannot be told from a step and comes with annotationNumber: null.

Two steps of the same day and code that INAPI tells apart only by their writ number (reference) are two entries. A step INAPI lists several times identically on the same day (no text or number to tell the copies apart) appears as many times as it is listed, and every copy carries the same annotationId, so that id is not unique within the list. Part of the history was stored before this was counted and shows such a step once until the application is read again. Madrid Protocol registrations have no «Expediente inicial (*)» entry: the e-file index prints it, but INAPI's public record page does not publish it, and it is not invented.

Error shape

Every tool error shares one shape: a tool result with isError: true whose text is an RFC 9457 application/problem+json — { "type", "title", "status", "detail", "code", "requestId" } — identical to the body the REST API returns for 4xx/5xx (REST≡MCP). A client can JSON.parse(text).code uniformly for any tool error.

⚠ requestId travels empty over MCP: the transport has no request id of its own, and the contract declares the field always present. Empty means "there is none", not "we forgot".

The codes you may see:

code Meaning
unauthorized Missing, invalid, revoked or expired key.
forbidden Valid key lacking the required scope (e.g. brands:read).
validation_error Invalid input (e.g. the "exactly one of businessDays or preset" rule).
not_found By-id lookup of an unknown (or another tenant's) resource. status 404, with the REST detail.
query_timeout 504: the query exceeded the time budget; retryable.
service_unavailable 503: dependency temporarily unavailable (the database out of resources, the watch store, search_trademarks with q when every channel fails —an empty result before—…); retryable.
coverage_not_declared 503 of list_holidays: the calendar declares no covered year; retryable.
application_terminal 422 of watch_application: the application exists but its status is terminal; it cannot be watched. The status travels in applicationStatus (and in detail).
watch_limit_reached 422 of watch_application: the org already has its plan's maximum of active watched applications; the maximum travels in limit.
watch_daily_limit_reached 429 of watch_application: the org used up the day's additions (Chile's day); carries limit and resetsAt (when it reopens), both at the top level of the body.
image_similarity_not_available 503 of find_similar_logos, a declared future operation: NOT retryable (unlike service_unavailable); it will not be available until an entry of the contract changelog announces it.
out_of_coverage Something outside the declared coverage was requested (a holiday-calendar range, a day's bulletin, a statute's text or a document version, the indicators). The error carries coverage with the real range.
internal_error Generic internal error — never leaks internal details.

Internal database or transport details are never included in the message.

⭐ status is the REST twin's: 401 for unauthorized, 403 for forbidden —with the same detail as REST, naming the missing scope (API key lacks required scope: brands:read)—, 400 for validation_error, 404 for not_found —with the same detail as REST (trademark 123 not found…)—, 422 for out_of_coverage. Until 2026-09-28 MCP said 400 for unauthorized and forbidden, and until 2026-09-29 also for not_found in 18 tools; the code did not change.

⭐ Since 2026-09-30 (ERR-1) every code has exactly one status, the same as REST, taken from a single table in the code. Until that date these came out as 400: query_timeout when the tool's total budget ran out (now 504), watchlist_too_large and unsafe_callback (now 422), idempotency_conflict (now 409) and the watch store's service_unavailable (now 503). list_holidays with no declared calendar said service_unavailable with 400; it is now coverage_not_declared with 503, as in REST. (watchlist_too_large, idempotency_conflict, concurrent_modification and watchlist_retired ceased to exist on 2026-10-04 with the watchlist tools, and no tool returns unsafe_callback any more.)

Connect your AI client

Generic Streamable-HTTP MCP client

Most MCP clients accept an HTTP server entry with custom headers. Point it at the endpoint and send your key in the X-API-Key header:

{
  "mcpServers": {
    "tarno": {
      "type": "streamable-http",
      "url": "https://mcp.tarno.cl",
      "headers": { "X-API-Key": "<API_KEY>" }
    }
  }
}

Claude Desktop (via the mcp-remote bridge)

Claude Desktop launches MCP servers over stdio, and its built-in custom connector UI only supports OAuth or no-auth servers — it has no field for a custom X-API-Key header. To reach a header-authenticated endpoint like Tarno, bridge to it with the mcp-remote helper, which runs locally over stdio and injects the header on each request.

Prerequisite: Node.js 18 or newer on your PATH (Claude Desktop invokes npx). Check with node --version.

1. Open your Claude Desktop config file (create it if it does not exist):

OS Path
macOS ~/Library/Application Support/Claude/claude_desktop_config.json
Windows %APPDATA%\Claude\claude_desktop_config.json

You can also reach it from Settings → Developer → Edit Config.

2. Add Tarno to mcpServers. Keep your key in an env entry (not inline in args) — this is the safe pattern and avoids a known Claude Desktop/Windows bug where a space inside an args value is mis-escaped when it invokes npx:

{
  "mcpServers": {
    "tarno": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote",
        "https://mcp.tarno.cl",
        "--header", "X-API-Key:${TARNO_API_KEY}"
      ],
      "env": {
        "TARNO_API_KEY": "<API_KEY>"
      }
    }
  }
}

Replace <API_KEY> with your operator-issued key. (Because an API key never contains spaces, the inline form "--header", "X-API-Key:<API_KEY>" also works — but the env pattern above keeps the key out of the args list and dodges the space bug.)

3. Restart Claude Desktop fully (quit and reopen). Tarno then appears as a connected tool source; you can ask Claude to search brands or fetch a brand's detail, and it will call search_trademarks / get_trademark for you.

Troubleshooting: if the server does not appear, confirm node --version works in a terminal, that the JSON is valid (no trailing commas), and that the key is a live, scoped key — an unauthorized/forbidden tool error means the key is missing, invalid, or lacks brands:read.

For a programmatic client (connect + callTool) using @modelcontextprotocol/sdk, see the code samples.

See also