GBIF Biodiversity MCP Server

Public Hosted Server: https://gbif-biodiversity.caseyjhand.com/mcp

Overview

GBIF species taxonomy, occurrence records, datasets, and publishers. Match names to the backbone taxonomy, search 3.9B+ occurrence records with Darwin Core filters, and browse dataset and publisher metadata from any MCP client. Runs as a stdio process, a local Streamable HTTP server, or the public hosted endpoint above.

Tools

Resources

Both resources are also reachable via gbif_get_species and gbif_get_dataset; many MCP clients are tool-only and never surface resources.

Related MCP server: MCP Registry Search

Capability reference

gbif_match_species tool

• Fuzzy matching handles minor typos and vernacular names; set strict: true for exact-only matching

• Returns taxonKey — the backbone key required by gbif_search_occurrences, gbif_count_occurrences, and gbif_occurrence_facets

• Confidence score 0–100; below 80 warrants review

• Full classification hierarchy with keys at each rank: kingdom, phylum, class, order, family, genus, species

• matchType NONE indicates no usable match — try removing strict mode or broadening the name

• Resolves synonyms: always returns the accepted backbone key regardless of which name form was queried; matchedTaxonKey carries the synonym's own key when the two differ

• kingdom disambiguates a name that appears in more than one kingdom, and is rejected when supplied blank: GBIF drops a blank one and matches against the whole backbone — Parus major resolves to taxon 9705453 with kingdom= exactly as it does with no kingdom at all, where Plantae resolves to 9711704 — so the undisambiguated answer would come back looking like a disambiguated one. Omit the field to match against the whole backbone

gbif_bulk_match_species tool

• The batch counterpart to gbif_match_species — built for checklist, inventory, and species-list workflows that would otherwise need one round trip per name

• Returns one result per input name, in input order; each carries taxonKey, matchType, and confidence

• Per-name isolation: an unmatched name yields matchType NONE and a per-name lookup failure yields matchType ERROR with the message and, when the failure was classified, a reason — neither sinks the rest of the batch

• Same synonym resolution as gbif_match_species: taxonKey is the accepted taxon, matchedTaxonKey the synonym it was queried under

• strict: true requires an exact match for every name; common names are not supported (use gbif_search_species)

gbif_get_species tool

• Full classification, authorship string, and vernacular (English) name when available

• taxonomicStatus: ACCEPTED, SYNONYM, DOUBTFUL — when SYNONYM, acceptedKey and accepted identify the current name

• numDescendants and numOccurrences for scope at a glance

• extinct field present only when explicitly flagged — not false on unlabeled taxa

• publishedIn carries the original description citation when available

gbif_search_species tool

• Accepts name fragments matching scientific and vernacular names

• Filter by rank, kingdom, family, or genus to scope browsing

• kingdom, family, and genus are given as names and resolved to a backbone key before the search runs, since /species/search scopes by key alone. The narrowest one supplied is what scopes — the three nest, and GBIF combines two keys with OR rather than AND. kingdom supplied beside family or genus disambiguates that name instead of scoping on its own: Prunella names both a bird genus and a plant genus and resolves to neither without it

• Names are matched exactly and capitalized as GBIF writes them, so paridae and Paridaee fail as unresolved_taxon_scope rather than being ignored. An alternative family name lands on the taxon it is a synonym of — Compositae scopes to Asteraceae. A family and genus in different lineages fail as conflicting_taxon_scope. The scope actually applied comes back in the enrichment

• isExtinct filter for extinct vs. extant taxa

• Scope to a specific checklist dataset with datasetKey — omit the field for the GBIF backbone. GBIF reads a higher-taxon key inside the checklist that key belongs to, so pairing datasetKey with a kingdom, family, or genus matches nothing unless the checklist is the backbone; the empty-result notice says so

• q, kingdom, family, genus, and datasetKey are rejected when supplied blank rather than dropped: a blank datasetKey returns the unfiltered backbone result, and q= returns the whole 46,623,754-name index where q= with a space returns nothing. Omit a filter to leave it off

• Paginated — limit up to 1000, use offset to walk through large groups

gbif_get_species_classification tool

• Root-first from kingdom down to the immediate parent of the queried taxon (kingdom → phylum → class → … → parent)

• The queried taxon itself is not included — use gbif_get_species for its own record

• Each entry: rank, canonical name, scientific name, taxon key

• Useful for building taxonomic trees or placing an unfamiliar taxon in context without manual backbone navigation

gbif_get_species_children tool

• Genera within a family, species within a genus, subspecies within a species

• Each child: key, name, rank, taxonomic status, common name, occurrence count, descendant count

• Paginated — limit up to 1000, iterate with offset for large groups like Coleoptera

gbif_search_occurrences tool

• Use taxonKey from gbif_match_species for reliable results — resolves synonyms automatically; scientificName filter does not

• Geographic filters: country (ISO 3166-1 alpha-2, uppercase), stateProvince, bounding box (decimalLatitude/decimalLongitude ranges as "min,max"), or WKT polygon (geometry)

• publishingCountry (ISO 3166-1 alpha-2, uppercase) is the country of the publishing organization, not of the observation — a different question from country, and the two disagree on most records: of 60,290,950 records observed in GB, 1,548,928 were published by US organizations

• Both country filters take the uppercase two-letter form only. GBIF parses a lowercase or alpha-3 code and then matches the verbatim string, so gb and USA return zero records instead of an error; the schema pattern turns that silent wrong answer into a validation error naming the accepted form. A two-letter code GBIF does not assign (XX) is rejected upstream by name and surfaces as invalid_filter

• stateProvince is matched verbatim — exact and case-sensitive, with no controlled vocabulary behind it. Take a value from a STATE_PROVINCE facet rather than guessing; an unmatched value returns zero records instead of an error, and the enrichment notice says so when it happens

• Every filter is checked whenever it is supplied, not only when it holds a value. GBIF ignores a parameter it is given with no value and answers 200 with the unfiltered scope, so a blank filter used to widen the query silently: stateProvince: "" returns all 60,290,950 records of a taxonKey=212 + country=GB scope where England returns 47,672,439 — the same total, and the same first record, as a call that sends no stateProvince at all. A blank or whitespace-only value now fails as invalid_filter instead. Omit a field to leave that filter off

• Temporal filters: year as single year or range, month (1–12) for seasonal queries

• basisOfRecord enum: HUMAN_OBSERVATION, PRESERVED_SPECIMEN, MACHINE_OBSERVATION, and more

• hasCoordinate to require or exclude georeferenced records

• occurrenceStatus — PRESENT (default), ABSENT, or ANY. GBIF indexes absence records (a survey that looked for the taxon and did not find it) alongside sightings; the default excludes them and the enrichment says so on every call

• iucnRedListCategory — CR, EN, VU, NT, LC, DD, EX, EW, CD

• Output per record adds taxonomicStatus, eventTime (with UTC offset), occurrenceStatus, and iucnRedListCategory

• Pagination capped at offset+limit = 100,001, the deepest page GBIF serves. The API has no cursor or scroll, so a larger result set is covered by partitioning it — facet by DATASET_KEY with gbif_occurrence_facets, then search each datasetKey on its own. Retrieving a set in one piece is not something this server can do: that needs GBIF's Download API with a GBIF.org account, or GBIF's monthly snapshot on AWS Open Data

gbif_count_occurrences tool

• Backed by /occurrence/search at limit=0 — no record payload, and the same endpoint gbif_search_occurrences queries, so the two agree on the same question. GBIF's dedicated /occurrence/count endpoint takes a closed parameter set that rejects occurrenceStatus and iucnRedListCategory outright

• Supported filters: taxonKey, country, publishingCountry, stateProvince, isGeoreferenced, datasetKey, year, occurrenceStatus, iucnRedListCategory. country and publishingCountry take the uppercase alpha-2 form only, and a blank filter is rejected rather than dropped, both for the reasons given under gbif_search_occurrences

• Counts sightings only by default, matching gbif_search_occurrences. For absence-heavy taxa the unfiltered figure is a different question entirely — Radicipes gracilis has 2,351,582 indexed records of which 79 are presences

• Use to assess result set size before deciding whether to paginate a full search. A count above 100,001 means paging cannot reach the end of it, and the enrichment notice says so and names the partition route

gbif_get_occurrence tool

• Complete Darwin Core record — all coordinate fields, administrative geography (continent, country, state/province, locality), dates

• occurrenceID, full classification (class/classKey), GADM administrative units (levels 0–3, each with a stable GID and name), and source identifiers

• occurrenceStatus — check it before reading the record as a sighting; ABSENT means a survey looked and found nothing, and the record still carries coordinates, a date, and a recorder

• taxonomicStatus, eventTime (with UTC offset), and iucnRedListCategory

• Collections metadata: institution code, collection code, catalog number

• Collector and identifier names, individual count, sex, life stage

• Associated media (images, audio, video) with URLs and license

• GBIF data quality issue flags for provenance assessment

gbif_occurrence_facets tool

• Facets: COUNTRY, STATE_PROVINCE, YEAR, BASIS_OF_RECORD, DATASET_KEY, KINGDOM_KEY, PHYLUM_KEY, CLASS_KEY, ORDER_KEY, FAMILY_KEY, GENUS_KEY, SPECIES_KEY, PUBLISHING_COUNTRY, MONTH, OCCURRENCE_STATUS, IUCN_RED_LIST_CATEGORY

• Scope with taxonKey, country, publishingCountry, stateProvince, year, geometry, basisOfRecord, datasetKey, occurrenceStatus, or iucnRedListCategory filters — so a COUNTRY, PUBLISHING_COUNTRY, or STATE_PROVINCE bucket can be passed straight back to drill into it. country and publishingCountry take the uppercase alpha-2 form only, and a blank filter is rejected rather than dropped, both for the reasons given under gbif_search_occurrences

• Aggregates sightings only by default, matching the search and count tools. To measure the presence/absence split itself, pass facet: OCCURRENCE_STATUS with occurrenceStatus: ANY

• Returns one page of values ranked by count descending — up to facetLimit (max 100), the top ones only while facetOffset is 0 — with no record payloads

• DATASET_KEY is the basis for splitting a result set too large for gbif_search_occurrences to page: every occurrence carries exactly one datasetKey, so its buckets sum to the scope's full total, and it has the cardinality to cut a large scope into pageable pieces. BASIS_OF_RECORD and PUBLISHING_COUNTRY are gap-free too and both have a matching filter on the occurrence tools, so either can drive a further split of a bucket still over the cap — but on that scope they return 9 and 41 buckets against DATASET_KEY's 550, too coarse for the first cut. A dimension a record can lack drops that record: a 60,290,950-record scope faceted by YEAR sums to 59,407,400, leaving 883,550 undated records out, and MONTH, STATE_PROVINCE, and SPECIES_KEY behave the same way — STATE_PROVINCE included, even though the occurrence tools can filter on it

• The scope filters here are narrower than gbif_search_occurrences accepts — no scientificName, month, bounding box, hasCoordinate, isInCluster, or coordinateUncertaintyInMeters. Buckets sum to a search's total only when both calls carry the same filters; re-apply the rest on each per-datasetKey search

• Page past the first facetLimit with facetOffset (advance by facetLimit per page) to walk high-cardinality facets like DATASET_KEY; enrichment echoes the applied facetOffset and sets moreValuesLikely when a full page suggests more values remain

• Core tool for distribution analysis ("which countries have the most records?") and trend queries ("how has observation volume changed since 2010?")

gbif_search_datasets tool

• Filters: free-text query, dataset type (OCCURRENCE, CHECKLIST, METADATA, SAMPLING_EVENT), publishingCountry (ISO 3166-1 alpha-2, uppercase), publishing and hosting organization UUIDs (lowercase)

• publishingCountry takes the uppercase two-letter form only, for the same reason the occurrence filters do: /dataset/search matches the verbatim stored code, so gb and GBR return zero datasets where GB returns 2,416. A two-letter code GBIF does not assign (XX) is rejected upstream by name and surfaces as invalid_filter

• Returns title, type, description, license, DOI, and record count. recordCount spans every occurrenceStatus, absences included — gbif_count_occurrences with the same key counts sightings only by default, so the two figures differ by design

• The description is a 300-character preview — descriptionTruncated flags when it was shortened, and gbif_get_dataset returns the full text

• An organization key from gbif_search_publishers chains in two ways, and they answer different questions. publishingOrg matches the organization whose data it is; hostingOrg matches the organization whose installation serves it. publishingOrg is almost always the one meant — Butterfly Conservation (0d72dd7f-6f05-46af-85c2-8b6e77ce5534) publishes 3 datasets and hosts none, while the National Biodiversity Network (07f617d0-c688-11d8-bf62-b8a03c50a862) hosts 984, those 3 among them, and publishes 1

• Supplying both organization filters intersects them rather than combining them, so the same key in both fields returns only what that organization published and serves. An empty result under both is called out by name in the response notice

• Both take the lowercase 8-4-4-4-12 hex UUID form and are checked before the request goes out, because GBIF answers the two silent ways of getting them wrong with a wrong result rather than an error: an empty organization value returns the whole 123,527-dataset index, and these two filters are matched case-sensitively, so an upper-cased key returns 0 where the same key lowercased matches. Both fail locally as invalid_filter, as a malformed value does

• The free-text q is rejected blank on the same ground, and the route makes the case plainly: q= returns all 123,527 datasets while q= with a single space returns none. One character between the whole index and an empty page is not a distinction to leave to chance — omit the field to browse without a term

• Paginated — limit up to 1000

gbif_get_dataset tool

• Full description, citation text (for academic reference), license, DOI

• Contacts with role, name, organization, and email — capped by contactLimit (default 10, max 100); contactsTotal/contactsReturned report the full count

• Temporal and geographic coverage ranges when the publisher declares them

• recordCount — the indexed occurrence total, matching what gbif_search_datasets reports, for every dataset type (a CHECKLIST reports 0). It spans every occurrenceStatus, absences included; gbif_count_occurrences with the same key counts sightings only by default, so the two figures differ by design

• numConstituents for aggregate datasets (e.g. iNaturalist, eBird)

• Use after gbif_search_datasets or when an occurrence record's datasetKey needs provenance detail

gbif_search_publishers tool

• Filter by name fragment or country

• country here is unconstrained on purpose: the registry endpoint matches the parsed country rather than the verbatim string, so gb, GBR, and GB all return the same 223 organizations. The uppercase-only rule the occurrence tools and gbif_search_datasets carry exists to close a silent zero that does not occur on this route. One value is rejected — an empty country, which the registry answers with all 3,561 registered organizations rather than an error; omit the field to search every country. Whitespace alone is left to the registry, which answers it with 400 Cannot parse … into a known Country, an error naming the value

• q is rejected blank in both forms, because the registry drops either one and answers with all 3,561 organizations against 460 for museum; omit the field to browse without a name term

• Returns organization key, title, and country. The key chains into gbif_search_datasets as publishingOrg for the datasets the organization published, or as hostingOrg for the ones its own installation serves. publishingOrg is the usual chain: of the first 25 GB organizations the registry lists, all 25 host no datasets while 13 publish one or two

• Paginated — limit up to 1000

gbif://species/{taxonKey} resource

• Taxon record as application/json — classification, authorship, synonymy status, vernacular name, descendant count

• taxonKey comes from gbif_match_species or gbif_search_species

• extinct present only when explicitly flagged, absent otherwise, matching gbif_get_species

• A key that resolves to no backbone taxon fails as not_found; a non-numeric key segment is rejected locally as a validation error, and a numeric one GBIF cannot parse comes back as invalid_filter

gbif://dataset/{datasetKey} resource

• Dataset record as application/json — title, description, citation, license, DOI, temporal and geographic coverage

• Contacts fixed at 10 (no configurable limit, unlike gbif_get_dataset's contactLimit); contactsTotal/contactsReturned report the full count

• recordCount spans every occurrenceStatus, absences included

• datasetKey comes from gbif_search_datasets or an occurrence record's datasetKey field; a key that resolves to no dataset fails as not_found

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

GBIF-specific:

• Full GBIF REST API v1 coverage: species taxonomy, occurrences, datasets, and publishers

• gbif_match_species as the entry point — resolves synonyms to backbone taxon keys used throughout

• Occurrence pagination guarded at GBIF's own offset+limit = 100,001 boundary — an over-cap request fails locally with a recovery hint naming the DATASET_KEY partition technique, instead of spending the retry budget on a deterministic upstream rejection, and a match larger than the cap says so on the first page rather than after hundreds of them

• WKT polygon geometry support for geographic occurrence queries

• Darwin Core field mapping with explicit provenance on sparse upstream fields

Agent-friendly output:

• gbif_match_species is the mandatory first step — all downstream tools document which key they expect

• Graceful sparse-field handling — optional fields absent from the API response are omitted rather than null-filled

• Discriminated error contracts with typed reasons, structured recovery hints, and when documentation per tool

Getting started

Public Hosted Instance

A public instance is available at https://gbif-biodiversity.caseyjhand.com/mcp — no installation required. Point any MCP client at it via Streamable HTTP:

{
  "mcpServers": {
    "gbif-biodiversity-mcp-server": {
      "type": "streamable-http",
      "url": "https://gbif-biodiversity.caseyjhand.com/mcp"
    }
  }
}

Self-Hosted / Local

Add the following to your MCP client configuration file.

{
  "mcpServers": {
    "gbif-biodiversity-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/gbif-biodiversity-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with npx (no Bun required):

{
  "mcpServers": {
    "gbif-biodiversity-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@cyanheads/gbif-biodiversity-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with Docker:

{
  "mcpServers": {
    "gbif-biodiversity-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/gbif-biodiversity-mcp-server:latest"]
    }
  }
}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http
# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher.

• No credentials — the GBIF endpoints this server calls are public.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/gbif-biodiversity-mcp-server.git

2. Navigate into the directory:

cd gbif-biodiversity-mcp-server

3. Install dependencies:

bun install

Configuration

All configuration is validated at startup via Zod schemas in src/config/server-config.ts. Key environment variables:

See .env.example for the full list of optional overrides.

Running the server

Local development

• Build and run the production version:

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:http
  # or
  bun run start:stdio

• Run checks and tests:

  bun run devcheck  # Lints, formats, type-checks, and more
  bun run test      # Runs the test suite

Docker

docker build -t gbif-biodiversity-mcp-server .
docker run --rm -p 3010:3010 gbif-biodiversity-mcp-server

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/gbif-biodiversity-mcp-server. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic

• Use ctx.log for logging, ctx.state for storage

• Register new tools and resources in the createApp() arrays

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.