@cyanheads/inaturalist-mcp-server

Overview

iNaturalist's index of 380M+ georeferenced citizen-science observations of plants, animals, and fungi. Search sightings by area, date, taxon, and annotation; read the community identification thread behind a record; chart when a taxon appears in a place; rank the species of an area; and check what a look-alike is most often confused with. Keyless and read-only, running as a stdio process or a local Streamable HTTP server.

Composes with servers covering institutional specimen records, botanical nomenclature, and geocoding — this one contributes the observation, identification-thread, and phenology layer.

Tools

Resources

Both resources mirror data also reachable through inaturalist_get_taxon and inaturalist_get_observation — useful for clients that don't surface MCP resources.

Related MCP server: musicbrainz-mcp-server

Capability reference

inaturalist_list_reference tool

• topic selects one table: controlled_terms, quality_grades, licenses, ranks, iconic_taxa, conservation_status_codes; source reports whether it came from iNaturalist or the published spec

• taxon_id applies only to controlled_terms and adds observed_usage — which annotation pairs identifiers have actually recorded for that taxon, with counts

• Every other tool's recovery hint routes here: an unrecognised filter value is not rejected upstream, it silently returns nothing

inaturalist_resolve_name tool

• type: taxon (name-prefix autocomplete) or place / project / user / any (scored cross-kind search); rank narrows taxa only; limit 1–30 (default 10)

• Taxon lookup matches a name prefix, not words inside a name — "monarch" hits where "monarch butterfly" misses

• A miss is a result: found: false with guidance naming why, rather than an error

• Each candidate carries kind and id — the identifier every other tool takes

inaturalist_find_places tool

• Exactly one of q (place-name prefix) or all four of nelat, nelng, swlat, swlng; neither or both fails as invalid_geography

• q returns places[]; the bounding box returns standard[] and community[] as separate lists

• Each place carries bbox, place_type, admin_level, ancestor_place_ids, location, and slug; boundary polygons are stripped, since one upstream response carries 247 KB of them

• per_page (1–30, default 10) binds the bounding-box arm only — the name-prefix endpoint publishes no page size, and its fixed page is disclosed through the truncation enrichment

inaturalist_search_observations tool

• An area is given in exactly one form — place_id, the lat+lng+radius triple in kilometres (radius ≤ 500), or the four-corner bounding box; partial or mixed fails as invalid_geography

• Filters: taxon_id, d1/d2, quality_grade, captive, term_id+term_value_id, iconic_taxa, hrank/lrank, csi, threatened/native/introduced/endemic, licensed/photo_licensed, and q+search_on

• Defaults to quality_grade: ["research"] and captive: false, echoed back as applied_filters on every call

• per_page 1–25 (default 10); page walks the first 10,000 results and cursor continues past it — passing both fails, and a cursor forces an id ordering, which is echoed

• include expands photos, annotations, sounds. identifications and comments are deliberately absent — one thread measures 28 KB, so the thread lives on inaturalist_get_observation

inaturalist_get_observation tool

• 1–10 ids per call, resolved in a single upstream request

• include defaults to ["identifications"]; comments, photos, annotations, and sounds are also available

• Partial success: ids that resolve return in observations, the rest in unresolved; the call fails as not_found only when nothing resolved

• Adds community_taxon and identification_disagreements_count on top of the projected search record

inaturalist_get_species_counts tool

• Distinct species for an area and period, ranked by observation_count — the "what lives here" answer without paging through individual sightings

• Same area forms and filters as the observation search; taxon_id narrows to a clade, such as the birds of a park

• per_page 1–50 (default 25), page for offset — upstream would serve 500 in one page, and the cap is sized by response bytes instead

• truncationCeiling carries the last count shown; the ranking is descending, so nothing left off the page exceeds it

inaturalist_get_histogram tool

• interval: month_of_year (default) and week_of_year fold every year into one seasonal curve; year, month, week, day, and hour bucket absolute dates, to which upstream applies its own default start date

• date_field: observed (default) or created

• taxon_id is optional — omit it to chart every taxon in the area

• Returns every bucket upstream produced in order, zeros included, plus their total — computed across every bucket upstream returned, even past the cap. day/hour over a wide date range can generate thousands of buckets, so the response is capped at 800, kept from the start of the range, with truncated/shown/cap disclosing the cut

inaturalist_get_leaderboard tool

• kind: observers (ranked by observations recorded, carrying species_count) or identifiers (identifications made); count_metric names what count measures

• per_page 1–250 (default 25), page for offset. Both endpoints rank only the top 500, so page × per_page past 500 fails as leaderboard_window_exceeded rather than returning a false zero-hit

• Takes the same area forms, taxon_id, d1/d2, and quality_grade as the observation search

inaturalist_get_similar_species tool

• The taxa a taxon_id is most often corrected from, ranked by misidentification_count — the field-identification check before committing to a look-alike

• An optional area, date range, quality_grade, and captive scope the confusion set to one region; omit them for the global set

• limit 1–50 (default 20), applied in-process — the endpoint publishes no page size and returns its whole set

inaturalist_get_taxon tool

• One taxon_id. Returns kind: "full" with the projected profile, or kind: "outline" listing each section and its byte size when the projection still overflows the budget

• Sections are summary, taxonomy, children, conservation, photos, encyclopedia; name them in sections to fetch a slice, and an unknown name fails as unknown_section

• The upstream record is 95 KB for a common species — per-country checklist membership is dropped, listed_taxa_count kept as a scalar, and ancestors, children, and conservation entries trimmed to their identifying fields

• A named section comes back whole at whatever size, so sum the outline's byte sizes before asking for several

inaturalist://taxa/{taxon_id} resource

• The projected taxon document as application/json, always whole — a resource read has no way to name sections, so use inaturalist_get_taxon when the outline path matters

• taxon_id comes from inaturalist_resolve_name; cached for six hours, matching the service's taxon TTL

inaturalist://observations/{observation_id} resource

• One observation with its identification thread expanded, as application/json

• observation_id comes from inaturalist_search_observations; cached for fifteen minutes, since a thread accrues identifications

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.

iNaturalist-specific:

• Keyless, read-only cover of the iNaturalist v1 API — observations, taxa, places, controlled terms, the similar-species graph, and the observer and identifier leaderboards

• Every response is projected in-process. Upstream accepts and ignores its own fields= parameter, so a two-record observation search arrives at 95 KB, a full upstream page of 200 at 4.3 MB, and a common taxon record at 95 KB before anything is trimmed

• Per-endpoint parameter allowlist — an unknown parameter name returns HTTP 200 and the entire global index, so nothing outside the allowlist is ever sent

• In-process rejection of every input upstream would silently widen: a lone lat, an unparseable d1, a term_value_id without its term_id, a page past the result window

• Self-paced outbound traffic with a per-UTC-day request budget, since the API returns no rate-limit headers to react to

Agent-friendly output:

• Applied defaults echoed on every call — quality_grade, captive, and the ordering a cursor forced — so an agent can see the filters that shaped its answer

• Zero-hit notices name the filter most likely responsible and the tool that decodes it, instead of an empty list

• Truncation disclosed unconditionally: truncated, shown, and cap on every path, plus a truncationCeiling where a descending ranking supports one

• Upstream free text — encyclopedia summaries, identification and comment bodies, place guesses, photo attributions — renders inside a markdown blockquote, marking it as third-party content rather than instruction

Licensing and attribution

The API is open; the records are not uniformly open.

• license_code is relayed verbatim and is nullable. A null means all rights reserved — it is never coerced to "", "unknown", or a default licence, and the rendered text spells the null case out in words.

• Photo attribution strings are relayed verbatim, never reformatted or shortened, and must be reproduced wherever the image is. A photo's own license_code is independent of its observation's.

• open is derived from the hosting domain: true for inaturalist-open-data.s3.amazonaws.com, false for anything else, because an unrecognised host is not evidence of an open licence. A licence change moves a photo between hosts, so the flag describes fetch time rather than a permanent property.

• Photos are linked, never proxied. No tool fetches image bytes or emits base64 image content — a URL with its attribution and licence is the whole contract.

• obscured: true marks a locality, not a sighting position. iNaturalist withholds true coordinates for threatened taxa, and the server never sends an Authorization header, so hidden coordinates stay hidden.

• Observer identity collapses to login. The upstream user object carries a real name, an ORCID, and counts; none of it is relayed.

Rate limits and response size

The published terms allow at most 100 requests per minute, ask clients to stay at or below 60, and ask for under 10,000 per day. No rate-limit headers come back, so pacing is entirely self-imposed: outbound requests start at least INATURALIST_MIN_REQUEST_INTERVAL_MS apart (1100 ms ≈ 54 per minute), at most INATURALIST_MAX_CONCURRENT_REQUESTS run in flight, and INATURALIST_DAILY_REQUEST_BUDGET bounds a UTC day. Exhausting the budget is a typed failure rather than a silent degradation.

Controlled terms (24 h), taxon profiles (6 h), places (6 h), the similar-species graph (6 h), and histograms (1 h) are cached in tenant-scoped storage. Observation search, species counts, leaderboards, and observation detail are never cached — freshness is what they are for.

Page-size maxima are sized by measured response bytes across structuredContent and the rendered text together, not by what upstream will serve:

Each default page fits the 24,000-byte budget a single document gets, and each full page fits 50,000. Nothing is unreachable at the lower caps — page and cursor reach the same records — so the smaller page costs one more call rather than any data.

Known limitations

• Upstream ignores its own fields= partial-response parameter, so every byte is fetched before being projected away. Projection saves the agent's context, not the network.

• total_results is an estimate over a live index. It drifts between calls seconds apart.

• place_type and admin_level have no published code table. The raw integers are relayed and display_name carries the meaning.

• A place crossing the antimeridian has a degenerate bounding box upstream. It is relayed as computed, not repaired.

• Obscured coordinates cannot be resolved, by design. A threatened-taxon record reports a locality with an accuracy radius in the tens of kilometres.

• A section named in inaturalist_get_taxon comes back whole however large it is — truncating a section the caller asked for by name is the failure the outline exists to prevent.

• The daily request counter is per process. A restart resets it, and two processes behind one egress IP do not share it.

• inaturalist_get_leaderboard can address only the top 500 entries, against the 10,000-result window on observation search.

Getting started

Add the following to your MCP client configuration file.

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

Or with npx (no Bun required):

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

Or with Docker:

{
  "mcpServers": {
    "inaturalist-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/inaturalist-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 (or Node.js ≥ 24.0.0).

• No API key required — the iNaturalist v1 API is keyless, and this server never authenticates.

• The published terms ask clients to identify themselves. A descriptive User-Agent with a contact URL is sent by default; keep one in any INATURALIST_USER_AGENT override.

Installation

1. Clone the repository:

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

2. Navigate into the directory:

cd inaturalist-mcp-server

3. Install dependencies:

bun install

4. Configure environment (optional):

cp .env.example .env
# edit .env to override defaults — no required vars

Configuration

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:stdio
  # or
  bun run start:http

• Run checks and tests:

  bun run devcheck  # Lint, format, typecheck, security
  bun run test      # Vitest test suite
  bun run lint:mcp  # Validate MCP definitions against spec

Docker

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

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/inaturalist-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

• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

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.