@cyanheads/brapi-mcp-server

BrAPI v2.1 (the Breeding API) data from Breedbase, T3, Sweetpotatobase, and any BrAPI v2-compliant server. Search studies, germplasm, observations, genotypes, images, locations, and variants — result sets beyond the per-call cap spill into a DuckDB-backed dataframe workspace that agents on the same session can query with SQL or hand off by name, and connections to multiple upstream servers can be held open in parallel under named aliases. Runs as a stdio process, a local Streamable HTTP server, or the public hosted endpoint above.

URI-addressable mirrors of the curated tool surface for clients that prefer resources. All resources use the default connection — multi-server workflows route through tools.

• baseUrl and auth are optional — when omitted, resolved from BRAPI_<ALIAS>_* env vars, then the built-in registry, then BRAPI_DEFAULT_*, so credentials never enter the LLM context
• alias (default default, pattern ^[a-zA-Z0-9_-]+$) registers multiple concurrent connections in one session
• Auth is a tagged union: none / bearer / api_key / sgn (Breedbase /token exchange) / oauth2 (client-credentials)
• Typed errors: auth_token_exchange_failed, auth_no_access_token
• Returns the full orientation envelope (identity, capabilities, content counts, attribution) — one call fully orients the agent; re-fetch on demand via brapi_server_info

• alias optional (defaults to the connection registered under default); forceRefresh (default false) bypasses the cached capability profile
• Typed error: unknown_alias
• Returns the same orientation envelope shape as brapi_connect

• endpoint required — one of studies, germplasm, observations, variables, images, variants, locations
• Each entry carries name, type (string / integer / number / boolean / date / string[] / integer[]), description, and an example value
• Typed error: unknown_endpoint (response carries availableEndpoints as recovery data)
• Catalog reflects the BrAPI v2.1 spec; individual servers may implement subsets

• Filters: crop, trialTypes, seasons, locations, programs, trials, studyNames, active, plus extraFilters passthrough
• loadLimit caps in-context rows; beyond it the full result set materializes as a canvas dataframe (query with brapi_dataframe_query)
• distributions cover programName, studyType, seasons, locationName, commonCropName
• Typed errors: unknown_alias, all_filters_dropped (every supplied filter was unsupported by the active dialect)
• Response enrichment: totalCount, returnedCount, appliedFilters, refinementHint, notice, warnings

• studyDbId required; resolves program, trial, and location FKs inline
• Companion counts: observationCount, observationUnitCount, variableCount — omitted (with a warning) rather than reported as a server-wide total when the upstream can't scope a count to the study
• Typed errors: unknown_alias, study_not_found

• Filters: names, germplasmDbIds, germplasmPUIs, accessionNumbers, crops, synonyms, collections, genus, species, plus extraFilters
• text is a client-side substring match against germplasmName, accessionNumber, defaultDisplayName, and registered synonyms — combine with a server-side filter to narrow the upstream pull first
• loadLimit caps in-context rows; beyond it the full result set materializes as a canvas dataframe
• distributions cover commonCropName, genus, species, collection, countryOfOriginCode
• Typed errors: unknown_alias, all_filters_dropped

• germplasmDbId required; returns attributes (/germplasm/{id}/attributes) and direct parents (/germplasm/{id}/pedigree)
• Companions: studyCount, directParentCount, directDescendantCount (from /germplasm/{id}/progeny) — signals for pedigree depth and observation coverage
• Typed errors: unknown_alias, germplasm_not_found

• 1–20 root germplasmDbIds, walked concurrently; direction is ancestors (default), descendants, or both; maxDepth 1–10 (default 3)
• Deduplicates nodes and breaks cycles; a 1,000-node safety cap sets truncated when reached
• Traversal stats: depthReached, rootCount, leafCount, cycleCount, deadEndCount
• loadLimit bounds the inline nodes/edges preview; beyond it both sets spill to JOINable canvas dataframes (nodesDataframe, edgesDataframe)
• Typed error: unknown_alias

• Filters: variables, variableNames, variablePUIs, traitClasses, ontologies, studies, methods, scales, crop, plus extraFilters
• text ranks the full upstream union via OntologyResolver (PUI / name / synonym / trait-class match) and fills the in-context window with matches first, unmatched rows for context — unlike brapi_find_germplasm.text, unmatched rows aren't dropped
• ontologyCandidates in the response carries the ranked matches with their match source
• loadLimit caps in-context rows; beyond it the full result set materializes as a canvas dataframe
• Typed errors: unknown_alias, all_filters_dropped

• Filters: studies, germplasm, variables, observationUnits, observations, seasons, programs, trials, observationLevels, timestampFrom/timestampTo, plus extraFilters
• loadLimit caps in-context rows; beyond it the full result set materializes as a canvas dataframe
• distributions cover observationVariableName, studyName, germplasmName, observationLevel, season
• Typed errors: unknown_alias, all_filters_dropped

• Filters: images, observationUnits, observations, studies, imageFileNames, mimeTypes, descriptiveOntologyTerms, plus extraFilters
• Metadata only — fetch bytes via brapi_get_image
• loadLimit caps in-context rows; beyond it the full result set materializes as a canvas dataframe
• Typed errors: unknown_alias, all_filters_dropped

• 1–5 imageDbIds per call
• Prefers /images/{id}/imagecontent; falls back to the metadata imageURL — source on each payload names which path served it
• Per-image errors[] for failed fetches and warnings[] for loaded-but-suspect content (e.g. a non-image MIME from the imageURL fallback) — a partial batch never fails as a whole
• Typed errors: unknown_alias, images_unsupported (server doesn't advertise /images)

• Filters: locations, locationNames, countryCodes (ISO 3166-1 alpha-3), countryNames (free-form English, resolved client-side to alpha-3), locationTypes, abbreviations, plus extraFilters
• Optional post-fetch bbox (minLat/maxLat/minLon/maxLon, all four required to activate); retries once with axes swapped when the spec-correct [lon, lat] reading yields zero matches on a server that stores [lat, lon], and reports coordinateAxisOrder: "swapped"
• loadLimit caps in-context rows; beyond it the full result set materializes as a canvas dataframe
• Typed errors: unknown_alias, all_filters_dropped

• Filters: variantSets, variants, references, referenceName + start/end (1-based inclusive/exclusive genomic region), plus extraFilters
• loadLimit caps in-context rows; beyond it the full result set materializes as a canvas dataframe
• distributions cover variantType, referenceName, variantSetDbId
• Typed errors: unknown_alias, all_filters_dropped

• Requires at least one of variantSetDbId, variantSetDbIds, germplasmDbIds, callSetDbIds, or variantDbIds — unfiltered pulls are rejected
• Upstream pull bounded by BRAPI_GENOTYPE_CALLS_MAX_PULL (default 100,000, max 500,000) via the async POST /search/calls → GET /search/calls/{id} pattern
• loadLimit bounds the inline preview; the full collected set materializes as a dataframe when it exceeds loadLimit
• Typed errors: unknown_alias, no_filters, search_endpoint_disabled (dialect marks this server's search route as known-dead)

• dataframe optional — omit to list all, or name one for full detail (columns, row count, provenance)
• Provenance (originating tool, baseUrl, query, expiry) is present only for auto-registered df_* dataframes, not user-derived ones from registerAs
• Typed error: list_all_disabled_on_shared_http — listing without a name is refused on a shared HTTP deployment without per-caller auth, since every caller shares one tenant workspace

• sql must be a single SELECT — writes, DDL, COPY, PRAGMA, ATTACH, and file reads are rejected at a three-layer gate (single statement → SELECT only → plan-walk allowlist); system-catalog reads (information_schema, pg_catalog, sqlite_master, duckdb_*) are denied separately
• LIMIT/OFFSET is the paging idiom; projection and aggregation (COUNT, GROUP BY, AVG) summarize without materializing every row
• registerAs (letters/digits/underscore, ≤63 chars) persists the result as a new dataframe; preview (≤1000) and rowLimit bound what's returned inline
• Typed error: sql_rejected (carries the granular gate reason on data.gateReason)
• Response enrichment: truncated, shown, cap, notice

• Opt-in via BRAPI_CANVAS_DROP_ENABLED=true — omitted from tools/list otherwise
• Idempotent: returns dropped: false (not an error) for an unknown name
• Dataframes also expire via TTL when left unmanaged, so explicit drop is only needed to free workspace memory immediately

• Opt-in via BRAPI_EXPORT_DIR, stdio-only — omitted from tools/list under HTTP transport or when unset
• format is csv, parquet, or json; optional columns (thin projection) or sql (full SELECT, mutually exclusive with columns) materializes a temporary derived table first
• filename rejects path separators and .. segments; omit for a timestamp-suffixed default
• Typed errors: export_dir_unset, dataframe_not_found, invalid_filename, mutually_exclusive_projection

• studies required (≥1) — study-anchored to avoid full-table scans; optional variables/germplasm subsets
• shape: wide (one row per germplasm, one column per variable) or long (one row per observation); aggregate: mean (default), median, first, or all (forces long form even when shape:"wide")
• Wide-matrix column names are SQL-safe identifiers derived from observationVariableDbId; variableLegend maps them back to display names
• Typed errors: unknown_alias, all_filters_dropped, no_observation_path (neither /observations nor /observationunits returned data)
• Response enrichment: truncated, shown, cap, notice

• germplasmDbId required; discovers the germplasm's studies automatically (capped at 200) unless an explicit studyDbIds set is supplied, which skips discovery entirely
• Per-variable aggregates: n, mean, median, sd (omitted when n < 2 or non-numeric), min/max, studyCount, studyDbIds, seasons
• Typed errors: unknown_alias, germplasm_not_found

• variantSetDbId required; format is matrix-json (dataframe only), vcf-lite (VCF-subset text in vcf, plus dataframe), or plink (.ped/.map text, plus dataframe)
• maxCalls/maxColumns can only lower the deployment ceilings (BRAPI_GENOTYPE_CALLS_MAX_PULL, BRAPI_GENOTYPE_MATRIX_MAX_COLUMNS), never raise them
• variantColumnLegend maps SQL-safe column names back to original variant IDs; truncated names which ceiling fired when the matrix is incomplete
• Typed errors: unknown_alias, no_filters, search_endpoint_disabled

• studyDbId required; 1–5,000 observation rows; observationDbId presence on a row routes it to PUT, absence to POST
• mode: "preview" (default) validates only and returns a POST/PUT routing breakdown; mode: "apply" asks the caller to confirm via a multi-round-trip input request, then writes and verifies post-state with a cheap count probe
• force: true skips the confirmation round — only for out-of-band-authorized writes
• Additive only — no observation is ever destroyed
• Requires BRAPI_ENABLE_WRITES=true to register; scoped to brapi:write:observations
• Typed errors: unknown_alias, observations_unsupported, study_not_found, post_unsupported, put_unsupported, user_declined

• path (relative BrAPI route, e.g. /samples) + optional params; last-resort escape hatch for endpoints no curated tool covers
• Emits a suggestion when a curated tool exists for the same endpoint
• Spills to a canvas dataframe when the upstream advertises more rows than loadLimit and the result is a list shape; skipped when the caller drives paging via params.page/params.pageSize
• Typed errors: unknown_alias, cross_origin_path (a full URL was passed instead of a relative route)

• noun (e.g. observations, calls, germplasm) + body posted verbatim to POST /search/{noun}; async polling resolved transparently, kind reports sync or async
• Emits a suggestion when a curated tool covers the same noun
• Same spillover behavior as brapi_raw_get
• Typed errors: unknown_alias, search_endpoint_disabled

• No parameters — reads the cached capability profile for the default connection
• Typed error: unknown_alias

• No parameters — raw /serverinfo + /calls profile (server identity, crops, supported services) for the default connection
• Typed error: unknown_alias

• Same payload as brapi_get_study, addressed by URI on the default connection
• Typed errors: unknown_alias, study_not_found

• Same payload as brapi_get_germplasm, addressed by URI on the default connection
• Typed errors: unknown_alias, germplasm_not_found

• Same payload as brapi_describe_filters; listing the resource collection returns one entry per supported endpoint
• Typed error: unknown_endpoint

• Canonical /variables/{id} record (trait, scale, method, ontology) on the default connection — the single-record counterpart to brapi_find_variables
• Typed errors: unknown_alias, variable_not_found

• Arguments: studyDbId required; alias optional
• Six-step playbook — orient via brapi_get_study, enumerate variables, pull observation coverage, quantify missing data, flag numeric outliers (IQR), and an optional pedigree walk on the top-observed germplasm
• Ends in a structured markdown report with a recommended-next-steps section

• Arguments: germplasmDbIds (comma-separated) and traitName required; alias optional (run once per alias for multi-server analyses)
• Seven-step playbook — resolve the trait to one or more observation variables, discover contributing studies, harmonize units/scales/methods across studies, then per-germplasm × per-study and across-study summary statistics
• Ends in a markdown report that cites every dataframe handle or filter map used, for reproducibility

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.

BrAPI-specific:

• Dataframe spillover — find_* tools cap in-context rows at loadLimit and materialize larger unions (up to 50,000 rows) as DuckDB-backed df_<uuid> canvas dataframes, queryable via brapi_dataframe_query
• Dialect adaptation — five per-server-family adapters (spec / brapi-test / breedbase / cassavabase / bms) translate v2.1 plural filter keys to the singular form each family honors, drop known-broken filters, and escalate to POST /search/{noun} when GET would silently downcast
• Multi-server session with a built-in known-server registry — ServerRegistry holds live connections under named aliases; six public Breedbase/T3 endpoints resolve out-of-the-box with no env vars
• Capability-aware, rate-limited calls — CapabilityRegistry caches /serverinfo and guards every call against unsupported endpoints; a per-connection concurrency cap and exponential-backoff retry cover 429/5xx
• Tagged-union auth (none / bearer / api_key / sgn session-token exchange / oauth2 client-credentials), resolved per alias from env vars so credentials never enter the LLM context

Agent-friendly output:

• Provenance on every dataframe — brapi_dataframe_describe reports the originating tool, baseUrl, and query for every auto-registered df_<uuid> table
• Graceful partial failure — brapi_get_image returns per-item errors[] and warnings[] rows instead of failing the whole batch when some images can't be loaded
• Discriminated output contracts — brapi_submit_observations returns a mode-discriminated union (preview / apply); brapi_export_genotype_matrix and the raw-passthrough tools carry typed format/kind fields callers branch on instead of parsing strings
• Response-shaping guidance — find_* tools echo appliedFilters, a refinementHint when results are broad, and typed notice/warnings so agents can see exactly what was queried and why a response looks the way it does

When a find_* tool's upstream total exceeds loadLimit, the full union materializes as a canvas dataframe and the response carries an inline dataframe handle ({ tableName, rowCount, columns, createdAt, expiresAt, … }). Upstream column names that aren't SQL-safe identifiers are sanitized, and a columnLegend on the handle maps each renamed column back to its original key.

1. brapi_find_observations { studies: ["s-422"] }
   → first-page rows inline + dataframe.tableName = "df_<uuid>" (when totalCount > loadLimit)
2. brapi_dataframe_describe { dataframe: "df_<uuid>" }
   → schema + provenance (originating tool, baseUrl, query, expiry)
3. brapi_dataframe_query { sql: "SELECT germplasmName, value FROM df_<uuid> WHERE observationVariableDbId = 'V1' LIMIT 100" }
   → typed columns + bounded rows

Dataframe names are capability tokens, not row-level ACLs — anyone holding the name within the same session or tenant bucket (see Deployment shapes) can read its rows. They auto-expire via TTL (BRAPI_DATASET_TTL_SECONDS, default 24h); set BRAPI_CANVAS_DROP_ENABLED=true to expose brapi_dataframe_drop for explicit cleanup.

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

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

Add the following to your MCP client configuration file.

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

Or with npx (no Bun required):

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

Or with Docker:

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

No env vars are required — the six built-in aliases (bti-cassava, bti-sweetpotato, bti-breedbase-demo, t3-wheat, t3-oat, t3-barley) resolve out-of-the-box, and agents can connect to any other BrAPI v2 URL at runtime via brapi_connect. For credentialed servers, prefer env vars over agent input so passwords, tokens, and API keys stay out of the LLM context — see Per-alias credentials.

• Bun v1.4.0 or higher (or Node.js v24+).
• @duckdb/node-api is a required dependency — supported on Linux/macOS/Windows × x64 plus Linux/macOS arm64 (no Windows arm64, no Cloudflare Workers).

1. Clone the repository:

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

2. Navigate into the directory:

cd brapi-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env
# edit .env if you need credentials or non-default settings

Every variable is optional.

Per-alias overrides follow the BRAPI_<ALIAS>_* pattern. See .env.example for the full list of optional overrides.

brapi_connect resolves baseUrl and auth from env vars when the agent omits them — credentials never enter the LLM context. Four layers of precedence:

1. Explicit agent input — always wins.
2. Per-alias env vars — BRAPI_<ALIAS>_* (uppercased, hyphens → underscores: my-server → BRAPI_MY_SERVER_*).
3. Built-in known-server registry — see Built-in aliases.
4. Default env vars — BRAPI_DEFAULT_*, only when the alias differs from default. Not layered on top of a built-in URL — defaults belong to the default server.

Each alias carries one credential family — auth mode is derived from which fields are set:

Mixing families within an alias raises a ValidationError.

# .env — attach write credentials to the built-in 'bti-cassava' alias
BRAPI_BTI_CASSAVA_USERNAME=alice
BRAPI_BTI_CASSAVA_PASSWORD=...
# (BASE_URL omitted — built-in registry covers it)

# Static API key as alias 'prod'
BRAPI_PROD_BASE_URL=https://my-brapi.example.com/brapi/v2
BRAPI_PROD_API_KEY=...
BRAPI_PROD_API_KEY_HEADER=X-API-Key

Then the agent calls brapi_connect({ alias: 'bti-cassava' }) — no baseUrl, no auth, no secrets in the prompt.

The server ships with a curated registry of public BrAPI v2 endpoints. Each resolves out-of-the-box; the orientation envelope surfaces license, citation, and homepage in its attribution block under Creative Commons Attribution.

Set BRAPI_<ALIAS>_BASE_URL to repoint at a staging mirror or fork (env wins over the built-in URL — hyphens in the alias become underscores in the env var, so t3-wheat → BRAPI_T3_WHEAT_BASE_URL). Set BRAPI_<ALIAS>_USERNAME etc. to attach credentials on top of the built-in URL — each Breedbase instance has its own user table, so write access requires separate registration on each upstream. Use BRAPI_BUILTIN_ALIASES_DISABLED=bti-cassava,t3-wheat to strip specific entries.

Citation: all six built-ins reference Morales et al. 2022, "Breedbase: a digital ecosystem for modern plant breeding." G3 12(7): jkac078. doi:10.1093/g3journal/jkac078.

# Hot-reload dev (Bun runs TS directly)
bun --watch src/index.ts

# Production
bun run rebuild
bun run start            # transport via MCP_TRANSPORT_TYPE (stdio default)
bun run start:stdio      # or pin explicitly
bun run start:http

# Checks
bun run devcheck         # lint + format + typecheck + security + changelog sync
bun run test             # Vitest
bun run lint:mcp         # validate MCP definitions

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

Defaults to HTTP transport, stateful session mode (engages the mcp-session-id lifecycle — precondition for BRAPI_SESSION_ISOLATION=true; hijack protection requires layering MCP_AUTH_MODE=jwt|oauth on top), logs to /var/log/brapi-mcp-server. OTel peer deps are installed by default — --build-arg OTEL_ENABLED=false to omit.

Two stateful layers scope by tenant and, by default, by MCP session: connection state (registered aliases, exchanged upstream tokens) and dataframes (df_<uuid> tables — possession of the name grants full read/write/drop within its bucket, auto-expires in 24h by default, provenance recorded). brapi-mcp-server runs in three shapes that pick where those buckets end:

Stdio is always one session, so isolation is moot there. Clients on MCP protocol revision 2026-07-28 are session-less by every transport (no ctx.sessionId), so they always land in the shared tenant workspace regardless of BRAPI_SESSION_ISOLATION — only the per-user-credentials shape isolates them.

Belt-and-braces under shared trust: brapi_dataframe_describe requires an explicit dataframe name (no list-all enumeration) and brapi_dataframe_query rejects system-catalog reads, so a caller without a known df_<uuid> name can't fish through either surface even in the shared-workspace shape.

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 tenant-scoped storage — no console, no direct persistence access
• Register new tools in the tools array of createApp() in src/index.ts
• Wrap upstream calls: validate raw → normalize → return output schema; never fabricate missing fields

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

Apache-2.0 — see LICENSE for details.