OpenBind MCP server

Public OpenBind experimental evidence and model/benchmark discovery for Bio MCP. Archetype: embedded, versioned data plus read-only Fragalysis REST. No API key.

Four operations, each with its mcp_ alias: openbind_search, openbind_execute, openbind_query_data, openbind_get_schema.

Use openbind_search to discover schemas. Example openbind_execute code:

return await api.get('/binding-events', {event_id: 'A71EV2A-x1776a'});

return await api.get('/affinity-fits', {used_in_analysis: 'true', limit: 5000});

Local pages default to 20 rows, maximum 5,000; follow next_offset until null. Live Fragalysis pages have a maximum of 100. Responses above 30KB use shared SQLite staging; take the handle and table names from _staging, then use the query/schema tools. workspace on execute opts into cross-server WorkspaceDO staging; pass that same workspace to query/schema calls (also returned in _staging.workspace). Snapshot queries have no runtime network dependency.

Worker bindings: OPENBIND_DATA_DO stores per-server staged data; WORKSPACE_DO connects the shared workspace Worker; CODE_MODE_LOADER runs sandboxed JavaScript. No secret environment variables are required.

Resources under resource://openbind/: catalog, query-notes, pagination-rate-limits, entity-model, dataset-map.

Scientific interpretation and provenance

The experimental protein is Coxsackievirus A16 2A protease, used as a surrogate for Enterovirus A71. Keep that distinction. Binding events, unique compounds and repeated kinetic fits have different denominators. Assayed CXSMILES and bound SMILES are separate fields. Missing affinities remain null. KD, kd and ka have different units and meanings. Structural non-observation does not establish biochemical inactivity; model confidence does not measure affinity.

The benchmark repository declares data CC0-1.0 and code Apache-2.0. Bundled source provenance records include that license declaration and commit-pinned SHA-256 hashes. External archives retain their own licensing metadata; inspect their records before reuse. Attribution: OpenBind Consortium; preprint DOI 10.64898/2026.08.27.747600. No participant-only data, hosted model inference or docking job submission is exposed.

Development and refresh

From the monorepo root:

pnpm --filter openbind-mcp-server dev
pnpm --filter openbind-mcp-server run type-check
pnpm exec vitest run servers/openbind-mcp-server/src
node scripts/verify-server.mjs openbind --local
node servers/openbind-mcp-server/test/structured-content-regression.mjs
# With the deployed shared workspace binding:
node servers/openbind-mcp-server/test/structured-content-regression.mjs https://openbind-mcp-server.quentincody.workers.dev/mcp --workspace
python3 scripts/openbind/import-release.py

The last command requires GitHub raw and Zenodo access, including the 124MB experimental ZIP. --cache supports previously downloaded originals and still verifies hashes. Refresh is deliberate: review release/target identity, licenses, artifact hashes and scientific labels before updating pins and fixtures.

Drift risk: high for evolving releases and Fragalysis; review every two weeks and on each public release. Search backing is a committed catalog. Gaps: August targets have discovery records rather than bundled rows; OB0 preview benchmark and bulky pose/MSA archives remain external. Consider a versioned OpenAPI spec when OpenBind publishes a stable supported API.

See project integration assessment for opportunities, scope limits and rollout validation.