Protein Mcp Server

Overview

Experimental (PDB) and predicted (AlphaFold) protein structures behind one surface. Search, fetch, align, and annotate structures and their ligands across RCSB, AlphaFold DB, 3D-Beacons, UniProt, InterPro, and Foldseek. No API key required.

Tools

Resources

Both resources are also reachable through protein_get_structure, for clients that don't surface resources.

Skills

Served over MCP as skill://protein-guide/SKILL.md; it also ships in plugin installs, the .mcpb bundle, the npm package, and the Docker image.

Capability reference

protein_search_structures _tool

• At least one of query, sequence (an mmseqs2 search; min_identity / max_evalue apply only with it), organism, method, max_resolution, or ligand (an exact component ID held as a bound copy); content_type is experimental, predicted, or all (default); up to 100 hits per page (default 25) via limit / start
• organism takes a scientific name, an NCBI common name, or a taxon ID, and matches that taxon plus every taxon under it; a word fragment matches nothing. method takes one of the 13 RCSB experimental methods in any letter case, or an alias (cryo-EM, cryoEM, EM, X-ray, xray, NMR for both NMR methods); anything else fails as unknown_method
• A sequence needs at least 10 residues; whitespace is stripped and case normalized, and an invalid one fails as invalid_sequence. Each hit names its source and an entry id you can pass to other tools, and carries organisms (a sequence hit's are the matched entity's). Sequence hits add identity, evalue, bitscore, and the aligned ranges, with queryLength at the response level
• Optional facets (method, organism, polymer_type, resolution, release_year, molecular_weight) add a breakdown capped at PROTEIN_FACET_BUCKET_CAP buckets per dimension; facets count the filtered set, so an active organism or method filter constrains its own facet

protein_get_structure _tool

• source: experimental (default) takes PDB entry IDs and the AF_* / MA_* computed-model IDs search returns (these come back as source: predicted); predicted takes UniProt accessions, isoform accessions (P04637-2), or AlphaFold DB entry IDs (AF-P04637-2-F1) for AlphaFold models with pLDDT/PAE; best_available takes the same identifiers (an entry ID reduces to its accession) and returns the PDB entry with the widest UniProt coverage (ties to the finest resolution), else the model with the highest confidence on the lDDT scale. Best-available records carry coverage, uniprotStart, and uniprotEnd; a pick can come from a non-PDB provider, and only a record with pdbId is a PDB entry. If the PDB check is unavailable, the pick ranks the federated models alone and notice says so
• A UniProt accession merged into another entry resolves to its successor's model with a notice; a demerged or deleted one lands in failed[] with its replacements
• IDs are trimmed and de-duplicated (case-insensitive); up to PROTEIN_MAX_BATCH_IDS are looked up per call (default 25) and the remainder comes back in ignored. One 50 s budget covers all upstream lookups in a call; under predicted and best_available, an ID it stops, or whose upstream lookup fails (an error status, an HTML page, a network failure), lands in failed[] with a reason to re-call it
• Unresolved IDs land in failed[]; when nothing resolves the call fails as upstream_unavailable (retryable, with data.failed) if every ID's lookup failed upstream or ran out of the 50 s budget, so none is known to lack a structure, and otherwise as all_failed (with data.failed); it fails mixed_id_types when the IDs don't fit source. Records include coordinateUrls for the files that exist, and organisms under experimental and predicted
• include_coords inlines coordinate text in 24,000-character windows: a single selected structure pages with coords_offset, passing back coordinatesNextOffset until it is absent; several selected files inline whole only when together they fit one window, otherwise an overflow outline lists each file's length to re-call with sections: [id]. coords_offset above 0 needs exactly one selected structure (coords_offset_needs_one_structure)

protein_find_similar _tool

• by: sequence runs a synchronous RCSB mmseqs2 search over experimental PDB entries only (protein_search_structures with a sequence also reaches computed models) from sequence, pdb_id, or uniprot, filtered by max_evalue / min_identity; by: structure runs an async Foldseek search from pdb_id (its mmCIF file) or uniprot (its AlphaFold DB model) against pdb100 + afdb50 unless databases says otherwise. uniprot takes a canonical or isoform accession, or a UniProtKB entry name such as HBA_HUMAN (resolved to its entry's accession), in both modes and is shape-checked before any request (invalid_accession). databases takes exactly these case-sensitive names: pdb100, afdb50, afdb-swissprot, afdb-proteome, bfvd_v3, cath50, mgnify_esm30, gmgcl_id, bfmd; any other fails schema validation (-32602), and a name Foldseek itself refuses fails as databases_rejected. A field only the other mode reads fails as mode_mismatched_field; up to 100 hits per page (default 25)
• Sequences follow the same rules as protein_search_structures, and a direct sequence wins over identifiers. A pdb_id sequence query searches one protein entity: the one holding chain (author chain, case-sensitive, by: sequence only, requires pdb_id), else the lowest-numbered protein entity of at least 10 residues; DNA and RNA entities are never searched. The response names it in queryEntityId, and a notice lists the entry's other protein entities. An unmatched chain fails as chain_not_found, and chain without pdb_id as chain_without_pdb_id
• Hits name their source (and database for structure hits), ranked by score; sequence hits add identity, evalue, bitscore, and the aligned ranges. Responses report totalCount, nextStart, and queryLength
• A merged uniprot accession answers for its successor with a notice; a demerged or deleted one fails as inactive_accession with its replacements. A by: structure coordinate file is limited to 16,000,000 characters, and a larger one fails as structure_too_large
• A structure hit carries Foldseek's target name and an id only when the target resolves to a PDB entry or UniProt accession; bfvd_v3, mgnify_esm30, gmgcl_id, and bfmd hits have none. Other structure-hit fields: chain (a -N suffix marks a symmetry copy), title, probability, organisms / taxId (the target's, where Foldseek classifies it), and the aligned spans queryStart, queryEnd, targetStart, targetEnd, targetLength. A sequence hit's organisms are the matched entity's
• A structure job still running at the poll budget returns status: computing with a ticketId; re-call with ticket_id to resume it. A job Foldseek accepted before the call's 50 s budget ran out comes back the same way. For a new search the poll window starts when the call does, so identifier lookups and the coordinate download come out of it. A multichain structure is searched one chain at a time via query (0-based, default 0) out of queryCount

protein_track_ligands _tool

• mode is find_ligand (query: a name or formula), structures_with_ligand (comp_id), or binding_site (pdb_id, optional comp_id); a missing mode input fails as missing_param. limit is 1–100 (default 25), and start pages the last two modes
• find_ligand returns component IDs with formula, weight, SMILES, and InChIKey, ranked by depositionCount, with totalCount and candidatesConsidered showing when the candidate pool was cut short; structures_with_ligand lists entries highest resolution first, each with title, method, and hostEntities (the polymer entities whose chains the depositor assigned the ligand to)
• Matching counts bound non-polymer copies only, not a residue inside a polymer chain or a sugar inside a branched glycan; depositionCount equals the totalCount of structures_with_ligand for the same component
• binding_site returns pocket residues with contact distances, numbered in both label (asymId, seqId) and author (authAsymId, authSeqId) namespaces, for ligand instances in a fixed order (ligand entity number, author chain, author residue number) so start / limit pages read each once; it works on experimental structures only

protein_compare_structures _tool

• 2 structures up to PROTEIN_MAX_COMPARE_STRUCTURES (default 10, max 25), each a pdb_id with an optional label-namespace chain; method is tm-align (default), fatcat-rigid, or fatcat-flexible; reference is first (default) or all_pairs; timeout_s (5–120) sets the per-pair poll budget
• Each pair row has a status (complete, computing, failed), tmScore, rmsd, alignedResidues, modeledResidues, and coverage; TM-score is normalized by a's length, so reversing a pair changes it. similarityScore and sequenceSimilarity appear when the upstream supplies them
• Structures past the cap return in ignored as { pdb_id, chain? }, not compared. A computing pair returns its job uuid; re-call with { a, b, uuid } in resume[] to poll it instead of resubmitting. A mismatched method or pair fails as resume_method_mismatch or resume_job_mismatch

protein_analyze_collection _tool

• group_by takes one of method, organism, polymer_type, resolution, release_year, molecular_weight, or two distinct ones for a cross-tab; scope with query, organism, method, max_resolution, and content_type (default experimental); organism and method take the same forms as in protein_search_structures, and an unknown method fails as unknown_method. interval sets a numeric bin width for resolution / molecular_weight or "year" for release_year
• Returns total and bucketsReturned; buckets carry label and count (numeric bins add rangeFrom / rangeTo), and every dimension reports missingValueCount
• bucket_limit (1–500, default PROTEIN_FACET_BUCKET_CAP) caps each dimension separately; truncated and notice say where results were cut

protein_get_annotations _tool

• A canonical uniprot accession, or a pdb_id resolved through its UniProt cross-reference (author chain picks one accession); include is features, domains, variants, or all (default); limit pages each class separately (1–200, default 50), with independent zero-based features_offset, variants_offset, and domains_offset (default 0)
• Returns UniProt features (domains, binding and active sites, PTMs, topology, processing products such as chains and signal peptides, mutagenesis, and more) and natural variants plus InterPro domains with GO terms; a PDB entry mapping to several accessions uses the lowest author chain and lists the rest in ambiguity
• An isoform accession fails as invalid_accession, naming the canonical one. An accession UniProt merged into another entry answers as its successor, with the merge named first in notice; a demerged or deleted one fails as inactive_accession with its replacements
• Each requested class carries featuresTotal, variantsTotal, or domainsTotal (InterPro's own membership count); a page with more records sets truncated and names the next offset in notice, and an InterPro walk that stops short of its count sets truncated and says how many memberships domains_offset can reach. One 50 s budget covers every upstream request in a call
• Failures are typed as missing_identifier, invalid_accession, accession_not_found, inactive_accession, no_uniprot_mapping, or chain_not_found

pdb://{entry_id} _resource

• entry_id is a PDB entry ID (e.g. 4HHB); returns application/json with title, methods, resolution, organisms, ligands, and polymerEntities carrying both authAsymIds and labelAsymIds
• Same data as protein_get_structure with source: experimental

af://{uniprot} _resource

• uniprot is a UniProt accession, an isoform accession (e.g. P04637-2), or an AlphaFold DB entry ID (e.g. AF-P69905-F1); returns application/json with meanPlddt, confidenceBuckets, cifUrl / pdbUrl / bcifUrl, and modelVersion, plus a notice when UniProt merged the accession into another entry (the successor's model answers)
• Failures are typed as invalid_accession, no_model, accession_not_found, or inactive_accession (with its replacements)
• Same data as protein_get_structure with source: predicted

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.

PDB / AlphaFold-specific:

• Search, fetch, and comparison treat experimental (PDB) and predicted (AlphaFold DB, ModelArchive, 3D-Beacons providers) structures the same way
• Keyless across RCSB, AlphaFold DB, 3D-Beacons, UniProt, InterPro, and Foldseek
• Collection analytics run on RCSB's facet engine and return bucket counts, not matching entries
• Foldseek and alignment jobs poll for up to PROTEIN_ASYNC_POLL_TIMEOUT_MS (default 30 s), then hand back a resume handle (ticketId, per-pair uuid) instead of blocking
• Chain namespaces stay separate: authAsymIds feed protein_get_annotations chain, labelAsymIds feed protein_compare_structures chain

Agent-friendly output:

• Provenance: every hit names its source (experimental / predicted) and the engine or database behind it, and protein_get_structure and protein_get_annotations carry an attribution block (see Upstream data licensing)
• Graceful partial failure: failed[] rows and per-pair status keep one bad ID or pair from failing the request
• Paging and advisories: paged tools return totalCount and nextStart, a start past the end is flagged in notice separately from zero matches, and every advisory joins into that one notice

Getting started

Public Hosted Instance

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

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

Self-Hosted / Local

Add the following to your MCP client configuration file. No API key is needed.

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

Or with npx (no Bun required):

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

Or with Docker:

{  "mcpServers": {    "protein-mcp-server": {      "type": "stdio",      "command": "docker",      "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/protein-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 v24+).
• No accounts or API keys: every upstream is public.

Installation

1. Clone the repository:

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

2. Navigate into the directory:

cd protein-mcp-server

3. Install dependencies:

bun install

4. Configure environment (optional):

cp .env.example .env# every variable has a working default

Configuration

Every variable is optional.

Each upstream's base URL can also be pointed at a mirror: RCSB_SEARCH_BASE_URL, RCSB_DATA_BASE_URL, RCSB_FILES_BASE_URL, RCSB_MODELS_BASE_URL, RCSB_ALIGNMENT_BASE_URL, BEACONS_BASE_URL, ALPHAFOLD_BASE_URL, MODELARCHIVE_BASE_URL, FOLDSEEK_BASE_URL, UNIPROT_BASE_URL, INTERPRO_BASE_URL. See .env.example for their defaults and the full list of optional overrides.

Running the server

Local development

• Build and run:

  sh

  # One-time buildbun run rebuild
  # Run the built serverbun run start:stdio# orbun run start:http

• Run checks and tests:

  sh

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

Docker

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

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/protein-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/AGENTS.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
• Export new definitions from the barrels in src/mcp-server/*/definitions/index.ts and register them in the createApp() arrays in src/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Upstream data licensing

Structure and annotation data comes from public databases, each under its own license. protein_get_structure and protein_get_annotations return an attribution block with the license, citation, and homepage of every source that contributed to that response. CC BY and CC BY-SA sources require attribution on redistribution; CC0 sources ask only for citation.

best_available models arrive through 3D-Beacons, so the block credits the provider that actually supplied the model; a provider with no curated entry gets a See provider terms fallback instead of a guessed license. InterPro and GO are credited separately, each only when it contributes. This table covers upstream data; the server's own code is licensed separately (see License).

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheckbun run test

License

Apache-2.0 — see LICENSE for details.