Start with the scientific task and source owner
Choose the workflow before the server. Specify the identifiers you have, the records or calculations you need, and the evidence your final answer must preserve. Compound identification, protein annotation, structure retrieval and materials screening require different contracts even when accessed through MCP.
Separate the upstream provider, integration maintainer and agent client. A third-party server is not the database it queries, and a documented client configuration is not an interoperability certification. This comparison reports source-described capabilities and proposes evaluation steps; it does not report executed installations or scientific testing.
For protocol background, read MCP in chemistry. Here, the decision is narrower: which tool surface answers your scientific question with acceptable provenance, permissions and operational requirements?
Classify the role before comparing features
The candidates occupy four roles:
- Scientific computation: RDKit MCP exposes toolkit functions; NovoMCP wraps calculations and configurable external services.
- Evidence retrieval: UniProt, RCSB, PubMed and Materials Project integrations retrieve records from named providers.
- Workflow-focused synthesis: PocketScout combines public evidence for known binding-site triage.
- Research interfaces: Benchling retrieves laboratory records; ChemCP supplies an embedded molecular viewer.
Do not confuse these roles with autonomy. Workflow prompts guide a client; they are not experimental protocols executed by the server. Likewise, Skill instructions are distinct from running software and do not install dependencies. Hosted endpoints add an operator boundary, while REST APIs offer another access surface—not additional scientific validation.
Compare the nine core candidates
Evidence for these nine candidates was observed on 2026-10-06 UTC. The table describes source documentation, not live service availability. Licence labels concern repository code only; upstream data and hosted-service terms need separate consideration.
| Resource | Domain and data source | Main tools or interface | Code licence / open-source evidence | Typical selection use |
|---|---|---|---|---|
| RDKit MCP Server (TandemAI) | Cheminformatics; RDKit library | Exposed RDKit functions, tool-listing utility, OpenAI CLI, evaluation suite | MIT licence text supplied | Access toolkit calculations after inspecting the actual inventory |
| UniProt MCP (cyanheads) | Protein evidence; UniProt REST | uniprot_search_proteins, entries, ID mapping, taxonomy, FASTA |
Apache-2.0 licence text supplied | Resolve protein identity and retain annotation provenance |
| RCSB MCP (cnyambura) | Structural biology; RCSB PDB API | get_pdb_entry, polymer metadata, downloads, custom queries |
README declares MIT; separate licence text unconfirmed, status unresolved | Inspect a known PDB entry and retrieve coordinate files |
| Materials Project MCP (benedictdebrah) | Materials data; Materials Project through mp_api |
search_materials, structures, electronic/phonon data, property and energy tools |
MIT licence text supplied | Search materials and inspect selected database properties |
| PubMed MCP (cyanheads) | Biomedical literature; PubMed, PMC, Europe PMC and fallbacks | Search, metadata, available full text, MeSH, citations | Apache-2.0 licence text supplied | Build a literature evidence set with provider and completeness labels |
| Benchling MCP (longevity-genie) | Laboratory records; Benchling workspace | Entry, sequence, project retrieval and entity search | MIT licence text supplied | Retrieve authorized research records, not control instruments |
| PocketScout MCP | Binding-site triage; UniProt, PDB, ChEMBL, PubMed, AlphaFold DB | Known ligand contacts, conservation, variants, pocket consolidation and prompts | MIT licence text supplied | Assemble a target briefing before deeper design work |
| NovoMCP | Computational chemistry; scientific packages and configured services | get_molecule_profile, searches, optional simulations, funnel auditing, MCP and REST |
Apache-2.0 wrappers/interfaces; BUSL-1.1 core, not uniformly open source | Coordinate calculations when required services are configured |
| ChemCP | Molecular visualization; browser-loaded RDKit.js | render_molecule from SMILES; interactive diagrams and descriptors |
Manifest declares ISC; separate licence text unconfirmed, status unresolved | Inspect structures inside an MCP Apps-capable host |
RDKit's ambition to expose every function is not demonstrated coverage. RCSB's organism-search tool returns search instructions, not retrieved matches. PocketScout's interpretation fields are software-generated context that should remain distinguishable from raw evidence.
Check credentials, runtime and interface requirements
The following conditions share the same 2026-10-06 UTC evidence date. Treat requirements as declared setup conditions, not tested compatibility.
| Candidate | Credentials | Documented running conditions and decision point |
|---|---|---|
| RDKit MCP | OpenAI key for the included CLI client | Python 3.10+ and pinned RDKit dependency; inspect tools before selecting operations |
| UniProt MCP | Upstream REST is keyless | STDIO, Streamable HTTP and hosted route; Bun prerequisite conflict requires revision-specific checking |
| RCSB MCP | No credential requirement specified in excerpts | Python 3.13+ metadata; standalone and Claude Desktop examples; review download destinations |
| Materials Project MCP | Materials Project API key | Local Python 3.12+ with uv, or running Docker environment |
| PubMed MCP | Optional NCBI key; contact email enables Unpaywall fallback | Bun 1.4+ or Node.js 24+; STDIO, Streamable HTTP and hosted route |
| Benchling MCP | API-enabled account, key and domain | Python 3.10+; transport documentation and entry-point default disagree |
| PocketScout MCP | Optional NCBI key | Python 3.11+ locally or documented hosted endpoint; metadata marks Alpha |
| NovoMCP | Local default omits authentication; hosted configuration differs | README specifies Python 3.10–3.12; optional scientific services gate capabilities |
| ChemCP | No server credential requirement specified | Node.js 18+; viewer requires MCP Apps support and CDN access |
UniProt's README minimum is Bun 1.3, whereas its manifest requires 1.4. Benchling's README describes a bare launch as HTTP, but metadata maps it to STDIO. Inspect the selected revision and launch route rather than guessing. ChemCP's web connector is experimental and content-security restrictions may block RDKit.js.
Match the task to the smallest useful tool surface
For cheminformatics, choose computation versus retrieval first. RDKit MCP is a toolkit interface; ChemCP is a viewer. The RDKit workflow guide helps frame the calculation stage.
For compound enrichment, PubChem MCP (cyanheads) describes identifier, formula and structure searches plus properties, safety records, bioactivity and conformers. PubChem MCP (PhelanShao) instead emphasizes get_pubchem_data, JSON/CSV/XYZ and structure-file downloads. Its fallback-generated coordinates must not be labelled retrieved PubChem conformers. Consult the PubChem integration guide when designing identifier handoffs.
For drug discovery, UniProt supplies protein identity, RCSB supplies structural records, and ChEMBL MCP (cyanheads) links compounds, targets, measured activity and assay provenance. PocketScout is a focused alternative for known-site triage; NovoMCP adds service-dependent computation. The protein structural-biology workflow provides a complementary stage-oriented view.
For literature research, select PubMed when query details, article metadata and available text are needed. For materials discovery, evaluate the third-party Materials Project server's explicit tools separately from the Materials Project API. The official mp-api project declares an optional MCP dependency and mpmcp entry point, but the excerpts do not establish its MCP tool inventory.
For lab automation, Benchling's documented scope is retrieval and search—not record editing or instrument control. For scientific data generally, provenance and completeness should outweigh the number of advertised tools. The chemistry API guide helps assess whether direct API access better fits a deterministic pipeline.
For the materials workflow, see AI for Materials Science.
Inspect input, output and permission contracts
Create a contract sheet for accepted identifiers, required fields, filters, units, pagination, files and error states. Check tools and resources separately: a tool-only client may not expose URI resources or an embedded viewer.
Preserve meaningful distinctions. PubChem separates ok, no_ghs_data and cid_not_found; absent classification is not evidence of safety. UniProt ID mapping distinguishes a running-job ticket from a completed-page continuation, and oversized entries can require section-specific follow-up. ChEMBL preserves missing potency as null and separates null_potency from ranked rows. Its optional DuckDB analysis has distinct preview and staged-data bounds, with table deletion separately opt-in.
Review network destinations, logs, storage and filesystem writes. Public source data does not make a confidential query harmless. RCSB downloads and PhelanShao caching require filesystem review; Benchling requires workspace access controls. NovoMCP's unauthenticated local default needs deliberate containment before network exposure. Approve only the operations required for the task.
Proposed example: a multi-server target evidence dossier
Proposed, untested workflow: prepare a dossier for human EGFR and three publicly identified compounds. Inputs are the gene symbol plus organism, compound names with available structure identifiers, measurement type IC50, and a bounded literature date range. No docking or generated coordinates are required.
- Resolve identity with UniProt. Retrieve the accession, organism, curation indicators and relevant cross-references. Stop if identity remains ambiguous; resume mapping tickets and continuations rather than submitting duplicate jobs.
- Resolve compounds with PubChem and ChEMBL. Save source-specific identifiers and structures. Stop for unresolved or conflicting identities; a shared name alone is insufficient reconciliation.
- Retrieve activity through ChEMBL. Resolve the target from the confirmed accession, collect compound–target measurements and inspect selected rows with
chembl_get_assay. Keep onestandard_type; retain measurement qualifiers (relations), units, organism and missing potency. - Inspect referenced structures through RCSB. Retrieve entry and polymer metadata for confirmed PDB cross-references. Decide whether coordinate downloads are necessary; do not treat structure presence as evidence of compound efficacy.
- Retrieve literature through PubMed. Save the effective query, identifiers, abstracts, correction/retraction links and full-text provider or unavailability labels.
The proposed output is an evidence table containing identifiers, assay context, structure references, citations, retrieval conditions and unresolved questions. PocketScout could be evaluated as an alternative triage branch, not assumed to reproduce this independently composed chain. Passing the dossier to DrugAgent would be a separate proposed downstream integration, not demonstrated interoperability.
Handle failures and scientific limitations explicitly
Evaluate a small non-sensitive sample against independently established expected records. Include ambiguous names, invalid IDs, missing fields, multi-page responses, unavailable full text and controlled provider failures. Preserve raw outputs and summaries; check whether partial failures and truncation survive summarization. Retry transient failures within a bounded policy, but correct invalid inputs instead of repeatedly sending them.
Scientific suitability is a separate gate. Similarity is not identity; one measurement type does not guarantee comparable assays. Retrieved coordinates, computed descriptors and service-dependent predictions have different evidential roles. PocketScout covers known sites, not novel pocket prediction, and its conservation method is local-context matching rather than full multiple sequence alignment. Materials property retrieval alone does not demonstrate experimental realizability. LLM-based evaluation judges and repository test instructions do not establish scientific validation.
Keep a compact decision checklist
- [ ] Define required inputs, outputs, units and source attribution.
- [ ] Confirm actual tool inventory and missing-data states.
- [ ] Record credentials, runtime, revision and client/transport pairing.
- [ ] Resolve documented runtime or transport conflicts before deployment.
- [ ] Bound network, workspace, storage, logging and file permissions.
- [ ] Evaluate pagination, partial failures and scientific interpretation separately.
- [ ] Keep code, database and hosted-service terms separate.
- [ ] Save remaining unknowns and reassess after interface changes.
Sources
Primary capability references: RDKit README, UniProt README, RCSB README, Materials Project MCP README, PubMed README, Benchling README, PocketScout README, NovoMCP README and ChemCP README.
Supplementary contracts: cyanheads PubChem, PhelanShao PubChem, ChEMBL, official Materials Project package metadata and official data-access documentation. NovoMCP's licensing distinction is supported by its top-level licence and core licence.