AI-assisted molecular analysis with MCP

chematic-mcp is a local MCP server that lets an agent call named cheminformatics operations on SMILES. It uses stdio by default and can expose the same tools through an opt-in loopback HTTP listener. The agent supplies language and workflow logic; chematic returns structured chemical results. It is not a remote service, a general reaction planner, or a substitute for experimental or clinical judgement.

Start locally

Install the published binary:

cargo install chematic-mcp --version 1.1.0 --locked

Then point the MCP client at chematic-mcp:

{
  "mcpServers": {
    "chematic": {
      "command": "chematic-mcp",
      "args": []
    }
  }
}

By default the server uses newline-delimited JSON-RPC over stdio. chematic-mcp --transport streamable-http serves the same tools at http://127.0.0.1:3000/mcp instead. That adapter is self-hosted and loopback-only by default, and it has no authentication layer. There is no hosted endpoint and no SLA. See the chematic-mcp README for setup and the protocol reference for legacy and modern envelopes, HTTP headers, limits, and conformance evidence.

What an agent can call

The current server exposes 20 tools. The useful groups are:

Need	Tools
Parse and identify a molecule	parse_smiles, canonical_smiles, smiles_to_moljson, moljson_to_smiles
Calculate screening properties	calc_properties, lipinski_check, sa_score, pains_check, brenk_check, admet_profile, boiled_egg
Compare structures	ecfp4, pairwise tanimoto, smarts_match, find_mcs
Produce bounded follow-up representations	representation_router, molecule_context_pack, generate_3d, retrosynthesis
Resolve a public chemical name	name_to_smiles

There is no built-in indexed similarity_search or standardize MCP tool. For a caller-owned collection, the agent or application must iterate pairwise tanimoto calls or use a binding-level index, while retaining input IDs and all rejected records. retrosynthesis is a bounded one-step BRICS disconnection, not a synthetic route recommendation. generate_3d and the ADMET-related tools have their documented bounded or experimental scope.

Example: inspect one input

For a SMILES supplied by the user, an agent can call the property and alert tools independently, then cite the returned fields instead of inventing a chemical interpretation. A modern tools/call request looks like this:

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "_meta": {
      "io.modelcontextprotocol/protocolVersion": "2026-07-28",
      "io.modelcontextprotocol/clientInfo": {"name": "my-agent", "version": "1.0"},
      "io.modelcontextprotocol/clientCapabilities": {}
    },
    "name": "calc_properties",
    "arguments": {"smiles": "CC(=O)Nc1ccc(O)cc1"}
  }
}

Successful modern calls include machine-readable structuredContent. Invalid chemistry input instead returns a successful RPC with isError: true and a structured error code such as INVALID_SMILES; invalid argument shape is JSON-RPC -32602. Treat either outcome as data in an agent workflow rather than silently retrying with altered structures.

Local-data boundary

Nineteen tools operate locally. name_to_smiles is the exception: it sends the supplied chemical name to PubChem over HTTPS. Do not use that tool for proprietary names. The stdio server currently has bounded request and tool inputs, but it does not implement JSON-RPC cancellation; avoid presenting long-running work as interruptible. Exact limits and failure behaviour are in Error and resource limits.

For a Python-created comparison prompt, see examples/ai_agent_demo.py. For version-pinned chemistry compatibility and validation scope, start with Validation and the RDKit migration guide.