Overview
RDKit MCP Server (TandemAI) is a Python MCP server intended to make RDKit cheminformatics functions accessible through natural-language interactions with language models. It acts as an interface to the upstream RDKit library, rather than replacing that library or providing a standalone autonomous agent. The README states a goal of exposing every function in RDKit 2025.3.1; this is a coverage ambition, not evidence that all functions are available or validated.
In a typical workflow, a user starts the server and connects an MCP-capable language-model client. The client can then use the RDKit tools exposed by the server while responding to requests. The repository also includes an OpenAI-powered command-line client for experimentation and a utility for listing available tools. Inputs include natural-language requests handled by the client and server configuration supplied through an optional settings file. Outputs include RDKit tool results and client responses; the supplied excerpts do not specify individual tool schemas or supported molecular input formats.
The repository provides an evaluation suite using pydantic-evals to assess tool outputs and agent responses. Its LLMJudge checks whether an agent used the tools appropriately and produced accurate results, with options to inspect detailed output, select cases, and export JSON results. These facilities support local evaluation but do not establish scientific accuracy, comprehensive function coverage, or production readiness. The package metadata identifies Python and a pinned RDKit dependency, while the repository code is licensed under MIT. That license does not establish terms for external model services or other dependencies.
Key Features
- Exposes RDKit functions through the Model Context Protocol for use by MCP-capable language-model clients.
- Includes an OpenAI-powered command-line client for rapid experimentation.
- Supports an optional server settings file, with settings.example.yaml identified as the configuration reference.
- Provides a tool-listing utility to inspect which RDKit tools the server exposes.
- Includes a pydantic-evals suite with LLMJudge assessment, case filtering, detailed output, and JSON result export.
Use Cases
- Intended evaluation: assess whether natural-language requests can invoke the appropriate exposed RDKit tools in an existing MCP client workflow.
- Intended evaluation: use the bundled CLI client to prototype a conversational interface to RDKit before integrating another client.
- Intended evaluation: run selected repository evaluation cases and inspect tool outputs alongside agent responses for a proposed cheminformatics workflow.
How to Use
- Read the official README to understand the server/client split and its local installation procedure. Check the package metadata for declared Python and dependency requirements; these are not tested-compatibility claims.
- Follow the README installation instructions from the repository checkout. Review the referenced settings.example.yaml before starting the server if configuration changes are needed.
- Start the server using the documented run_server.py procedure, optionally supplying the settings file. Connect an MCP-capable client; the README links to the Claude Desktop quickstart as an example.
- Use the documented list_tools.py utility to inspect exposed tools before designing requests. For the bundled OpenAI client, configure your own credentials securely and follow the run_client.py instructions without publishing credentials.
- Install the documented evaluation dependencies and run the evaluation suite with the server active. Inspect detailed inputs and outputs, filter relevant cases, or export JSON results. Treat LLMJudge assessments as evaluation evidence to examine, not independent scientific validation.