sds-mcp-server

An MCP server that lets Claude (or any MCP client) search for Safety Data Sheets (SDS) by chemical or product name, using chemicalsafety.com's SDS search.

Tools

search_sds_by_name

Search for SDS records by chemical or product name.

Argument	Type	Default	Description
chemical_name	string	—	Chemical or product name to search for, e.g. "Acetone"
is_contains	bool	false	Match names containing the text instead of requiring an exact match (useful for misspelled or partial names)
limit	int	25	Max results to return (common chemicals can return 100+ matches)

Returns a list of records with id, product_name, manufacturer, cas_number, msds_number, revision_date, has_sds, and sds_url (a direct link to the SDS document, usually a PDF).

Synonym matching (IncludeSynonyms) is intentionally not exposed — it's the source endpoint's slowest query mode and was the main cause of Claude Science's local-connector calls stalling past its response-latency comfort zone, with no real benefit for named chemicals. is_contains covers the useful "fuzzy match" case within normal response times.

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Using with Claude Code

Copy .mcp.json.example to .mcp.json and update the paths to point at this repo's .venv/bin/python3 and server.py:

cp .mcp.json.example .mcp.json

Then edit the paths inside, restart Claude Code in this directory, and approve the project MCP server when prompted.

Using with Claude Science

Claude Science connects to local MCP servers via a custom connector, which runs under a macOS sandbox (sandbox-exec) with restricted process execution, filesystem access, and network access. Getting this working needs both a sandbox exception (below) and the connector itself.

1. Allow this repo through the sandbox

Claude Science reads optional settings from ~/.claude-science/config.toml (create the file if it doesn't exist — it's not created by default). Add this repo's path to [sandbox] user_read_paths so the sandbox permits reading server.py and its dependencies, and add chemicalsafety.com to [sandbox.network] allowed_domains so the sandboxed process can actually reach the search endpoint:

[sandbox]
user_read_paths = [
    "/Users/you/sds-mcp-server",
]

[sandbox.network]
allowed_domains = [
    "chemicalsafety.com",
]

This file is only read at startup, so restart Claude Science after editing it. See the configuration file reference for the full set of config.toml keys.

2. Add the connector

1. Run the setup steps above so .venv/ exists with dependencies installed in this repo.
2. In Claude Science, go to Settings > Connectors > Add connector > Local command.
3. Fill in:

   • Name: sds-lookup

   • Command (single field, the whole command line): the absolute path to this repo's venv Python interpreter, including its version number, followed by the absolute path to server.py, e.g.

     /Users/you/sds-mcp-server/.venv/bin/python3.14 /Users/you/sds-mcp-server/server.py
     

     Run ls .venv/bin | grep python3\\. to find your exact interpreter name (e.g. python3.14). Using the unversioned python or python3 symlink instead fails with sandbox-exec: execvp() of '.../.venv/bin/python' failed: Operation not permitted — the sandbox's exec check doesn't resolve the extra symlink hop those names add, but the version-numbered binary resolves directly and is covered by the user_read_paths grant from step 1.

4. Click Add, then approve the search_sds_by_name tool when Claude first tries to use it (or set it to Always allow on the connector's page under Tools).

No PYTHONPATH or system-Python juggling needed — once the repo is in user_read_paths, the sandbox can execute the venv's own interpreter directly, and it already knows where its own site-packages are.

Notes

The chemicalsafety.com search endpoint is undocumented and returns 403 unless the request includes browser-like User-Agent, Origin, and Referer headers — this is already handled in sds_search.py.

Roadmap

Not yet implemented:

• Search by CAS number
• Filter by manufacturer
• Fetch and extract full SDS document content (most links are PDFs)