SDS data for AI agents.

What the server offers#

The Model Context Protocol (MCP) lets an AI assistant call tools through a standard interface. Our server uses the Streamable HTTP transport, is stateless, and sits at https://safetydatasheetapi.com/mcp. It is listed in the official MCP registry as com.safetydatasheetapi/sds-extraction.

• extract_sds: extract one SDS from a public HTTPS file_url or from file_base64.
• submit_bulk_extraction: submit up to 10 documents as one asynchronous job.
• get_bulk_job: read the status and per-file results of a job.
• get_account: read the prepaid balance, price per document and 30-day usage for the key.

Connect a client#

Point an MCP client at the URL and send your API key as a bearer token. Discovery calls, initialize and tools/list, work without a key so clients and registries can see the tools. Calling a tool without a valid key returns 401 with a pointer to the protected-resource metadata. Get a key in the dashboard.

The same rules as the REST API#

Every tool forwards to the same handlers as the REST API, so authentication, rate limits, size limits and error messages are identical. Each extraction is billed from your prepaid credit at $0.25 per 40 pages, and failed extractions are not billed. Because a model decides when to call a tool, get_account is useful: an agent can check the balance before it submits a large batch. File URLs are fetched by the server, which accepts HTTPS addresses only and limits redirects and download rates.

What an agent should do with the result#

An SDS drives decisions about storage, transport and exposure. Give the agent a rule: when warnings is not empty or confidence_score is low, report the record as unverified and ask a person to review it, instead of acting on it. The server's own instructions say the same. Results are compact, with the record only; XML and CSV are available through the include_formats option.

Useful prompts#

• "Extract this SDS and list the H statements and the UN number."
• "Submit these eight SDS URLs as one job, then summarize which have warnings."
• "How much credit is left, and how many documents did we process this month?"

The clients and Postman collection and the MCP manifest have the full tool definitions.

Related reading#

• AI agents and MCP in the docs
• Clients and Postman collection
• Agent skill index