Skip to main content
SEC API exposes filing and research tools through a hosted MCP endpoint. Pick the client that will actually run the work, connect with an API key owned by that runtime, then prove one narrow filing lookup before you give an agent a larger assignment. GET https://api.secapi.ai/mcp is public discovery metadata. It confirms the deployed connection shape, but it does not authorize a tool call. Each client guide below configures an authenticated MCP connection separately. ChatGPT MCP compatibility is not currently published or verified; see the Custom GPT status page rather than treating the discovery endpoint as a working setup path.
Create an API key in the SEC API dashboard and keep it in the secret store that actually runs the client. Do not put a production key in browser code, a repository, an agent prompt, or a shared desktop configuration. A desktop IDE configuration is appropriate for a developer-owned development key; a deployed agent should read its key from the server-side runtime.

​
Choose a path

​
IDE clients

  • Continue keeps a remote MCP server in its YAML or MCP server configuration. Use it for an interactive, developer-owned workflow.
  • Windsurf reads a global JSON MCP configuration and can interpolate an environment variable into the request header.
  • Claude Desktop, Claude Code, and Cursor have their own installation instructions.

​
Application frameworks

  • LangChain discovers remote MCP tools and supplies them to a Python or TypeScript agent.
  • LlamaIndex adapts an MCP client into LlamaIndex FunctionTool objects.
  • Vercel AI SDK converts remote MCP tools to an AI SDK tool set. Keep this code on the server, not in a client component or route response.

​
ChatGPT

  • Custom GPT explains why ChatGPT MCP setup is not currently published or verified, and directs builders to supported clients.

​
Prove one filing result

Use this as the first prompt after your client shows SEC API tools:
Use filings.latest with ticker: AAPL and form: 10-K. Return the company name, form, filing date, accession number, filing URL, requestId, and traceparent. Do not infer a filing from a ticker alone.
The tool schema defines accessionNumber, filingDate, and filingUrl for this response; requestId and traceparent are part of the MCP output envelope. A good first result names the issuer, links its filing, and has an identifier another person can reopen. Keep those fields with a cached answer, generated report, or review record. For extraction or search tools, retain provenance, freshness, and source fields when returned rather than replacing them with an uncited model summary.

​
When you move beyond a test

Tool discovery and tool calls are separate lifecycle steps. Refresh the tool list when you deploy or when a tool schema changes; do not hard-code a copied tool definition indefinitely. Bound the work an agent can request with issuer, form, date, and result limits before it fans out. For failures, inspect the MCP JSON-RPC error and error.data.code. Correct validation errors, rotate or fix a rejected key on 401, resolve billing or quota state on 402, and honor Retry-After on 429 or a temporary 503. For mcp_tool_timeout, reduce the request scope before one bounded retry. Keep the request ID with a repeated failure. See the Error code catalog and Request diagnostics.

​
Next step

Open the page for the client you are using, complete its first filing lookup, then turn that verified call into the smallest useful agent workflow.