What it is
A public, read-only MCP server from OpenAI that lets your coding agent search and read OpenAI developer docs directly from your editor.
Gabriel’s notes
Docs MCP is a public Model Context Protocol (MCP) server hosted by OpenAI that provides read-only access to OpenAI developer documentation (search + page content). The server endpoint is published as a streamable HTTP MCP server at https://developers.openai.com/mcp, and it’s intended to let agents pull the freshest docs into context while you work—without you playing the “copy/paste/screenshot the docs” mini-game.
Quick take: If you’re tired of your coding agent confidently inventing OpenAI API details from six months ago, this is the boring, correct fix: give the agent a doc faucet it can turn on whenever it needs receipts.
I saved this under Dev & code because it’s the most direct way I’ve seen to reduce “agent hallucination risk” on OpenAI-specific implementation details—schemas, parameters, and product behavior—while staying in-flow in the editor.
Good fit if you want to:
- Keep your agent’s OpenAI API knowledge up to date without manual linking or context dumping.
- Ask for a specific request schema / parameter behavior and get an answer grounded in official docs.
- Use MCP in tools like Codex (CLI/extension), VS Code agent workflows, or Cursor, and standardize “docs-first” behavior.
- Reduce “it compiled yesterday” debugging caused by doc drift, renamed parameters, or deprecated endpoints.
Pricing snapshot (auto-enriched)
OpenAI doesn’t list a separate price for using the Docs MCP server itself on the Docs MCP page (so: Unknown / not confirmed). Practically, the cost you’ll feel usually comes from the model/tool you’re using (e.g., your LLM usage in Codex/Cursor/VS Code) and any OpenAI API calls you make in your actual code, which are billed per OpenAI’s API pricing.
Work-use / compliance snapshot (auto-enriched)
The Docs MCP server is explicitly documentation-only and does not call the OpenAI API on your behalf. That’s good: fewer surprise side effects and a simpler threat model.
That said, you’re still sending queries to a remote server over the network. What exactly is logged/retained for Docs MCP queries is Unknown / not confirmed from the Docs MCP page alone—so treat prompts you route into it like any other external request: don’t include secrets, private keys, or proprietary code you wouldn’t paste into a ticketing system.
If you’re also using the OpenAI API in your workflow, note that OpenAI states API data is not used to train or improve OpenAI models unless you explicitly opt in, and they document retention controls separately.
Alternatives (auto-enriched)
- Cursor “@Docs”: built-in documentation context feature inside Cursor; great if you live in Cursor already, but it’s editor-specific rather than a general MCP endpoint.
- Context7 MCP: an MCP-based documentation retrieval service for a broad set of third-party libraries/frameworks; better for non-OpenAI docs, but it’s a different provider with its own data/privacy model.
Before you adopt it:
- Add a one-line rule to your agent’s project instructions (e.g.,
AGENTS.md) that says “use Docs MCP for OpenAI questions first,” so it doesn’t “forget the faucet exists.” - Decide what you’ll allow into MCP queries (e.g., no secrets, no customer data). Write it down. Future-you will thank you.
- In multi-MCP setups, keep server names short and unambiguous so the agent picks the right tool.
Sources
- https://developers.openai.com/learn/docs-mcp
- https://developers.openai.com/api/docs/guides/your-data
- https://openai.com/api/pricing/
- https://help.openai.com/en/articles/12584461-developer-mode-and-mcp-apps-in-chatgpt-beta
- https://docs.cursor.com/en/guides/advanced/working-with-documentation
- https://context7.com/docs/security/data-privacy