MCP for agents

The first docs platform agents can read.

Every CodeSummary site ships with a Model Context Protocol endpoint. Wire it up once in Claude Code, Cursor, ChatGPT, VS Code, or Zed; the agent gets cited answers about your product and a live map of your codebase, not guesses.

how it works

Three steps. Repeats forever.

Step 1
Open the MCP tab
Every site has a Sites > MCP tab with the endpoint URL ready to copy, plus a token for team-only sites.
Step 2
Wire it up in your client
One command in Claude Code, one config block in Cursor / VS Code, one paste in ChatGPT desktop. Standard MCP HTTP transport.
Step 3
Ask the agent
“How do I authenticate?”, “Where is the retry logic?”, “Walk me through the onboarding flow.” One ask call returns a cited answer from your docs; orient and lookup answer codebase questions from a live map of your source.

What MCP is

A standard, not a vendor lock-in.

Model Context Protocol is an open spec Anthropic introduced for letting AI clients pull live context from external tools instead of guessing from their training cutoff. It's now supported by every major agent surface, not just Claude.

CodeSummary's endpoint goes well past doc search. ask answers a question about your product in one call, grounded in your docs and cited. orient hands a coding agent a ranked plan of the files and code relevant to its task. lookup returns where a symbol is defined and everywhere it is used, in milliseconds. Plus full-text search and page fetchacross every doc, and the OpenAPI spec when the site has one.

The endpoint

One URL per site. Token-scoped if it's private.

Every site exposes its endpoint at/products/<id>/mcpon whatever host the site lives on (your custom domain or our default).

Public sites don't need a token. Team-only sites mint a per-token URL from the dashboard so you can give different agents different scopes. Revoke a token from the same UI; access cuts off instantly.

Wire it up

One command, copy-pasteable.

We copy the right command for your client into your clipboard from the dashboard's MCP tab. The example to the right uses the public docs at docs.acme.com as a stand-in.

Team-only sites add a token header to the request; the dashboard generates the right snippet for that case too.

Claude Code

claude mcp add codesummary \
  https://docs.acme.com/products/<id>/mcp

Cursor / VS Code / Zed (~/.cursor/mcp.json)

{
  "mcpServers": {
    "codesummary": {
      "url": "https://docs.acme.com/products/<id>/mcp"
    }
  }
}

ChatGPT desktop

Settings → Connectors → Add MCP server → paste URL.

For coding agents

Your agent stops exploring and starts working.

Coding agents burn most of their time and tokens figuring out where things live before they can change anything. CodeSummary parses your code on every push into a symbol map the agent can query directly: definitions, call sites, and a ranked view of what matters for the task at hand.

Setup installs itself. Every docs PR includes a short rules block for your repo's agent instructions, so merging the PR is the whole integration. From then on, every agent on the team routes its codebase questions through the endpoint.

ask("how do refunds work?")

A grounded, cited answer in one call. No file spelunking.

orient("add rate limiting to the API")

The files and code relevant to the task, ranked, with a plan.

lookup("createInvoice")

The definition and every call site, in milliseconds.

The smart part

An endpoint that answers, not a folder that searches.

Most MCP servers hand the agent raw pages and make it do the reading. Ours has an intelligence layer inside: ask a question, and the endpoint reads your docs and code map on its side of the wire, then returns one short answer with its sources and file paths attached. The thinking is included in your plan.

No more "based on my training data" preambles. No more out-of-date framework names. The answer is whatever your docs and code say, today, with citations your agent can verify.

dashboard.codesummary.io

screenshot

GIF: ask Claude Code a question, get a grounded answer from the docs