MCP overview

An agent reads, drafts and commits. A person publishes.

MCP overview

Staleless exposes your documentation over the Model Context Protocol, so an agent can work on it with tools instead of by scraping a website.

The endpoint

https://mcp.staleless.com/mcp

Streamable HTTP, authenticated by a bearer token you mint in the app. The token names one project, so there is no header to get wrong.

Configuring a client

{
  "mcpServers": {
    "staleless": {
      "url": "https://mcp.staleless.com/mcp",
      "headers": {
        "Authorization": "Bearer ${STALELESS_PROJECT_TOKEN}"
      }
    }
  }
}

Danger

Put the token in an environment variable, never the literal string in a file you commit. A token pasted into a tracked configuration file reaches your repository the next time somebody commits.

Three scopes

Scope	What it permits
`docs:read`	Read the tree, pages, history, diffs, and run the linter.
`docs:write`	Stage a create, update, move or delete. Never touches git.
`docs:commit`	Write history on a branch, and open a pull request.

Mint the narrowest token that does the job. An agent that only answers questions needs docs:read and nothing more.

What an agent cannot do

Publish. Merging is not offered to a machine at all. docs_push returns a pull request link for a person to read.

This is not a limitation to work around. An agent writing to the live site routes around the editorial model, and the editorial model is the product.

The staged layer

docs:write tools stage. A staged change is recorded so it can be reviewed and then committed. Nothing enters git history until docs_commit, and nothing reaches a reader until a person merges.

So an agent can be wrong three times and leave no trace.

The intended sequence

docs_branch  →  docs_read  →  docs_write (repeatedly)
             →  docs_lint  →  docs_diff  →  a HUMAN reads it
             →  docs_commit  →  docs_push  →  a HUMAN merges

See The agent workflow for what each step is for, and Tools for every tool.

A public MCP server

A project can also expose a read-only MCP server for its published documentation, so anybody’s agent can read your docs. A private project’s MCP requires a token. An admin can turn either off.