# The agent guide

PalmaPlan exposes a full MCP server: 101 tools, 10 resources and 5 prompts (the exact
count, verified at every build, is in [the generated reference](reference-tools.md)). An
agent — Claude Code, Claude Desktop, or any other MCP client — connects to it without
writing a line of code against an API: the MCP server *is* the API, in the schedule's own
words rather than a REST schema's.

This part of the wiki has two readers at once, and says so plainly: a human connecting an
agent to their instance, and the agent itself, which can read it raw (unrendered) to learn
how to use it without being walked through it by hand.

- **[Connecting Claude Code](claude-code.md)** — the command, browser authorization
  step by step, the named token that comes out of it, what to check when it stalls.
- **[Connecting Claude Desktop](claude-desktop.md)** — stdio, the configuration, the
  address pitfall.
- **[Tokens and scopes](jetons-et-portees.md)** — read / write, one token per client,
  revocation, what an agent cannot do.
- **[The embedded view](vue-embarquee.md)** — `show_schedule`, MCP Apps, and what an
  agent that doesn't render it gets instead.
- **[The tools, resources and prompts reference](reference-tools.md)** — generated from
  the server itself, never written by hand.
- **[For agents](pour-les-agents.md)** — the machine path, with no human detour:
  discovering, authenticating, building a schedule, reading one.

Everything described here is checked against the MCP server's code (`packages/mcp/src/`)
and the API's OAuth (`packages/api/src/http/oauth.ts`), or replayed for real against a
local instance. Where something hasn't been replayed under real conditions yet, the page
says so — rather than presenting it as settled.