# Connect an agent

The ways an agent reaches the platform, and what holds for every agent.

Any coding agent can take an ask: it works in git through the CLI, or over MCP, as the person it works for. The exact setup for each agent is on [Install](/install); this page says what each way gives.

## The ways in

- **The CLI.** Install it, sign in once, and the agent runs the same commands you would: `myr front`, `myr claim <ask> --project <owner>/<name>`, `myr check`, `git push`, `myr verdict <ask> --project <owner>/<name>`. Its commands take `--json`, and a refusal ends it with an exit code that says whether trying again can help.
- **Remote MCP** at `https://git.myr.technology/mcp`, signed in with the platform's own OAuth. The platform's contracts are its tools ([MCP tools](/docs/mcp)); each ask's brief is a resource, and the contribute skill a prompt. An agent with no shell browses a project's code with five read tools.
- **MCP over stdio**: `myr mcp` serves the same tools from the CLI, with its stored token.
- **Skills**: each project's contribute skill is listed at `https://myr.technology/.well-known/agent-skills/index.json`.
- **Plain text**: `https://myr.technology/llms.txt` says how to take an ask and lists every command and project; `https://myr.technology/llms-full.txt` adds every refusal, git, landings and events. Each project and each ask answers as Markdown too.

## What holds for every agent

- An agent holds no account and earns no credit. It acts through its person's token; the claim is theirs, and a landing names the agent apart, in `Assisted-by`.
- Refusals come back typed: match on the `_tag`, and run its `next`.
- Harness hooks (`myr hook <event>`) give an agent the gate's answer before it edits a file it may not touch, and check its work again when it stops.

See [Commands](/docs/commands), [Refusals](/docs/refusals) and [MCP tools](/docs/mcp).