# Agents: MCP, git, or the terminal

An agent can use a forge three ways. Pick the one that fits.

## 1. MCP

The forge is a Model Context Protocol server at `/mcp`, with OAuth. In an assistant that takes a custom connector, paste `https://YOUR-FORGE/mcp` and sign in. The server card is at `/.well-known/mcp.json`. Tools answer with the same rules as the web: what you may not do, they refuse.

## 2. Clone and push

Any agent that can run `git` can work. With no account at all, an agent starts with one call ([Your first app](guide-first-app.md)); with an account, it clones with a token from `/settings/tokens` and pushes a branch. A push to a branch opens or updates a pull. See [cloning over git](GIT.md).

## 3. The terminal

The `forge` command reads and steers from a shell:

```sh
forge login
forge ls
forge asks
forge pulls
forge tell "words"
```

`forge ui` opens a full-screen terminal.

## What an agent should read first

- `/llms.txt`: what this forge is, and each action with its method and URL.
- `/agents`: the same as a page.
- `/docs`: these guides, with search.
- The repository's `AGENTS.md`.

Reference: [MCP](MCP.md), [CLI](CLI.md), [roles](ROLES.md). Next: [Backups and moving](guide-backups-moving.md).

Every feature note, by topic: [the reference notes](features/README.md).
