Autocoding Agents

Point Claude at your repo and a fleet of agents reads your tasks, writes the code, runs your tests, and opens pull requests, overnight, on your own Claude Max/Pro subscription (no per-token API charges).

1. What you need

A redu.cloud account . You sign in with it when you connect the MCP (OAuth).
A GitHub token for the repo you want worked on: a fine-grained token scoped to that one repo, with Contents and Pull requests both set to Read and write.

⚠️ Must be Read and write, not Read-only, a read-only token clones fine but then silently can’t push or open PRs. (The #1 setup mistake.)
Claude Code with a Claude Max or Pro subscription.

Keep the GitHub token in a .env file, and add .env to your .gitignore.

2. Set it up

In VS Code, Cursor, or any IDE

Open the repo you want the agents to work on, then paste this into the Claude Code chat:


Add the redu.cloud MCP server: claude mcp add --transport http redu https://mcp.redu.cloud/mcp --client-id redu-mcp

Reload the window when it asks, so the connection loads. Then paste:


Set up autocoding agents for this repo with 3 workers.

Claude provisions the fleet and can draft your ROADMAP.md right there in the open repo. (Connecting is one-time, after that, the second prompt is all you need for any repo.)

Or in the terminal

claude mcp add --transport http redu https://mcp.redu.cloud/mcp --client-id redu-mcp

A browser opens so you sign in to your redu.cloud account. If it does not prompt, run /mcp to start the login.

Then tell your agent:


Set up autocoding agents for owner/your-repo with 3 workers.

If Claude Code blocks the setup with a warning about sending your GitHub token to an external service, that’s its safety guard, and it’s expected here: redu.cloud needs your token to clone the repo and open PRs from the worker VMs. To allow it, switch Claude Code so it asks before each action (press Shift+Tab to cycle permission modes, or run /permissions), then re-run the prompt and approve the setup_agent_fleet call. (A smoother one-click GitHub connection is planned.)

3. Authorize, then go

Setup takes ~7 minutes. Two things left:

Authorize, run the one-line command the setup returns, in your terminal. It logs you in once (claude setup-token) so the agents use your subscription. Do it once; it lasts.
Add tasks to ROADMAP.md (Claude can write these for you), one per line, exactly this shape:
```
- [ ] **Add a /health endpoint**, returns 200, with a test *(agent-ready)*
```
Then tell Claude:
```
trigger the agent batch
```

PRs start landing within minutes. Agents do best on focused, testable tasks (an endpoint, a test, a bug fix) and poorly on vague or huge ones. When you’re done, just ask Claude to “tear down the fleet” so the VMs stop billing.

Windows: run the terminal and authorize commands in WSL or Git Bash.