clodfarm: a farm of Claude Code agents you can steer from the Claude app

Key Highlights

clodfarm (say it out loud and it becomes "cloud farm") is a self-hosted orchestration farm for Claude Code agents. You plant a "mission," and it splits the task across a team of sub-agents that get to work automatically and divide the labor, while watching each account's real 5-hour and weekly usage caps so it never runs itself over the limit. The name is a joke, but the problem it solves is serious: scaling one agent into many without melting your rate limits or waking up to a bill you did not expect.

What It Does

Using it feels like running a farm. You submit a mission, clodfarm breaks it into subtasks and assigns them to multiple Claude Code sub-agents running in parallel, with progress laid out on a dashboard you can actually watch. Its most practical feature is "rate-limit self-discipline": it paces work according to each account's actual rate caps to avoid tripping limits and getting throttled mid-job. You can also remotely steer the farm from the Claude app, so a long task can keep moving while you are away from the keyboard, and you can intervene when a sub-agent goes off track instead of discovering the mistake hours later at the end.

Technical Details

The backend is written in Python, and self-hosted deployment relies on AWS, Docker, and DynamoDB, giving it a cloud-native stack that fits teams already on those services and already used to operating containers. It targets users who "want several Claude Code agents working for me at once, but don't want to manually manage rate limits and division of labor." The UI features pixel-art farm visuals, a token counter, and account cards so status is visible at a glance; the docs cover both container and serverless deployment shapes, which means you can start small on one machine and grow into a managed setup later without rewriting everything or abandoning the dashboard you already rely on.

Versus Alternatives

Compared with simply throwing a prompt at one agent, clodfarm's value is the integration of "task splitting plus parallel agents plus rate-limit orchestration." It shares a direction with various multi-agent frameworks that dispatch work to sub-agents, but it treats Claude Code's rate constraints as a first-class concern, sparing you from writing your own throttling logic and fitting real multi-account usage better. Most toy orchestrators assume infinite quota; clodfarm assumes the opposite, which is the world the rest of us actually live in with metered API accounts that punish careless parallelism with hard limits and cool-down periods.

Industry Impact and Use Cases

It suits people with Claude Code capacity who want to batch repetitive coding or research tasks. Simply put, it upgrades "one person watching one agent" into "one farm managing a herd of agents," which is attractive for individuals and small teams that want to scale Claude Code engineering, and makes it easier to split long tasks and run them to completion steadily. For a solo developer, that can mean a background fleet finishing the boring parts; for a small team, it can mean one mission fanned out across accounts without anyone manually babysitting the limits or copying context between sessions by hand.

Practical Notes

To try it, deploy the container via the provided Docker setup and connect the Claude Code accounts you are willing to lend to the farm, then start with one small mission to learn how it splits work. Watch the token counter and the account cards before you scale up, because the whole point is respecting the 5-hour and weekly caps, not fighting them. Keep missions narrow and well-scoped so sub-agents have clear boundaries and less chance to collide. Review the output of each sub-agent before you merge it, since parallelism speeds things up but does not remove the need for a human to accept the result. Check the MIT license for your own usage, and self-host so your code and keys stay under your control. Start with a throwaway mission the first time, so you learn the dashboard and the splitting behavior before you point the farm at work you care about. Once the farm is steady, promote a real low-risk task to it and compare the result with doing the same thing by hand, so you learn where parallelism helps and where a single careful agent is still better.