I built agent fleets five times before this one worked. Named it Chump so I'd remember who was doing the learning.
what it is
A coordinator for a fleet of coding agents.
Concretely, it's a multi-agent fleet coordinator. It takes natural-language task descriptions, filed to a SQLite registry, and turns them into merged pull requests, autonomously, without human review at every step. A fleet of coding agents (Claude Code, opencode, Codex CLI, or anything that can push a branch) picks tasks from the queue, claims atomic file ownership via lease files, commits work to isolated git worktrees, and ships through a merge-queue pipeline.
Chump is not an LLM wrapper. The models are interchangeable and optional. The thing that makes agents productive is the coordination layer: the task registry, the lease graph, the ambient event stream, the per-task briefing memory.
why now
The model problem is close to solved. The coordination problem is not.
Commodity hardware is crossing a threshold. A $500 mini-PC can serve a capable 7B model today. The model access problem for solo developers is close to solved.
The coordination problem is not. Spawn agents against the same codebase without a substrate and you get chaos: they overwrite each other's changes, claim the same task twice, ship conflicting PRs, and stall on flaky tests with no recovery. Every hour of agent time gets wasted on collision recovery that a 20-line lease file would have prevented.
Chump is the coordination layer that was missing, built to run on hardware a solo developer already owns, with models they can run locally, with no API bill.
who it's for
People who already watched "spawn more agents" fall apart.
Engineers who understand what a merge conflict costs when six agents hit the same file at once. People with a real backlog, hundreds of filed tasks, not a handful, who need something to grind through it overnight without babysitting.
Chump is not for people who want to prompt an AI once and review the result. A chat interface handles that. Chump is for people who want a fleet that runs for hours without input, produces real PRs, and tells you what shipped, what was skipped, and why.
Big and actively built. 14 crates on crates.io (the reusable libraries under Apache-2.0, the coordinator itself AGPLv3 so a cloud giant can't host it and rent it back to you), plus a 4-platform Homebrew tap. Young: early days on stars and adopters, and I audit its own hype in the open. Real, not finished.