drover — run a team of coding agents in parallel
drover is an MIT-licensed macOS toolkit for developers who run teams of coding agents in parallel in Herdr, with task dispatch, isolated Git worktrees and cross-harness review.
It is for developers operating more than one coding CLI on the same repository. It is not a coding model, hosted IDE or security sandbox. It coordinates Claude Code, Codex, OpenCode and other configured Herdr harnesses; you supply their accounts and your project's gates.
Website · Get started · FAQ · Alternatives · LLM-readable index · Cite a version
[!CAUTION]
Drover was extracted from one operator’s fleet, not a general reliability benchmark. Agents use your local credentials and configured native permissions. Inspect install.sh --dry-run, configuration and permission settings before enabling automation. macOS is the supported operating environment.
What happened in one real fleet
These are historical observations from the maintainer’s fleet, not promised improvements or independent benchmarks. The public lessons preserve their dates; raw operational ledgers are not distributed.
- 2026-09-07 — 44-minute median dispatch gap. Across 650 gaps between dispatches to the same seat, the median was 44 minutes and p90 was 285 minutes, while tasks took 10–20 minutes. The lead had been polling in eight-minute sleeps. Drover replaced that handoff with explicit completion announcements. Dated source.
- 2026-09-07 — 10 “done” seats produced no work. After OpenCode credits ran out, ten agents reported done with no commit, report or error. That incident is why status alone does not release ownership and budget checks inspect provider-limit evidence. Dated source.
- 2026-09-18 — the $952 daily-meter bug.
opencode stats --days 1 counted lifetime session cost for sessions touched in the window. At one checked instant it read $856.01 while the message cost inside 24 hours was $212.60. The guard changed to summing message cost since local midnight. Dated source.
A 15-second fixture demo
Download the asciinema recording and run asciinema play demo.cast; read the transcript. It records real fleet-claims-audit invocations against synthetic local fixtures, first with matching ownership and then conflicting seats. It does not show a live fleet, paid agent calls or measured throughput.
drover vs alternatives
First-party descriptions read 2026-10-19. These are workflow differences, not benchmark rankings. Projects evolve; check their linked sources.
| Project | Documented workflow | Source |
|---|
| drover | A local Herdr fleet loop: task dispatch, budget/ownership evidence, cross-harness review and gated publication. | Operating model, source |
| herdr-synapse | A Herdr plugin with a shared team board, human-owned charter and idle-gated message delivery. | First-party source |
| Claude Squad | A terminal app using tmux and Git worktrees to run coding agents on separate tasks and review changes. | First-party source |
| Conductor | Its official docs describe parallel coding agents in isolated workspaces, with diff review and pull-request workflows. | First-party source |
| Vibe Kanban | Its README describes kanban planning, agent workspaces and inline diff review; it also announces the project is sunsetting. | First-party source |
| Claude Code agent teams | Anthropic documents a lead session coordinating Claude Code teammates through shared tasks and inter-agent messages. | First-party source |
Conductor’s public release repository has no product README, so the row cites its official docs; its tutorial README is a learning project. Claude Code’s own README links its official docs; the agent-teams behavior above comes from that feature’s documentation. Source audit.
FAQ
How do I run multiple Claude Code agents in parallel?
Install Herdr and authenticate your coding CLIs, inspect Drover’s install dry-run, configure a lead plus separate coder/reviewer seats in examples/config, then run fleet up. Each assigned coder works in its own Git worktree. Start with two seats and one task; inspect the result before enabling watcher automation.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
