herd

Herd your agents.

GitHub-native orchestration for AI coding agents.
One command. Parallel workers. One reviewed PR.

Getting Started · Install · Design Docs

herd plan "Add user authentication"

That's it. Herd decomposes your feature into tasks, dispatches AI agents to work them in parallel on self-hosted runners, consolidates the results, runs an automated review, and opens a single PR — ready for you to merge.

You walk away after herd plan. You come back to a reviewed PR.

How it works

graph TD
    A[herd plan 'Add auth'] --> B

    subgraph B[Tier 0 — parallel]
        W1[Worker #42<br>auth model]
        W2[Worker #43<br>login route]
    end

    B --> C[Integrator consolidates]

    subgraph D[Tier 1 — depends on Tier 0]
        W3[Worker #44<br>auth middleware]
    end

    C --> D
    D --> E[Integrator consolidates]
    E --> F[Batch PR opened]
    F --> G[Agent review]
    G --> H[You merge]
    F --> I[You close without merging]
    I --> J[Non-done issues cancelled, cleanup]

Workers run your configured agent (Claude Code, OpenCode, Codex, Cursor) in headless mode. If one fails, the Monitor detects it and retries automatically. The system is self-healing.

Quick start

brew install herd-os/tap/herd

cd /path/to/your/repo
herd init
herd plan "Add dark mode support"

See the full setup guide for runner configuration and options.

Why "Herd"?

Anyone who's tried to coordinate multiple AI agents knows the feeling — it's like herding cats. Each one is powerful on its own, but getting them to work together without stepping on each other? That's the hard part.

Herd tames the chaos.

Documentation

Status

In active development. The core system is functional end-to-end: planning, dispatch, parallel worker execution, tier-based DAG scheduling, conflict resolution, automated agent review with fix cycles, and health monitoring.

License

Apache License 2.0 — see LICENSE.