# Hive

> An open-source async multi-agent CLI pipeline that turns a rough idea into a merge-ready pull request through nine inspectable stages with markdown artefacts at every step.

Hive is an open-source (MIT) command-line tool built by Ivan Kuznetsov that automates the journey from a rough software idea to a merged pull request. It runs an async multi-agent pipeline in the background — brainstorm, plan, execute, review, finalize — and surfaces only the moments that need a human answer. The project is written in Ruby, ships as the `hive-cli` rubygem, and is available on macOS via Homebrew, Arch Linux via AUR, and glibc Linux via a signed install script.

## What It Is

Hive is an agentic harness for software development workflows. It sits between a developer's rough idea and a finished pull request, orchestrating a sequence of AI coding agents (Claude, Codex, or Pi by default) through nine discrete stages. Each stage writes a durable markdown artefact — idea, brainstorm, plan, reviews, PR metadata — that the next stage can read, edit, or hand to another agent. The result is a fully inspectable, file-based state machine rather than a black-box chat thread.

## The Nine-Stage Pipeline

The pipeline is the core abstraction. Every task is a folder; the folder's location on disk is its state. Moving a task from one stage directory to the next is the approval gesture:

- **1-inbox** — capture a rough idea (one sentence is enough)
- **2-brainstorm** — an agent pins down what you actually want
- **3-plan** — scope, implementation units, and risks are fixed
- **4-execute** — code is written in an isolated git feature worktree
- **5-open-pr** — the branch is pushed and a draft PR opened
- **6-review** — CI fixes, multi-agent review, triage, and fixes
- **7-artifacts** — release and handoff artefacts are gathered
- **8-finalize** — the PR body is refreshed and marked ready
- **9-done** — the task is archived with cleanup commands

The project calls this practice "compound engineering": making each step's output strong enough for the next one to run autonomously.

## Autonomous Features: Patrol and Babysitter

Beyond the core pipeline, Hive ships two opt-in autonomous modes. **Patrol** maps feature slices across a repository, reviews them, validates fixes, and opens PRs only for the ones that pass. **Babysitter** keeps open PRs green and mergeable — it runs bounded repair attempts and auto-rebases behind a moving `main` branch, then hands off when it is stuck. Both modes run through the same daemon that advances the main task queue.

## Interfaces: TUI, Telegram Bot, and Agent CLI

Hive offers three interaction surfaces. The **TUI** (`hive tui`) is a two-pane terminal dashboard showing registered projects and the live task queue; pressing `n` captures a new idea and `Enter` opens the relevant markdown file in the user's editor. The **Telegram bot** uses long polling (no webhook or public URL required) to let users capture ideas, check the queue, and approve work from a phone — including photos, documents, and transcribed voice notes. The **agent CLI** exposes every workflow verb with `--json` output and typed envelopes so that Claude Code, Codex, or any other coding agent can drive Hive programmatically. Hive also publishes an OpenClaw skill (`/hive`) on ClawHub for guided setup and command dispatch.

## Update: v0.3.0

The latest release is **v0.3.0**, published on 12 June 2026. The repository was created in April 2026 and has been actively pushed to as recently as June 2026, with 37 open issues reflecting active development. The project self-describes as alpha-stage — the homepage labels the Box and Bench sub-tools as "alpha" — and the README flags the `tmux` Claude launch mode as an "experimental workflow" with rough edges. The install script pins to v0.3.0 and each release is signed with cosign keyless attestation.

## Tradeoffs to Know

The homepage is explicit about the cost profile: Hive is "token-heavy" by design. The default configuration runs many subagents and several coding agents per task. The project recommends a Claude Max subscription plus ChatGPT Pro for Codex to get the best results, and notes that cost-sensitive users may try the Pi agent with a recent Kimi model for parts of the workflow, though that integration is not yet performance-tuned. Hive is also strictly local — there is no hosted web app or managed service. It requires a background daemon running on the user's own machine, Ruby 3.4, git ≥ 2.40, authenticated `claude` and `gh` CLIs, and tmux when using the default Claude launch mode.

## Features
- Nine-stage async multi-agent pipeline from idea to merged PR
- Folder-as-agent state machine with markdown artefacts at every stage
- Multi-agent support: Claude, Codex, and Pi
- Background daemon advances tasks in parallel
- TUI dashboard for monitoring and human-input gates
- Telegram bot for remote idea capture, queue checks, and approvals
- Patrol mode for autonomous repository review and PR creation
- Babysitter mode to keep open PRs green and auto-rebased
- JSON output envelopes for all CLI verbs for agent-driven automation
- OpenClaw /hive skill for guided setup and command dispatch
- Isolated git feature worktrees per task
- Cosign keyless attestation on every release
- Voice note idea capture via Telegram with OpenAI-compatible transcription
- hive doctor diagnostics and hive status --json for programmatic inspection

## Integrations
Claude (Anthropic Claude Code), Codex (OpenAI), Pi agent, GitHub (gh CLI), Telegram, OpenClaw / ClawHub, tmux, Homebrew, AUR (Arch Linux), systemd-user (Linux autostart), launchd (macOS autostart), Kimi model (via Pi agent), OpenAI-compatible Whisper transcription API

## Platforms
MACOS, LINUX, WEB, API, CLI

## Pricing
Open Source

## Version
v0.3.0

## Links
- Website: https://hivecli.sh
- Documentation: https://hivecli.sh/docs
- Repository: https://github.com/ivankuznetsov/hive
- EveryDev.ai: https://www.everydev.ai/tools/hive-cli