Orchestrate Agent Colonies with Hive

Most multi-agent frameworks ask you to draw a graph of agents and wire their inputs and outputs. Hive — an 11k-star open-source runtime — has one execution primitive instead: the agent loop. A queen agent runs it, and every worker is a clone of it with its own task, fewer tools, and a strict budget.

Why This Skill Matters

Tell a queen what you need and she does the first piece herself to learn what the job takes, then hands the rest to as many parallel workers as it needs. Every result lands in a shared tracker — a SQLite file on disk at ~/.hive/colonies/<name>/tracker/tracker.db — that the queen queries with SQL to see what is done and what is left. Because state lives on disk, a crash or restart resumes exactly where things stopped.

Hive ships thirteen persona queens, six of them active out of the box: Growth, RevOps, Content, Lead Generation, Outbound, and Brand & Design. It runs on any model LiteLLM supports, including coding subscriptions such as Claude Code and OpenAI Codex, plus local models through Ollama.

Installation

You need Python 3.11+, Node.js 20+, and git:

git clone https://github.com/aden-hive/hive.git
cd hive
./quickstart.sh

On Windows, run .\quickstart.ps1 instead. The quickstart installs uv and ripgrep if missing, stores your API key encrypted under ~/.hive, and opens the dashboard at http://127.0.0.1:8787. Next time, run hive open from the repository.

One trap from the README: pip install -e . at the repo root installs a placeholder that will not run — use the quickstart. Without the repository, pip install "aden-hive[tools]" installs Hive from PyPI, then hive open starts it.

Real Workflow: Research Five Product Launches

The README's own demo shows the shape of a colony run. A growth team wants to know how five developer tools' launches did on Hacker News:

1. Hand the task to a queen. Type the goal on the home screen and choose the Head of Growth queen.
2. She asks the question that changes the answer — every launch post, or only each product's original launch? You pick, and she carries on.
3. She proposes a colony; you approve it. Five products means five parallel jobs.
4. She does one herself. She takes one product, works out what counts as a genuine launch post, checks numbers against the Hacker News API, and saves the method as a reusable skill.
5. She fans out. A playbook starts one worker per product; each follows her skill and writes its row to the tracker.
6. She checks every row before answering. In the demo she catches a worker that counted a post from a product's old name, clears the row, and the final chart shows that product as not found rather than zero.

Nothing is wired up in advance — the queen shapes the colony as the work unfolds, and the tracker, not a context window, keeps count.

Tips

• run_worker returns at once and runs four workers at a time by default; the rest queue, and each report arrives as a new turn in the queen's conversation.
• hive mcp add attaches any MCP server, and its tools join the same allowlists as the built-ins.
• The Browser Bridge extension lets agents work in your own Chrome, with each worker getting its own tab group.
• Colonies can wake on cron, interval, or webhook triggers; opt-in Sentinel watches a stopped queen and escalates to you through the inbox, Telegram, or Slack.
• Every worker runs under hard turn and tool-call limits, so a stuck worker stops itself — but there are no dollar spending limits yet.

When Not to Use This

For a single prompt or a one-off script, the README says a plain agent is simpler. Hive earns its keep when the work has many similar units — leads, tickets, repositories, documents — that must run in parallel, survive restarts, and end in results you can query and audit.

See the leaderboard for more skills.