Development, issues, pull requests, and releases are maintained in this repository, including the gateway source. The default branch can contain unreleased work; npm installs the latest published release. See CONTRIBUTING.md and the release guide. The BSL license is unchanged.

Hive

Run Claude Code, Codex, Gemini, OpenCode, Qwen, Pi, and other CLI agents as a visible local team. Hive gives you one browser workbench where an Orchestrator plans and delegates while workers implement, review, test, research, and report back — all as real PTY processes on your laptop.

Use Hive when one agent is not enough, but a pile of terminal windows is not a workflow.

🌐 Website: hivehq.dev/en/ · 中文

English · 简体中文

Hive is local-first, runs on 127.0.0.1, and is intended for anyone who already runs CLI agents. The latest stable release is on npm and the badge above resolves to it.

Why Hive

CLI agents are powerful, but coordinating several of them manually is awkward:

• Long-running sessions are spread across terminals.
• Splitting work across agents — implementation/review/testing, research/drafting/fact-checking, or any other division of labor — needs a routing layer you don't have.
• Worker progress disappears into scrollback.
• Restart recovery depends on each CLI's native session behavior.

Hive adds the coordination layer without replacing the CLIs. The Orchestrator is a real agy / claude / codex / opencode / gemini / hermes / qwen / pi process, not a scripted PM. Workers are real CLI agents too. Hive injects a small team command into their shells, so they can dispatch, report, and keep a shared markdown task graph at <workspace>/.hive/tasks.md.

Use It For

Ship a PR with a reviewer in the loop

Ask the Orchestrator to implement a change, spawn a reviewer, and keep the review feedback visible before you merge. The coder edits; the reviewer checks the diff; the Orchestrator decides what still needs work.

Ship the settings search bugfix. Use one worker to implement it and another to
review edge cases before the final report.

Run a parallel bug hunt

Give several workers separate slices of a flaky behavior: one reads the server path, one checks the UI path, one looks for regressions in recent commits. You watch the reports converge instead of juggling terminals.

Find why mobile reconnect sometimes stalls. Split server transport, browser UI,
and recent commit history across separate workers.

Research, draft, and fact-check without losing the thread

Let one worker gather sources, another draft, and a reviewer check claims. The task graph and reports stay in one workspace, so the handoff is inspectable instead of trapped in chat scrollback.

Write a technical note on our release flow. Have one worker collect evidence,
one draft, and one verify every command and file reference.

Try the demo first

Don't have an agent CLI installed yet? Run hive, open the printed URL, and click Try Demo in the first-run wizard. The demo is a pure client-side replay: a fake terminal types through planning, team spawn, team send, worker reports, and .hive/tasks.md checkoffs. It does not need network access, an installed CLI, or any sign-in, and it never calls demo workspace server routes.

Quick Start

Prerequisites:

• Node.js 22.18+ (22.x), or 24+.
• At least one supported agent CLI installed, authenticated, and available on PATH.

Install and start Hive:

npm install -g @tt-a1i/hive
hive

Hive uses Node's built-in SQLite and platform-specific PTY binaries. First installation works with npm 12 defaults and with --ignore-scripts; no build tools or script approvals are required. Keep optional dependencies enabled so npm can install the binary for your platform.

Open the printed local URL, usually http://127.0.0.1:9483/. Use hive --port 0 when you want Hive to ask the OS for any free local port.

To upgrade in place:

hive update

hive update installs the new version with --ignore-scripts, preserving the original npm prefix. It never changes your npm script policy. After npm exits 0 it probes SQLite and a short-lived PTY in a fresh process using this Hive CLI Node (process.execPath). The npm child gets that Node directory prepended on PATH for the update only; a custom --prefix is the package target and is not required to contain a Node binary. It will not print a successful update if that probe fails. Restart any in-flight Hive process to pick up the new version. If you installed Hive with pnpm or yarn, upgrade through the same package manager — otherwise the new npm copy will shadow your existing install.

Install Hive as an app (optional):

Open http://127.0.0.1:9483/ in Chrome, Edge, or Brave and click the install icon at the right edge of the browser's omnibox. The PWA launches in its own dock-anchored window without browser chrome and shows Add Workspace / Try Demo shortcuts from the dock right-click menu. Firefox and Safari currently don't implement the install-prompt protocol, so the omnibox icon only appears in Chromium-based browsers.

The Hive daemon must still be running for the PWA to do anything; if the runtime isn't reachable when you launch the app, you'll see a "Hive runtime is not running" page that auto-reloads once hive is back on 127.0.0.1. The PWA install scope is keyed by origin, so hive --port 9484 installs as a separate app from hive --port 9483. To uninstall, visit chrome://apps, right-click the Hive tile, and choose Remove from Chrome….

Hive asks the browser to confirm before closing the tab or PWA window so an accidental close shortcut (Cmd-W on macOS, Ctrl-W on Windows/Linux) doesn't drop your session. Modern browsers gate that prompt on prior page interaction — if you open the PWA and immediately press the close shortcut without clicking or typing anywhere first, it still closes cleanly. That's a browser policy, not a Hive bug.

First-run flow:

1. Create a workspace from a project folder.
2. Choose an Orchestrator preset.
3. Hive creates <workspace>/.hive/tasks.md, starts the Orchestrator PTY, and injects the internal team command into the agent session.
4. Add workers from the Team Members panel.
5. Ask the Orchestrator to delegate work. It sends tasks with team send <worker-name> "<task>"; workers report back with team report.

For stronger automation, enable the experimental Workflows toggle in settings. The Orchestrator can then author and run multi-agent workflows that fan out across implementation, review, testing, or other stages. The topbar Workflows panel shows runs, phase results, logs, schedules, and stop controls. The same panel also lets you choose which CLI workflow-created agents use by default and which CLIs they are allowed to use.

How It Works

Browser UI on 127.0.0.1
  tasks, team, terminals, reports
          |
          | HTTP + WebSocket
          v
Hive runtime
  SQLite metadata, PTY lifecycle, task dispatch
          |
          +-- Orchestrator PTY
          |     can call: team send, team list, team report
          |
          +-- Worker PTY
          |     can call: team report
          |
          +-- Worker PTY
                can call: team report

Workspace task graph:
  <workspace>/.hive/tasks.md

Three details matter:

• Agents are real CLI processes, not simulated subagents.
• team is injected only inside Hive-managed agent sessions by prepending the package's internal bin directory to PATH; it is not installed as a global command.
• The task graph is a markdown file in the workspace, so you can inspect or edit it outside the app.

Agent Presets

Hive does not install thes