tt-a1i/hive

What it solves

Hive provides a coordination layer for users who run multiple CLI-based AI agents (such as Claude Code, Gemini, or Qwen) locally. It eliminates the need to manually juggle multiple terminal windows by providing a single browser-based workbench where an Orchestrator agent can plan and delegate tasks to Worker agents.

How it works

Hive runs as a local daemon on 127.0.0.1 that manages real PTY (pseudo-terminal) processes of the agents installed on your machine. It injects a specialized team command into these agent sessions, allowing them to communicate via dispatching tasks and reporting progress. The state of the collaboration is tracked in a shared markdown task graph located at <workspace>/.hive/tasks.md, which can be edited both within the Hive UI and by external editors.

Who it’s for

Developers and power users who already have AI agent CLIs installed and authenticated on their local machines and want to coordinate a team of agents for complex workflows like parallel bug hunting, implementation and review cycles, or research and drafting.

Highlights

• Multi-Agent Orchestration: Use one agent as an Orchestrator to delegate work to specialized Worker agents (e.g., coder, reviewer, tester).
• Local-First Architecture: Runs entirely on your local machine; agents are real CLI processes, not simulations.
• Broad CLI Support: Includes presets for Claude Code, Codex, Gemini, OpenCode, Qwen, Pi, and others, with support for custom executables.
• Experimental Workflows: Ability for the Orchestrator to author and run multi-stage, multi-agent workflows with dedicated logging and controls.
• Remote Access: Optional end-to-end encrypted tunnel for accessing the local workbench from a mobile device via a self-hosted gateway.
• PWA Support: Can be installed as a Progressive Web App for a dock-anchored experience.