A local, operator-controlled meta-harness for AI coding sessions. Launch, supervise, and compare coordinated CLI agents (Claude, Codex, and others) without handing topology decisions to an opaque control plane.
- Hive Mode: An Opus Queen coordinates manager-launched coding principals, with capability-aware native delegation inside supported harnesses
- Fusion Mode: Parallel competing implementations with best-pick resolution
- Solo Mode: One directly supervised coding agent for focused work
- Legacy Swarm Compatibility: Existing Swarm sessions remain programmatically launchable outside the primary flow
- Session Persistence: Save and resume sessions across app restarts
- Multi-CLI Support: Works with Claude Code, Codex, OpenCode, and more
- Real-time Monitoring: Watch all agents work simultaneously with live terminal output
- Git Integration: Automatic branch management and coordination
Download the latest release from the Releases page:
- NSIS Installer:
Hive Manager_x.x.x_x64-setup.exe(recommended) - MSI Installer:
Hive Manager_x.x.x_x64_en-US.msi
Requirements:
- Node.js 18+
- Rust 1.70+
- Windows 10/11
# Clone the repository
git clone https://github.com/rdfitted/hive-manager.git
cd hive-manager
# Install dependencies
npm install
# Development mode
npm run tauri dev
# Production build
npm run tauri build- Launch Hive Manager
- Click New Session in the sidebar
- Select your project directory
- Choose a primary launch type (Hive, Fusion, or Solo)
- Configure the topology, workspace strategy, agents, and delegation policy
- Click Launch
The default managed topology. An Opus Queen coordinates coding principals that Hive Manager launches and displays. A direct new Hive starts with one generic Codex gpt-5.6-sol coding principal; built-in feature and bug templates can preconfigure backend and frontend specializations. The operator's CLI, model, and role selections are authoritative.
Launch multiple agents working on the same task in parallel. Compare approaches and pick the best solution.
Launch one agent directly when a managed multi-agent topology would add no value.
Swarm remains programmatically compatible for existing callers and sessions, but it is not part of the primary launch flow.
Hive Manager keeps two delegation layers explicit:
- Managed principals (macro layer) are launched, displayed, and supervised by Hive Manager. The operator chooses their CLI, canonical model ID, role, workspace, and delegation policy.
- Native children (micro layer) may be created inside a capable Claude or Codex harness. They inherit the parent's Assignment Contract and cannot expand its authority, path ownership, or delivery obligations.
shared_cell is the recommended workspace strategy for a new collaborative Hive; isolated_cell gives each managed principal an explicit worktree when the operator wants separation.
Native delegation policy is separate from capability inference. The current card comes from Hive Manager's CLI adapter profile, not a live binary/version probe: disabled always turns delegation off; auto permits only adapter-declared support; encouraged records explicit operator authorization without rewriting an unknown capability as supported. Optional child and depth values are carried into the assignment as guidance; hard concurrency enforcement remains owned by the native harness.
Canonical model IDs are gpt-5.6-sol and fable; GPT-5.6 Sol and Fable 5 are display names. Hive Manager normalizes the legacy Codex value gpt-5.6 at launch so sessions and templates saved by older builds keep working. Older models remain selectable. Built-in defaults are recommendations, never hidden overrides of operator choices.
When Master Planner is used, it is contract-only: it converts the objective into bounded Assignment Contracts and stops before implementation.
| CLI | Behavior | Notes |
|---|---|---|
| Claude Code | Action-Prone | Anthropic's official CLI. Supports native delegation; Opus is the recommended Queen model. |
| Codex | Explicit-Polling | OpenAI's CLI. Supports native delegation; the GPT-5.6 tiers are gpt-5.6-sol (flagship/deep, the recommended coding-principal model), gpt-5.6-terra (balanced), and gpt-5.6-luna (fast). Hive task activation uses a durable polling loop. |
| OpenCode | Explicit-Polling | Open-source alternative. |
| Qwen | Instruction-Following | Follows instructions literally, respects role boundaries naturally. |
| Droid | Interactive | TUI mode with /model command for model selection. |
| Cursor | Interactive | Runs via WSL. Uses global model setting. |
Behavior profiles guide CLI-specific prompt hardening. Capability cards separately report adapter-declared harness support; delegation policy records operator permission.
- Action-Prone: Proactive agents that need strong constraints to stay in their lane
- Instruction-Following: Literal interpreters that respect role boundaries naturally
- Explicit-Polling: Agents that need bash loops for coordination
- Interactive: TUI-based agents with different prompt injection
Sessions are stored in %APPDATA%/hive-manager/sessions/.
App configuration is in %APPDATA%/hive-manager/config.json.
# Run in development mode with hot reload
npm run tauri dev
# Type checking
npm run check
# Build for production
npm run tauri build- Frontend: SvelteKit 5, TypeScript
- Backend: Rust, Tauri 2
- Terminal: xterm.js with PTY support
Contributions are welcome! Please feel free to submit a Pull Request.
MIT License - see LICENSE for details.