Task Hounds – a local-first, inspectable multi-agent dev workspace

Task Hounds, a local-first, inspectable multi-agent development workspace powered by OpenCode, has been released as an MIT-licensed open-source project on GitHub. The tool turns one human goal into a visible development loop where a Manager plans, a Worker implements, and a Reviewer verifies each task, with all directives, plans, todos, reports, and agent state stored locally in a SQLite-backed database. It offers a web dashboard, Windows desktop app, Docker support, and an experimental Android client, designed for developers who want agent autonomy without giving up control or context.

Work like a dog. Ship like a pack. A local, inspectable multi-agent development workspace powered by OpenCode. Website https://task-hounds.com · GitHub https://github.com/catowabisabi/task-hounds · Demo https://www.youtube.com/watch?v=pu-Rt8Ye4EQ&t=174s · Issues https://github.com/catowabisabi/task-hounds/issues Task Hounds turns one human goal into a visible development loop. Give the pack a Human Directive ; the Manager plans, the Worker implements, and the Reviewer checks the result before the next task begins. Unlike a black-box coding assistant, Task Hounds keeps the work inspectable. Directives, plans, todos, reports, agent state, and reusable OpenCode sessions are stored locally, while the dashboard shows what every agent is doing in real time. It is designed for developers who want agent autonomy without giving up control or context . | Role | Responsibility | |---|---| You | Set the durable project mission, add ideas, and redirect the work at any time. | Manager | Understand context, maintain the plan, and assign one concrete task at a time. | Worker | Implement the selected task and report files changed, tests, and known issues. | Reviewer | Inspect the result for bugs, UX problems, edge cases, and safety risks. | Chat | Let you discuss the project and interact with the system directly. | php flowchart LR H "Human Directive" -- M "Manager<br Plan & delegate" M -- W "Worker<br Implement" W -- R "Reviewer<br Verify" R -- M M -- D "Live dashboard<br Todos, reports & state" W -- D R -- D | Input | Meaning | Lifecycle | |---|---|---| HUMAN DIRECTIVE | The durable project or session mission. | Copied into each new session in the same project. The agent loop never edits or deletes it; only a human can change it. | HUMAN NEW THOUGHT AND SUGGESTION | Direction, questions, product taste, concerns, or ideas. | The Manager digests it, may turn it into todo items, then marks it processed while preserving its history. | HUMAN SUGGESTED NEW TASK OR ITEM | An explicit feature or work item. | The Manager adds it to the plan and todo system when appropriate, then marks it processed while preserving its history. | The loop runs on an explicit message contract: the Manager digests input, decides, plans, and releases exactly one task; the Worker implements and files a structured report; the Reviewer verifies and routes feedback back to the Manager. Machine-readable todo JSON is validated and repaired before any work is released. Full specification — complete loop, role/data-flow diagram, time sequence, and hard rules: docs/architecture/agent-loop-contract.md Local first — your workspace, database, runtime state, and logs stay on your machine. Inspectable by design — follow live thinking, tool activity, todos, reports, and review feedback. Persistent context — SQLite-backed project state and reusable role sessions survive across loops. Clear responsibilities — planning, implementation, and review are handled by separate agents. Human steerable — change direction with durable directives, thoughts, and suggested tasks. Multiple ways to run — web dashboard, Windows desktop app, Docker, and an experimental Android client. Open source — MIT licensed and ready to adapt. git clone https://github.com/catowabisabi/task-hounds.git cd task-hounds docker compose up Then open http://localhost:8765 http://localhost:8765 . The image builds the dashboard for you — no Python or Node setup required. Project data persists in ./data . - Windows recommended for the managed runtime and desktop build - Python 3.11+ - Node.js 20+ - npm git clone https://github.com/catowabisabi/task-hounds.git cd task-hounds .\installation.cmd pip install -r requirements.txt pip install . installation.cmd installs the pinned, managed OpenCode runtime used by Task Hounds. cd ui/web npm ci npm run build cd ../.. Copy-Item .env.example .env Task Hounds keeps the legacy POWER TEAMS environment-variable prefix for compatibility. Review .env.example before adding provider keys or exposing the API beyond localhost. task-hounds-serve --port 8765 Equivalent: python -m task hounds api --port 8765 . Open http://localhost:8765 http://localhost:8765 , create or select a workspace, write a Human Directive, then choose Start Loop or Run Once . By default, projects are created under ~/task-hounds-projects . Override with the TASK HOUNDS PROJECTS DIR environment variable. Existing installs that already use C:\task-hounds-projects keep working unchanged. Task Hounds will not begin autonomous work without a pending Human Directive. For more detail, see the Getting Started guide /catowabisabi/task-hounds/blob/main/docs/guides/getting-started.md . Task Hounds talks to models through OpenCode's Anthropic-compatible providers. Three are bundled — bring any one key and the pack runs: | Provider | Models | Env key | |---|---|---| | MiniMax default | MiniMax-M2.7 | OPENCODE API KEY MINIMAX | | Moonshot Kimi | kimi-k3 1M context , kimi-k2.7-code , kimi-k2.7-code-highspeed | OPENCODE API KEY KIMI | | Alibaba Bailian | Qwen3.x, GLM, Kimi K2.5 | OPENCODE API KEY BAILIAN | The default model for every role is minimax-coding-plan/MiniMax-M2.7 . Kimi K3 ships with a 1M-token context window and thinking on by default — a great fit for long agent loops. Get a key from the Kimi Open Platform https://platform.kimi.ai/console/api-keys , then in .env : OPENCODE API KEY KIMI=your kimi api key TASK HOUNDS OPENCODE MODEL=kimi-coding-plan/kimi-k3 Restart, and Manager, Worker, and Reviewer all run on Kimi K3. You can also paste the key in the dashboard's Runtime panel, or pick the model per role in the binding editor. Each role resolves its model independently, so you can put the strongest reasoner where it matters: TASK HOUNDS MANAGER OPENCODE MODEL=kimi-coding-plan/kimi-k3 TASK HOUNDS REVIEWER OPENCODE MODEL=kimi-coding-plan/kimi-k3 Worker stays on the default MiniMax-M2.7 Role bindings set in the dashboard stored in the DB override these env vars. docker build -t task-hounds . docker run --rm -p 8765:8765 -v "$ pwd /data:/app/data" task-hounds pip install git+https://github.com/catowabisabi/task-hounds.git task-hounds --port 8765 This installs the backend server and CLI. The web dashboard is served when running from a repo checkout or Docker; a pip-only install exposes the HTTP API. .\build exe.ps1 The portable Electron build is written to ui/desktop/dist/ . The experimental React + Capacitor client is in ui/mobile/ . It connects to the same backend and shares projects, sessions, todos, chat, and agent state. Private access through Tailscale Serve https://tailscale.com/docs/features/tailscale-serve is recommended; do not expose the backend directly to the public internet. See ui/mobile/README.md /catowabisabi/task-hounds/blob/main/ui/mobile/README.md for setup instructions. SQLite is the runtime source of truth for project sessions, directives, todos, reports, suggestions, and agent state. Compatibility files under core/runtime/ are local runtime mirrors and fallbacks. task-hounds/ ├── core/ │ ├── db/ SQLite schema, migrations, and backups │ ├── runtime/ Local runtime state and logs not committed │ └── task hounds api/ Backend: FastAPI server, agent workflow, OpenCode runtime │ ├── api/ HTTP API and dashboard server │ ├── db/ DB access layer │ ├── opencode/ Managed OpenCode runtime and client │ └── workflow/ Manager/Worker/Reviewer loop, contracts, repair ├── ui/ │ ├── web/ React + Vite dashboard │ ├── desktop/ Electr