cd /news/ai-agents/task-hounds-a-local-first-inspectabl… · home topics ai-agents article
[ARTICLE · art-83438] src=github.com ↗ pub= topic=ai-agents verified=true sentiment=· neutral

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.

read5 min views1 publishedAug 2, 2026
Task Hounds – a local-first, inspectable multi-agent dev workspace
Image: source

Work like a dog. Ship like a pack.

A local, inspectable multi-agent development workspace powered by OpenCode.

Website · GitHub · Demo · 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.
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. 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, 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.

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, 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

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 is recommended; do not expose the backend directly to the public internet.

See 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
── more in #ai-agents 4 stories · sorted by recency
── more on @task hounds 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/task-hounds-a-local-…] indexed:0 read:5min 2026-08-02 ·