Tack

A self-hosted task board for AI agent teams. Model-agnostic REST API. Any agent framework, any model, any scale.

Pin your tasks, track your agents.

What is this?

Tack is a lightweight Kanban board designed for human-AI collaboration. Your AI agents create, update, and move cards through a simple REST API. You see everything in a drag-and-drop web interface. Drag a card to "Approved" and your agents pick it up. Need a decision? Agents queue structured options with clickable buttons. You click, they resume.

No vendor lock-in. No cloud dependency. SQLite backend. Self-host in 30 seconds.

Status: alpha. Developed and tested internally on sandboxed development machines. If you deploy this: inspect the code, run in a VM or isolated environment, and back up your data before upgrading. This has not been independently security audited. See SECURITY.md for details.

Auth note: If you set TACK_API_KEY, the server will require the key for all write operations via the API. However, the web UI does not currently send the API key with its requests. This means drag-and-drop, card creation, and other UI write actions will be rejected by the server when a key is set. Agent API calls (which include the key in headers) will work correctly. Web UI auth support is planned.

Quick Start

pip install fastapi uvicorn
python server.py

Open http://localhost:8795 in your browser.

Docker

docker compose up -d

Or build manually:

docker build -t tack .
docker run -d -p 8795:8795 -v tack-data:/data tack

For AI Agents

Give your agent the contents of TOOL.md as context. It contains:

• REST API reference with all endpoints
• Column and priority definitions
• OpenAI function-calling tool definitions (works with any compatible framework)
• Usage guidelines (when to create cards, when to move them)

The tool spec is model-agnostic. It works with OpenAI, Anthropic, Ollama, llama.cpp, or any system that supports function calling or HTTP tool use.

API

Method	Endpoint	Description
GET	/api/board	Get all cards organized by column
GET	/api/cards	List cards (filter by ?column=, ?assignee=, ?fields=)
POST	/api/cards	Create a card
GET	/api/cards/{id}	Get a single card
PATCH	/api/cards/{id}	Update card fields
POST	/api/cards/{id}/move	Move card to a column
POST	/api/cards/{id}/claim	Claim card (prevents other agents from grabbing it)
POST	/api/cards/{id}/release	Release a claim
POST	/api/cards/{id}/notes	Add a note to the card's trace log
GET	/api/cards/{id}/notes	Read card notes
POST	/api/cards/{id}/decision	Request a human decision (with structured options)
POST	/api/cards/{id}/decide	Submit a decision
GET	/api/decisions	List all pending decisions
POST	/api/cards/{id}/approval	Request approval for a plan (yes/no)
POST	/api/cards/{id}/approve	Submit an approval or denial
GET	/api/approvals	List all pending approvals
POST	/api/batch	Execute multiple operations in one call
DELETE	/api/cards/{id}	Delete a card
GET	/api/activity	Recent activity log
GET	/api/changes	Poll for changes since a timestamp
GET	/health	Health check

Agent-Native Features

• Sparse-field reads — ?fields=id,title,column_name returns only the fields you ask for. Cuts response size ~5x, saves tokens on every poll.
• Atomic claiming with TTL — POST /claim locks a card so two agents don't work on the same task. Claims auto-expire after a configurable TTL. 409 conflict with recovery hint if already claimed.
• Per-card trace log — Append-only notes on each card. Agents document reasoning, findings, and handoff context. Makes agent-to-agent handoffs possible without losing context.
• Human-in-the-loop decisions — Agents post structured questions with options. Humans see clickable buttons in the UI. Click to decide, card moves forward. GET /api/decisions gives you the decision queue.
• Plan approvals — Agents propose a plan, humans approve or deny. Simpler than decisions — no options to choose from, just yes/no with optional comments. Denied cards move to blocked with the reason logged. GET /api/approvals gives you the approval queue.
• Batch operations — POST /api/batch with an array of actions. Create, move, update, claim, release, and note in a single call.
• Webhooks — Planned but not yet implemented. See Spur for webhook event relay in the meantime.

Columns

Column	Purpose
Inbox	New tasks, not yet reviewed
Approved	Human approved — agents pick these up
In Progress	Actively being worked on
Awaiting Decision	Needs human input (with structured options)
Awaiting Approval	Agent proposed a plan, needs yes/no
Deferred	Postponed — auto-returns to Awaiting Decision after TACK_DEFER_DAYS
Done	Completed
Blocked	Cannot proceed

Configuration

Environment variables:

Variable	Default	Description
TACK_HOST	127.0.0.1	Bind address
TACK_PORT	8795	Port number
TACK_DB	./data/board.db	SQLite database path
TACK_DONE_ARCHIVE_DAYS	7	Auto-hide done cards older than N days from board view
TACK_DEFER_DAYS	7	Days before deferred cards auto-return to Awaiting Decision
TACK_API_KEY	(none)	Optional API key for write operations (reads remain open)

Keyboard Shortcuts

• n — New card
• / — Focus search
• Esc — Close modal / clear search
• Double-click card to edit

Known Limitations

• Web UI does not support API key auth. If TACK_API_KEY is set, the web UI cannot perform write operations (create cards, drag-and-drop, edit). API clients that send the key in headers work fine. Fix planned for next release.
• Webhooks not yet implemented. See Spur for webhook event relay.
• No built-in HTTPS or rate limiting. Use a reverse proxy for production deployments. See SECURITY.md.

Dependencies

• Python 3.10+
• FastAPI
• Uvicorn
• SortableJS (loaded from CDN in the frontend)

License

MIT