Agent Kanban

agent-kanban

Local-first kanban for AI-agent workflows. SQLite, FastAPI, no auth, no cloud.

A self-hosted kanban board designed to be driven by AI coding agents (Claude Code, Cline, opencode, Open WebUI, any function-calling LLM). Drag-drop in browser, REST + MCP for agents. Multi-project, light/dark theme, markdown plan-file import, automation rules.

Dark theme	Cyberpunk profile	Horizon profile

Toggle theme with t, cycle profiles with p. Or pin a theme/profile via URL: ?theme=light&profile=cyberpunk.

Features

• 9 workflow columns: Backlog → Approved → Analyst → In progress → Testing → UAT → Done, plus Blocked / Cancelled.
• Multiple projects in one DB, URL-routed (/p/{slug}), each with optional Claude Code-project directory binding.
• Light/dark theme switcher (t), density toggle (d), collapsible columns, sidebar.
• Search + filter chips, keyboard shortcuts (/, n, r, \, Esc).
• Inbox watcher: drop a .md into kanban_data/inbox/, get a card in 5 sec. Point KANBAN_INBOX_DIR at any folder you like — even ~/.claude/projects/<proj>/memory/inbox/ for Claude session captures.
• PLAN.md import (loose-mode parser, one-shot): any ## Section → cards in backlog with section-name in description.
• Automation rules (kanban_data/rules.json): task_moved, task_idle, blockers_done, task_count_in_status → move, comment, assign, set priority, archive, run a command. Polling rules act once per stay in a column; WIP limits, a pause switch. Hot-reloaded.
• Agent launcher (examples/agent-launcher/): card lands in Approved → headless Claude Code works it to Testing, with a timeout, one run per card, max_concurrent / max_runs, and the card blocked with a reason when a run goes wrong.
• Outbound webhooks (kanban_data/webhooks.json): POST to Slack / Telegram / any HTTP endpoint on task_created/task_moved/task_commented/task_updated. 3 formats: generic, slack, telegram.
• MCP server (kanban_mcp/) with 16 tools for Claude Code / Cline, strict arguments and compact answers; agents can't close cards themselves (configurable). python -m kanban_mcp.connect <dir> --project <id> writes the .mcp.json + CLAUDE.md block.
• One event stream: every change — UI, REST, MCP agent, script, rule — lands in the task history, and rules and webhooks react to it the same way.
• REST API + auto-generated OpenAPI for opencode / Open WebUI / any LLM with function calling. Scripts sign their changes with X-Kanban-Actor.
• Archive instead of delete: archived cards leave the board but keep their status. Links to single cards: /t/T-042.

Quickstart

5 minutes from clone to an AI agent doing your work for you: see QUICKSTART.md.

Bare minimum (one command after install):

# Install uv once: `curl -LsSf https://astral.sh/uv/install.sh | sh` or `brew install uv`
git clone https://github.com/<your-user>/agent-kanban.git
cd agent-kanban
uv run python -m kanban_ui                     # http://localhost:7777

uv run auto-creates a .venv/, installs deps from pyproject.toml (~5 sec on a warm cache), and launches the server.

python3.12 -m venv .venv
.venv/bin/pip install -r requirements.txt
.venv/bin/python -m kanban_ui                  # http://localhost:7777

requirements.txt is kept in sync with pyproject.toml for Dependabot and pip-only environments.

Optional — auto-start on login (macOS):

bash scripts/install_launchd.sh install        # macOS launchd
bash scripts/install_launchd.sh status
bash scripts/install_launchd.sh uninstall

Logs: ~/Library/Logs/agent-kanban/{stdout,stderr}.log.

AI agent integrations

Pick the integration matching your agent:

The kanban serves MCP over stdio (python -m kanban_mcp, for Claude Code / Cline), streamable HTTP at /mcp (Cursor, new Cline, MCP Inspector) and SSE at /sse for clients that only speak the older transport.

Static OpenAPI schema: docs/openapi.yaml. Interactive Swagger: http://localhost:7777/docs.

Real-world flows: docs/USECASES.md — 11 use cases (solo dev, team Slack, legacy import, multi-project, agent session, auto-launch, etc.).

Why

• vs. Trello/Jira/Linear — local-first; no SaaS account, no rate limits, your data on your disk.
• vs. plain TODO.md — drag-drop, history, multi-project, MCP integration, automation rules.
• vs. building your own — it has been run daily with Claude Code agents across several projects; the CHANGELOG for v0.2 lists what that surfaced and how it was fixed.

Configuration (env vars)

Project layout

agent-kanban/
├── kanban_store/    SQLite store + schema (5 migrations) + maintenance CLI
├── kanban_ui/       FastAPI web UI + automation/{events,inbox,rules,webhooks,plan_md}
│   └── static/      index.html · styles.css · app.js · vendor/Sortable
├── kanban_mcp/      MCP server (16 tools) + connect helper
├── seed/            generic example tasks
├── examples/        llm-tool-calling demo scripts
├── docs/            INTEGRATION.md, openapi.yaml, screenshots
├── tests/           pytest tests
├── scripts/         install_launchd.sh
├── snapshots/       JSON snapshots (gitignored except .gitkeep)
└── kanban_data/     runtime DB, inbox/, rules.json (gitignored)

Maintenance

python -m kanban_store.maintenance doctor             # health report, read-only
python -m kanban_store.maintenance dedupe-history     # dry run; --apply to delete (backup first)
python -m kanban_store.maintenance restore-archived   # done cards an old rule relabelled "cancelled"
python -m kanban_store.maintenance vacuum             # reclaim space (stop the server first)

Upgrading from 0.1.x: see CHANGELOG.md.

Roadmap

• Plan-file sync that survives edits on both sides (today: one-shot import).
• GitHub Issues importer (already has source-config table; needs gh API client).
• Multi-user mode with simple cookie-based auth.
• Optional Postgres backend.

License

MIT