Agent Kanban
An AI-native kanban board MCP server where agents pull tasks via the Model Context Protocol. It provides tools for task discovery, claiming, progress updates, and review workflows.
README
Agent Kanban
English · Русский
An AI-native kanban board where agents (Codex, Hermes, etc.) pull tasks via
the Model Context Protocol. The board is passive — it never spawns or controls
agents. You point your agents at the board via MCP config, and they
self-serve tasks through get_next_task / claim_task.
Why?
Most agent orchestrators push work: a dispatcher decides who does what and sends it. Agent Kanban inverts this — the board is a passive MCP server. Agents discover and claim work when they're ready, exactly like a developer picks a ticket from a real kanban. This decouples "what needs doing" (the board) from "who does it and when" (the agent), and works with any agent that speaks MCP — no glue code per agent.
Features
- Pull-based MCP — board exposes 11 tools (
get_next_task,claim_task,post_progress,request_review, …) over the streamable-HTTP transport - Hard assignment — reserve a task for a specific agent; others don't see it
- Live progress feed — agents stream diffs, text, artifacts, errors; you comment back and the agent reads your feedback on the next turn
- Bilingual UI — Russian / English, with a one-click language toggle
- Review diffs — the board collects
git diff base...branchonrequest_reviewand renders it inline (Shiki syntax highlighting) - Design system — dark-first, single-amber-accent UI built on tokens
- Self-hosted — FastAPI + PostgreSQL + React; ships as a Docker image
Screenshots
<table> <tr> <td width="50%" align="center"><b>Board (dark)</b></td> <td width="50%" align="center"><b>Card detail</b></td> </tr> <tr> <td><img src="docs/screenshots/board.png" alt="Board view"></td> <td><img src="docs/screenshots/card-detail.png" alt="Card detail"></td> </tr> <tr> <td align="center" colspan="2"><b>First-run setup</b></td> </tr> <tr> <td align="center" colspan="2"><img src="docs/screenshots/login.png" alt="Login / setup screen" width="50%"></td> </tr> </table>
Architecture
See docs/superpowers/specs/2026-07-05-agent-kanban-design.md for the full
design. Phase 1 MVP scope is documented in
docs/superpowers/plans/2026-07-06-agent-kanban-phase1-mvp.md.
Quickstart (local dev)
Prerequisites
- Python 3.11+
uv(install:curl -LsSf https://astral.sh/uv/install.sh | sh)- Node.js 20+ and
pnpm - PostgreSQL 16+ running locally
1. Database
createdb kanban # or use docker: see docker-compose.yml
cp .env.example .env # DATABASE_URL defaults to port 5436 (matches docker-compose ak-pg)
uv run kanban migrate
2. Backend
uv sync --extra dev
uv run kanban serve
3. Frontend (separate terminal)
cd web
pnpm install
pnpm dev # http://localhost:5173, proxies to :7331
4. Point an agent at the board
Add to your agent's MCP config:
Codex (~/.codex/config.toml):
[mcp_servers.kanban]
url = "http://localhost:7331/mcp"
Hermes (~/.hermes/config.yaml):
mcp_servers:
kanban:
url: http://localhost:7331/mcp
Then instruct your agent: "Check the kanban board for tasks via get_next_task."
Agent workflow
- Create a task in the UI. It starts in
todo. - Drag it to the
READYcolumn. It's now available to agents viaget_next_task. - Instruct your agent (e.g. Codex, Hermes) to check the board:
- Call
get_next_taskto discover work. - Call
claim_taskto take it. - Call
post_progressto report what it's doing. - Call
request_reviewwhen ready for your review, orcomplete_taskwhen done.
- Call
- Watch progress events stream into the card detail view in real time.
- Add comments to give the agent follow-up instructions; the agent reads them via
get_comments.
Agents must pass their identifier as the agent argument to mutation tools
(claim_task, post_progress, complete_task, request_review, post_comment,
post_artifact, set_task_branch, set_task_pr). The board authorizes mutations
by checking claimed_by == agent.
Coding tasks (git/PR)
For tasks that touch a git repo, set repo_path and base_branch when creating
the task. The board does NOT create branches or PRs — your agent does that with
its own git tools. The board records what the agent reports and renders a review
diff.
Agent workflow for a coding task:
claim_task— receivesrepo_pathand (if set)base_branch.- Create a branch in
repo_pathwith the agent's git tool. set_task_branch(task_id, agent, branch)— record it so the UI shows it and the diff can be collected.- Commit work on that branch.
request_review(task_id, agent, summary)— the board runsgit -C <repo_path> diff <base>...<branch>once and stores the result as a diff event visible in the card's progress feed.- Open a PR with the agent's GitHub tool, then
set_task_pr(task_id, agent, pr_url, "open"). - When merged,
set_task_pr(task_id, agent, pr_url, "merged")thencomplete_task.
If repo_path, base_branch (or the project's default_branch), or branch is
missing, diff collection is skipped silently. If git diff fails, an error event
is recorded instead, but the review request itself still succeeds.
Authentication
The board requires authentication. Two kinds of principals:
- Users (humans): log in with username + password via the web UI. Sessions are signed cookies.
- Tokens (agents): opaque bearer tokens, managed in the Admin panel. Each token is bound to an
agent_name.
First run
On first startup with an empty database, the UI shows a setup screen: choose an admin username + password (8+ chars) and submit. Alternatively, set AGENT_KANBAN_BOOTSTRAP_ADMIN_PASSWORD to create the admin headlessly (for automation). After setup, log in and go to Admin → Tokens to mint tokens for your agents.
Pointing an agent at the board
Agents authenticate via a bearer token. In your agent's MCP config:
Codex (~/.codex/config.toml):
[mcp_servers.kanban]
url = "http://your-host:7331/mcp"
# Codex reads headers from config in newer versions; otherwise set the auth via env.
headers = { Authorization = "Bearer <your-token>" }
Hermes (~/.hermes/config.yaml):
mcp_servers:
kanban:
url: http://your-host:7331/mcp
headers:
Authorization: Bearer <your-token>
The token's agent_name MUST match the agent argument you pass to MCP tools. A token minted with agent_name=codex can call claim_task(agent="codex") but NOT claim_task(agent="hermes").
Production env vars
SESSION_SECRET— signing key for session cookies. REQUIRED in production; set a long random string.PUBLIC_URL— the public base URL (e.g.https://kanban.example.com). Controls cookieSecureflag.AGENT_KANBAN_BOOTSTRAP_ADMIN_PASSWORD— first-run admin password (optional; if unset, the UI setup screen creates the first admin).
Docker
docker compose up -d
Runs the app on :7331 with Postgres in an internal-only container (not exposed to
the host). The app connects to it over the compose network as postgres:5432.
For local dev against a host-visible Postgres, see the standalone container note
in "Database" above (port 5436).
Contributing
Contributions are welcome. The design spec and phased implementation plans live
under docs/superpowers/ — they're a good map of the architecture and intent.
# clone, then:
uv sync --extra dev # backend deps + test tooling
cd web && pnpm install # frontend deps
uv run pytest -q # 112+ tests
cd web && pnpm build # type-check + bundle
Please open an issue first for non-trivial changes, so we can align on scope.
Credits
The UI is built on a design system inspired by Sergey Dmitriev's brand book — amber as the single UI accent, dark-first, with semantic-only status colors.
License
MIT © Sergey Dmitriev
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.