kanban-mcp

A database-backed kanban board that AI coding agents use via MCP (Model Context Protocol). Track issues, features, todos, epics, and diary entries across all your projects — with a web UI for humans and 40+ tools for agents.

What It Does

• Persistent project tracking — issues, features, todos, epics, questions, diary entries stored in SQLite (default) or MySQL/MariaDB

• Status workflows — each item type has its own progression (backlog → todo → in_progress → review → done → closed)

• Relationships & epics — parent/child hierarchies, blocking relationships, epic progress tracking

• Tags, decisions, file links — attach metadata to any item

• Semantic search — find similar items using local ONNX embeddings (optional; downloads nomic-embed-text-v1.5 from HuggingFace on first use, ~140MB — the first query will block until download completes)

• Activity timeline — unified view of status changes, decisions, updates, and git commits

• Export — JSON, YAML, or Markdown output with filters

• Web UI — browser-based board at localhost:5000

• Session hooks — inject active items into AI agent sessions automatically

Related MCP server: kanban-lite

Quick Start

Requires Python 3.10+.

Linux / macOS:

curl -fsSL https://raw.githubusercontent.com/multidimensionalcats/kanban-mcp/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/multidimensionalcats/kanban-mcp/main/install.ps1 | iex

The script prompts interactively (backend choice, etc.) then installs pipx and kanban-mcp, sets up the database, runs migrations, and prints your MCP config. SQLite is the default — just press Enter. No database server required.

Want MySQL/MariaDB instead? Add --mysql (or -MySQL on Windows):

curl -fsSL https://raw.githubusercontent.com/multidimensionalcats/kanban-mcp/main/install.sh | bash -s -- --auto --mysql

Manual install (no script):

pipx install kanban-mcp
kanban-cli --project "$(pwd)" summary   # SQLite DB auto-created on first run

Then add the MCP server to your AI client — see MCP Client Setup.

Prerequisites

• Python 3.10+

• pipx (recommended) — installed automatically by the install script if missing

• MySQL 8.0+ or MariaDB 11+ (optional) — only needed if you choose MySQL over the default SQLite backend

Installation

The Quick Start one-liner is the fastest path. Below are alternative install methods and additional options.

Option 1: pipx (recommended)

pipx installs into an isolated virtualenv while making commands globally available. This avoids PEP 668 conflicts on modern distros and ensures hooks work outside the venv.

# SQLite backend (default, zero dependencies)
pipx install kanban-mcp

# With semantic search
pipx install kanban-mcp[semantic]

# With MySQL backend
pipx install kanban-mcp[mysql]

# Everything (MySQL + semantic)
pipx install kanban-mcp[full]

The SQLite database is created automatically on first run — no extra setup needed. For MySQL, see Database Setup.

Upgrade later with:

pipx upgrade kanban-mcp

Option 2: pip

pip install --user kanban-mcp

Note: On modern distros (Debian 12+, Fedora 38+, Arch, Gentoo), bare pip install is blocked by PEP 668. Use --user, --break-system-packages, or prefer pipx.

Option 3: From source (development)

git clone https://github.com/multidimensionalcats/kanban-mcp.git
cd kanban-mcp
pip install -e .[dev]

Note: If PEP 668 blocks the install, use a venv: python3 -m venv .venv && source .venv/bin/activate first. Be aware that hooks run via /bin/sh, not the venv Python — you'll need to use full paths to the venv's console scripts in your hook configuration.

Option 4: Docker (MySQL/MariaDB + web UI)

Note: Docker compose runs MySQL/MariaDB, not SQLite. Use this if you want a containerized MySQL setup.

The install script can start MySQL/MariaDB via Docker for you (./install.sh --auto --mysql --docker or choose Docker when prompted). To run the compose stack manually:

1. Start the containers (MySQL 8.0 + web UI on port 5000):

   git clone https://github.com/multidimensionalcats/kanban-mcp.git
   cd kanban-mcp
   docker compose up

   Migrations run automatically on web container startup. Credentials are configurable: KANBAN_DB_USER=myuser KANBAN_DB_PASSWORD=secret docker compose up

2. Install the MCP server on the host — Docker only provides the database and web UI. MCP clients spawn the server as a subprocess, so it must be installed locally:

   pipx install kanban-mcp[mysql]

3. Configure your MCP client — see MCP Client Setup. The database is exposed on port 3306 so the host-side MCP server can connect.

Database Setup

kanban-mcp uses SQLite by default — no setup required. The database file is created automatically on first run at ~/.local/share/kanban-mcp/kanban.db (or $XDG_DATA_HOME/kanban-mcp/kanban.db). You do not need to run kanban-setup for SQLite — it is only necessary if you want a custom database path or MySQL.

kanban-setup --auto defaults to SQLite. To choose a custom path:

kanban-setup --auto --backend sqlite --sqlite-path /path/to/db

Note: kanban-setup --with-semantic installs the semantic search Python packages. This is only needed if you installed without [semantic] initially (e.g. pipx install kanban-mcp). If you already installed with kanban-mcp[semantic], you don't need this flag. Works with any backend.

MySQL/MariaDB (optional)

If you need MySQL/MariaDB instead of SQLite:

Automated (interactive)

kanban-setup

Prompts for database name, user, password, and MySQL/MariaDB root credentials (including root password), then creates the database, runs migrations, and writes credentials to ~/.config/kanban-mcp/.env.

Note: On Debian/Ubuntu, default-mysql-server installs MariaDB, which defaults to auth_socket for the root user. Socket auth only works when the OS user matches the MySQL user (i.e. running as OS root). For non-root users, provide the MySQL root password when prompted — this is the normal path.

Automated (non-interactive / AI agents)

The --auto flag skips all interactive prompts. Without it, kanban-setup will prompt for each value.

# With root password (most common)
kanban-setup --auto --backend mysql --mysql-root-password rootpass

# With explicit credentials via environment variables
KANBAN_DB_NAME=kanban KANBAN_DB_USER=kanban KANBAN_DB_PASSWORD=secret \
  MYSQL_ROOT_PASSWORD=rootpass kanban-setup --auto --backend mysql

# With CLI args
kanban-setup --auto --backend mysql --db-name mydb --db-user myuser --db-password secret

# Socket auth (only works when OS user matches MySQL user, e.g. running as root)
kanban-setup --auto --backend mysql

Important: MYSQL_ROOT_PASSWORD is required for non-interactive use unless you are running as OS root. Socket auth (auth_socket) only works when the OS user matches the MySQL user — this is uncommon outside of CI or Docker. On Debian/Ubuntu, MariaDB defaults root to auth_socket — set MYSQL_ROOT_PASSWORD or use the manual SQL setup below.

Install script reference

The install scripts can be run from the repo or downloaded standalone:

./install.sh                                        # interactive (asks backend, installs pipx/kanban-mcp)
./install.sh --auto                                  # non-interactive, SQLite (default, zero config)
./install.sh --auto --mysql                          # non-interactive, local MySQL/MariaDB (socket auth)
MYSQL_ROOT_PASSWORD=rootpass ./install.sh --auto --mysql  # non-interactive, MySQL with root password
./install.sh --auto --mysql --docker                 # non-interactive, MySQL via Docker
./install.sh --auto --mysql --db-host HOST           # non-interactive, remote MySQL
./install.sh --upgrade                               # upgrade existing Docker install

.\install.ps1                         # Windows interactive
.\install.ps1 -Auto                   # Windows non-interactive (SQLite)
.\install.ps1 -Auto -MySQL            # Windows MySQL
.\install.ps1 -Auto -MySQL -Docker    # Windows MySQL via Docker
.\install.ps1 -Auto -MySQL -DbHost HOST  # Windows remote MySQL
.\install.ps1 -Upgrade                # upgrade existing Docker install

Manual

Manual setup is a good alternative if database root auth is problematic (e.g. socket auth issues, restricted access).

-- As MySQL/MariaDB root user:
CREATE DATABASE kanban CHARACTER SET utf8mb4 COLLATE utf8mb4_unicode_ci;
CREATE USER 'kanban'@'localhost' IDENTIFIED BY 'your_password_here';
CREATE USER 'kanban'@'%' IDENTIFIED BY 'your_password_here';
GRANT ALL PRIVILEGES ON `kanban`.* TO 'kanban'@'localhost';
GRANT ALL PRIVILEGES ON `kanban`.* TO 'kanban'@'%';
FLUSH PRIVILEGES;

Note: On MariaDB, 'kanban'@'%' does not match localhost socket connections — you need both the @'localhost' and @'%' users.

Run the migration files in order:

mysql -u kanban -p kanban < kanban_mcp/migrations/001_initial_schema.sql
mysql -u kanban -p kanban < kanban_mcp/migrations/002_add_fulltext_search.sql
mysql -u kanban -p kanban < kanban_mcp/migrations/003_add_embeddings.sql
mysql -u kanban -p kanban < kanban_mcp/migrations/004_add_cascades_and_indexes.sql

Configuration

Credentials

kanban-setup writes database credentials to a .env file in the user config directory:

• Linux/macOS: ~/.config/kanban-mcp/.env (or $XDG_CONFIG_HOME/kanban-mcp/.env)

• Windows: %APPDATA%\kanban-mcp\.env

All install methods (pipx, pip, source) use this same location. You can also set credentials via environment variables or your MCP client's env block.

Precedence (highest to lowest): MCP client env block → shell environment variables → .env file. In practice, just use one method — the .env file from kanban-setup is simplest.

Warning: If you previously used MySQL and switch to SQLite, remove or rename the old .env file at ~/.config/kanban-mcp/.env. Leftover KANBAN_DB_USER/KANBAN_DB_PASSWORD/KANBAN_DB_NAME values will silently trigger MySQL auto-detection. Alternatively, set KANBAN_BACKEND=sqlite explicitly to override.

MCP Client Setup

The kanban-mcp server speaks JSON-RPC 2.0 over stdin/stdout (standard MCP STDIO transport). Any MCP client can use it. If kanban-setup already wrote your .env file, you only need the command — no env block required.

If you need to pass credentials explicitly (e.g. the client doesn't inherit your shell environment), add an env block:

"env": {
  "KANBAN_DB_HOST": "localhost",
  "KANBAN_DB_USER": "kanban",
  "KANBAN_DB_PASSWORD": "your_password_here",
  "KANBAN_DB_NAME": "kanban"
}

Claude Code

Add to ~/.claude.json (global) or .mcp.json (per-project):

{
  "mcpServers": {
    "kanban": {
      "command": "kanban-mcp"
    }
  }
}

Claude Desktop

Add to ~/.config/Claude/claude_desktop_config.json (Linux) or ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "kanban": {
      "command": "kanban-mcp"
    }
  }
}

Gemini CLI

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "kanban": {
      "command": "kanban-mcp"
    }
  }
}

VS Code / Copilot

Add to .vscode/mcp.json (per-project):

{
  "servers": {
    "kanban": {
      "command": "kanban-mcp"
    }
  }
}

Note: VS Code uses the key servers, not mcpServers.

Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.kanban]
command = "kanban-mcp"

Cursor

Add to .cursor/mcp.json (per-project):

{
  "mcpServers": {
    "kanban": {
      "command": "kanban-mcp"
    }
  }
}

Other MCP Clients

For any other MCP-compatible tool: point it at the kanban-mcp command with STDIO transport. With the default SQLite backend, no env configuration is needed. If using MySQL and the tool can't read the .env file (e.g. it doesn't inherit your shell environment), pass the KANBAN_DB_* variables via the client's env configuration.

Hooks

Hooks are what make the agent use the board automatically. Without them, the agent only interacts with kanban-mcp when you explicitly ask it to.

Two hooks ship as console scripts, installed alongside kanban-mcp:

kanban-hook-session-start — Runs at session start. Reads the project directory from the hook's stdin JSON (cwd field), looks up the project in the database, and prints any in-progress items. This output gets injected into the conversation context, so the agent knows what's active without being told.

kanban-hook-stop — Runs at session end. Lists items still in progress and suggests creating a diary entry, updating statuses, or adding progress notes. This nudges the agent (and you) to keep the board current.

Both hooks exit silently if the project isn't tracked or the database is unreachable — they never block a session.

Which clients support hooks?

¹ VS Code Copilot reads Claude Code's hook configuration — if you already configured hooks for Claude Code, VS Code Copilot will use them too.

Claude Code, Gemini CLI, and VS Code Copilot all use the same hook format. Gemini CLI also sets CLAUDE_PROJECT_DIR as a compatibility alias, so the kanban hooks work across all three without modification.

Configuration

Hooks run via /bin/sh (Linux/macOS) or cmd (Windows), which do not read shell profiles — you must use absolute paths.

Find your paths:

# Linux/macOS
which kanban-hook-session-start   # typically ~/.local/bin/kanban-hook-session-start
which kanban-hook-stop

# Windows
Get-Command kanban-hook-session-start | Select-Object -ExpandProperty Source
# typically C:\Users\<you>\pipx\venvs\kanban-mcp\Scripts\kanban-hook-session-start.exe

Merge into your client's settings file (~/.claude/settings.json for Claude Code and VS Code Copilot, ~/.gemini/settings.json for Gemini CLI):

Linux/macOS:

{
  "hooks": {
    "SessionStart": [
      { "hooks": [{ "type": "command", "command": "/home/you/.local/bin/kanban-hook-session-start" }] }
    ],
    "Stop": [
      { "hooks": [{ "type": "command", "command": "/home/you/.local/bin/kanban-hook-stop" }] }
    ]
  }
}

Windows:

{
  "hooks": {
    "SessionStart": [
      { "hooks": [{ "type": "command", "command": "C:\\Users\\you\\pipx\\venvs\\kanban-mcp\\Scripts\\kanban-hook-session-start.exe" }] }
    ],
    "Stop": [
      { "hooks": [{ "type": "command", "command": "C:\\Users\\you\\pipx\\venvs\\kanban-mcp\\Scripts\\kanban-hook-stop.exe" }] }
    ]
  }
}

If you already have hooks configured, add the kanban entries to your existing arrays — don't replace them.

Tip: install.sh and install.ps1 print a ready-to-use config snippet with your resolved paths after setup completes.

Usage

First session

Once installed and configured, open your AI agent in a project directory. The first time you do this:

1. The session start hook fires but exits silently (it doesn't know about this project yet)

2. The agent needs to call set_current_project with your working directory's absolute path — this auto-creates the project in the database

3. From this point, all kanban tools work against this project

4. At session end, the stop hook lists any in-progress items and suggests logging progress

On subsequent sessions in the same directory, the start hook injects your active items into the conversation automatically — the agent picks up where you left off.

There is no "create project" command. Projects are created implicitly the first time set_current_project is called for a directory. If the agent doesn't call it on its own, ask it to — or the hooks will handle project context once the project exists in the database.

Note: Paths are resolved (symlinks, ., ..) before hashing, so --project . and --project $PWD refer to the same project. The same applies to set_current_project.

Three interfaces

kanban-mcp provides three ways to interact with the same data:

MCP tools — 40+ tools the AI agent calls during conversation. This is the primary interface. The agent creates items, tracks dependencies, advances statuses, and logs progress as part of your normal workflow. You don't need to tell it to — the session hooks provide context and the agent uses the tools naturally.

Web UI — a browser-based kanban board for humans.

kanban-web                    # http://127.0.0.1:5000
kanban-web --port 8080        # custom port
kanban-web --host 0.0.0.0     # network-accessible (no auth — use with care)
KANBAN_WEB_PORT=8080 kanban-web  # port via env var

kanban-web runs in the foreground. To run it persistently, use a process manager (e.g. systemd, screen, tmux) or the Docker compose stack which includes the web UI.

macOS note: Port 5000 is used by AirPlay Receiver on modern macOS. If kanban-web fails to bind, use --port 5001 or set KANBAN_WEB_PORT=5001.

The board shows all status columns (backlog → todo → in_progress → review → done → closed) with drag-and-drop between them. Use the project dropdown in the header to switch between projects. Cards show priority, tags, epic membership, blocking relationships, and progress bars for epics.

CLI — kanban-cli for terminal queries, scripts, and exports. Every command takes --project PATH (the same directory path you'd use with your agent) and an optional --format {text,json}.

CLI reference

All commands take --project PATH with the absolute path to your project directory, and an optional --format {text,json}.

# What's currently being worked on?
kanban-cli --project /path/to/project active

# What's in the backlog?
kanban-cli --project /path/to/project todos

# Project overview — counts by type and status
kanban-cli --project /path/to/project summary

# Search items and updates
kanban-cli --project /path/to/project search "authentication"

# Semantic search (requires [semantic] extra)
kanban-cli --project /path/to/project semantic-search "auth problems" --limit 5

# Export the full board
kanban-cli --project /path/to/project export --format markdown
kanban-cli --project /path/to/project export --format json --metrics --relationships

# Get children of an epic
kanban-cli --project /path/to/project children 42 --recursive

# Context dump (what the session hooks use internally)
kanban-cli --project /path/to/project context

# Rebuild embeddings after bulk changes
kanban-cli --project /path/to/project rebuild-embeddings

Tip: Paths are resolved (symlinks, ., ..) before hashing, so --project . and --project $PWD refer to the same project.

Upgrading

Database migrations run automatically on startup — no manual migration step needed after upgrading.

pipx:

pipx upgrade kanban-mcp

pip:

pip install --user --upgrade kanban-mcp

From source:

cd kanban-mcp && git pull && pip install -e .[dev]

Docker:

./install.sh --upgrade
# Windows: .\install.ps1 -Upgrade

This re-downloads the latest Docker files, rebuilds the web image from PyPI, and restarts the containers. Migrations run automatically on container startup.

For AI Coding Agents

Claude, Gemini, GPT, Codex, and other agents: start here.

If your human has asked you to install kanban-mcp, follow these steps exactly. Do not improvise. The --auto flag handles all prompts.

1. Install and set up the database

The default path uses SQLite — zero config, no database server needed:

curl -fsSL https://raw.githubusercontent.com/multidimensionalcats/kanban-mcp/main/install.sh | bash -s -- --auto

If your human specifically wants MySQL/MariaDB, add --mysql:

# MySQL via Docker:
curl -fsSL https://raw.githubusercontent.com/multidimensionalcats/kanban-mcp/main/install.sh | bash -s -- --auto --mysql --docker

# Local MySQL (set MYSQL_ROOT_PASSWORD if needed):
curl -fsSL https://raw.githubusercontent.com/multidimensionalcats/kanban-mcp/main/install.sh | bash -s -- --auto --mysql

# Remote MySQL:
curl -fsSL https://raw.githubusercontent.com/multidimensionalcats/kanban-mcp/main/install.sh | bash -s -- --auto --mysql --db-host HOST

For Windows (PowerShell — downloaded to a file because irm | iex doesn't support passing flags like -Auto):

irm https://raw.githubusercontent.com/multidimensionalcats/kanban-mcp/main/install.ps1 -OutFile install.ps1
.\install.ps1 -Auto

This installs pipx (if needed), installs kanban-mcp, creates the database, runs migrations, and writes config to ~/.config/kanban-mcp/.env.

2. Add MCP config

Write the config entry for whichever tool you are:

The server entry (JSON example):

{
  "kanban": {
    "command": "kanban-mcp"
  }
}

No env block is needed — credentials are read from ~/.config/kanban-mcp/.env which the install script created in step 1.

3. Verify

kanban-cli --project /path/to/humans/project summary

4. Configure hooks (recommended)

Hooks are what make you use the board automatically. Without them, you only interact with kanban-mcp when the human asks you to.

• kanban-hook-session-start — Runs at session start. Reads the project directory, queries in-progress items, and injects them into the conversation so you know what's active.

• kanban-hook-stop — Runs at session end. Lists items still in progress and reminds you to update statuses or create a diary entry.

If your client supports hooks (Claude Code, Gemini CLI, VS Code Copilot, Copilot CLI, Cursor): resolve the absolute paths via which, read your client's hook config file, merge the kanban hook entries (preserving any existing hooks), and write back. Do not overwrite other hook entries. See the Hooks section above for per-client config formats and paths.

Entry Points

kanban-setup flags

MCP Tools Reference

Project Management

Item CRUD

Status Workflow

Progress Updates

Relationships & Hierarchy

Tags

File Links & Decisions

Search & Export

Timeline

Item Types & Workflows

Contributing

git clone https://github.com/multidimensionalcats/kanban-mcp.git
cd kanban-mcp
python3 -m venv .venv && source .venv/bin/activate
pip install -e .[dev]

# Run tests (uses in-memory SQLite by default, no setup needed)
pytest

# Run tests against MySQL (requires MySQL/MariaDB running)
KANBAN_BACKEND=mysql KANBAN_DB_HOST=localhost KANBAN_DB_USER=kanban KANBAN_DB_PASSWORD=secret KANBAN_DB_NAME=kanban_test pytest

# Run frontend JS tests (requires Node.js — optional, only touches web UI code)
npm install && npm test

Changelog

See CHANGELOG.md for release notes. Check your installed version with:

pipx list | grep kanban-mcp
# or
pip show kanban-mcp

License

MIT