Why do AI coding agents need a kanban board?

Agents work best with clear, scoped tasks. A kanban board gives them a queue to pull from, a place to report progress, and a shared source of truth with you. Instead of pasting context into every prompt, the agent reads the board over MCP and picks up the next task.

Which agents can connect to Kanboard?

Any MCP-compatible client: Claude Code, Claude Desktop, Codex, Cursor, Windsurf, and more. Kanboard's MCP server is remote (Streamable HTTP), so there's nothing to install — one command to connect.

Can multiple agents work the same board?

Yes. Each agent connects with the same API key (or its own key) and sees the same board. A common pattern is one board per repo: you triage into To Do, agents pull tasks into In Progress, and completed work lands in Done for your review.

How is this different from giving an agent a markdown TODO file?

A TODO file lives in one repo and drifts out of date. A kanban board is shared across sessions, machines, and agents, has real state transitions, and gives you a visual overview of what every agent is doing — without opening a terminal.

How do I stop an agent from doing destructive things on the board?

Agents act through API keys you create in Kanboard's settings and can revoke at any time. MCP clients like Claude Code also ask for permission before tool calls, so you approve board writes until you trust the workflow.

Kanban for AI Agents - A Task Board Claude Code Can Use

How to use a kanban board as the coordination layer between you and your AI coding agents — plan the work, dispatch agents over MCP, and track progress across columns.

Agents changed how work gets planned

AI coding agents like Claude Code and Codex can implement whole features from a well-written task. The bottleneck has moved from writing code to planning and reviewing it. That makes your task board more important, not less — it's where you decide what agents work on, and where you see what they did.

But most project management tools were built for humans clicking through UIs. Agents need programmatic access: a way to read the board, claim a task, and report progress. That's exactly what the Model Context Protocol (MCP) provides.

The workflow

Create a board for your project. One project per repo works well. Keep the default columns: To Do, In Progress, In Review, Done.
Connect your agent over MCP. Create an API key in Settings → Integrations, then run: claude mcp add --transport http kanboard https://kanboard.io/api/mcp --header "Authorization: Bearer YOUR_KEY". Codex and Cursor connect with a similar one-liner.
Plan the work yourself. Break the feature into small, scoped tasks in To Do — the same discipline that helps human teams helps agents even more. Put acceptance criteria in the task description.
Let agents pull tasks. Tell your agent: 'Check the board and pick up the top task in To Do.' It calls get_board, moves the task to In Progress, does the work, and moves it to Done.
Review in the In Review column. Have agents park finished work in In Review instead of Done. You review the diff or PR, then move the card yourself — the column becomes your approval gate.

Why kanban fits agents so well

Pull-based: agents pull the next task when they're free — no sprint ceremonies, no assignment overhead.
Small, scoped tasks: the same task-splitting discipline that keeps humans unblocked keeps agents on track and reviewable.
Visible state: columns show you at a glance what every agent is working on across sessions and machines.
An approval gate: the In Review column separates “the agent thinks it's done” from “you agree.”

Set it up in two minutes

Kanboard ships a remote MCP server with tools for listing projects, snapshotting boards, and creating, updating, and moving tasks. Follow the MCP setup guide to connect Claude Code, Codex, or Cursor with one command. Client pages: Claude Code kanban, Codex kanban, Cursor kanban. The hosted-board model is on AI agent board.