VS Code Agent Kanban

Task management for the AI-assisted developer — plan, track, and implement inside VS Code without losing context.

Install from VS Code Marketplace  |  View on GitHub

Anyone who works regularly with AI coding agents will recognise the problem: you go deep into a feature with Copilot — exploring approaches, debating trade-offs, building shared understanding. Then you hit the context limit, clear chat, or come back a week later, and that entire history is gone. You're starting over.

Modern AI coding agents are powerful but stateless. Every session is a blank slate. Agent Kanban solves this by giving every task a persistent Markdown file that travels with your code, and a clear plan / todo / implement workflow that integrates directly with GitHub Copilot.

Youtube (Quick Demo)

VS Code Agent Kanban board and Copilot Chat

Main Features

  • Kanban board in the Activity Bar — visual overview of Todo, Doing, and Done lanes, configurable via board.yaml
  • Markdown task files — every task is a .md file with YAML frontmatter, committed to your repo alongside your code
  • @kanban chat participant — structured plan / todo / implement commands direct Copilot through the work
  • GitOps friendly — one file per task, diffable and mergeable, no opaque binary state

Markdown Files as the Source of Truth

Every task lives in a .agentkanban/tasks/ folder as a .md file. YAML frontmatter tracks the title, lane, and timestamps. The body is a structured conversation log using [user] and [agent] markers — the full planning and implementation history, in plain text:

---
title: Implement OAuth2
lane: doing
created: 2026-03-08T10:00:00.000Z
---

## Conversation

[user]
Let's plan the OAuth2 implementation...

[agent]
Here's my analysis of the two OAuth2 approaches...

This is intentionally simple and readable. No proprietary formats, no databases, no opaque state — just text files you can open, edit, search, and commit.

The Plan → Todo → Implement Workflow

Select a task with @kanban /task, then drive the agent with simple verbs in Copilot agent mode:

  • plan — the agent reads the task file and writes a structured plan back into it. Steer it with context: plan focus on error handling or plan assume we're already using Prisma
  • todo — generates a checkbox TODO list from the plan, written into a companion todo_*.md file
  • implement — the agent works through the plan and TODOs, making code changes using Copilot's built-in tooling

You can chain them: todo implement will generate TODOs and immediately start implementing. When work stalls or context gets unwieldy, clear chat, run @kanban /task again, and you're restored to exactly where you left off.

Works With Your Existing Tooling

Agent Kanban doesn't bundle its own agent harness. It contributes a @kanban chat participant that works with GitHub Copilot's native agent mode. That means:

  • Copilot's tool calling, diff previews, terminal commands, and file edits all use Copilot's native UI
  • Your existing AGENTS.md, skills, and custom instructions continue to work unchanged
  • You keep your existing Copilot subscription and workflow

GitOps Friendly by Design

The entire .agentkanban/ folder is designed to be committed to version control. This gives you:

  • Permanent task history — every decision, plan, and implementation conversation is preserved in your repository
  • Diffable, mergeable state — one file per task, board config in YAML; they diff and merge naturally
  • Shared team context — pull the latest and open the board; you're immediately up to date on what the AI is working on
  • Audit trail — a git-tracked record of the planning and decision-making process for AI-assisted development

The Kanban Board

The visual board in the VS Code Activity Bar gives you a bird's-eye view of all tasks. Default lanes are Todo, Doing, and Done — lanes are configurable in board.yaml. Tasks are created with @kanban /new My Task or with the + New Task button. Drag cards between lanes as work progresses. Tasks in the Done lane are excluded from /task selection, keeping the active list clean.

Getting Started

  1. Install Agent Kanban from the VS Code Marketplace
  2. Click the Kanban icon in the Activity Bar
  3. Create your first task: @kanban /new My Feature in Copilot Chat
  4. Select it: @kanban /task My Feature (fuzzy matched)
  5. In Copilot agent mode, type plan

The extension handles the rest — creating the task file, injecting instructions, and keeping the board in sync.