Trackboi Review: Markdown-Powered Kanban Built for AI Coding Agents

Every AI coding agent forgets. Close a Claude Code session, reopen it the next morning, and the model has no idea which feature you were three-quarters of the way through. Teams usually patch this with a Linear board or a Notion page — but those live outside the repository, behind an API the agent reaches unreliably, if at all. Trackboi takes a different bet: store every task as a plain markdown file committed inside the project, where a coding agent already has full read and write access.

We dug into the project to see whether “markdown as a database” holds up once a real coding agent is driving the board.

What Trackboi actually is

Trackboi is an open-source Kanban tool with no database, no hosted dashboard, and no separate app to keep running. The board is a set of markdown files in your repository. Each task is a file, and its status — backlog, in progress, done — is recorded in the file itself rather than in an external system. Moving a task across the board means editing text and committing it.

That design has one immediate consequence: the board is versioned by git, exactly like your code. Every status change is a diff. Every completed task carries an author and a timestamp in the commit log. There is no separate sync step and nothing to reconcile, because the board is not a copy of the truth — it is the truth, sitting in git next to the code it describes.

For a human, that is mildly convenient. For an AI agent, it removes an entire category of failure. The agent does not need an MCP server, an OAuth token, or a brittle screen-scrape to learn what to work on. It reads a file, the same way it reads everything else in the repo.

Why plain markdown wins for agent handoffs

The pitch only lands if you have watched an agent lose the thread mid-task. Here is what file-based task tracking fixes:

No integration to break. Every coding agent — Claude Code, Cursor, Codex — can already open, search, and edit files. A tracker built from files needs zero adapter code. Wire an agent into a SaaS board instead and you are one API change away from a silent break.

The board is grep-able. An agent starting cold can search the task directory and reconstruct state in a single step: what is in progress, what is blocked, what shipped last. No pagination, no rate limits, no auth dance.

Git history is the audit trail. When two agents — or an agent and a human — hand work back and forth, the commit log already records who changed which task and when. You get provenance for free, which matters more as a larger share of the diff is machine-written.

It plugs into AGENTS.md. The AGENTS.md convention places a single instructions file at the repo root that compliant agents read before acting. A file-based board is its natural companion: point AGENTS.md at the task directory and tell the agent to update task status as part of finishing work. Instructions and work queue then live in the same tree, in the same format.

The honest counterpoint: a markdown board has no live UI, no notifications, and no burndown chart. If your team runs on a shared visual board that non-technical stakeholders watch, that is a real loss. Trackboi is built for people whose tasks are already born and buried inside a git repo.

Trackboi vs Vibekanban

Both tools answer the question “how do I give an AI agent a task board,” but they answer it from opposite ends.

Vibekanban is an application. You run it, and it hands you a Kanban UI for dispatching and supervising agent runs — often several in parallel. The board is a control surface you watch. Trackboi is a file format. There is no app to launch; the board is whatever your editor and git already show you. The board is a data structure you commit.

Neither is strictly better. If your day is spent orchestrating multiple agents at once and watching them work, an application with a real UI earns its keep. If you mostly want one agent to resume cleanly and leave a paper trail inside the repo, the file-based approach has less machinery to go wrong and nothing to host.

Should you adopt it

Trackboi is a young project, and “young open-source tool” carries the usual caveats: expect rough edges, read the open issues before you depend on it, and do not assume the file format is frozen.

The core idea, though, is sound and not really tied to one tool: tracking agent work in plain, committed text is more robust than tracking it in a system the agent can only see through a straw. Even if you never install Trackboi itself, the pattern — one markdown task file per unit of work, updated as part of the diff — is worth stealing for any repo where agents do real work.