Ever sat through an AI agent generating 15,000 lines of code in one shot, only to stare at a diff so massive you just say “looks fine” and merge blind?
Yeah, me too. And that’s exactly the problem Backlog.md solves.
I’ve been testing it for the past hour on a Ryzen 9 workstation, and honestly? This isn’t another Kanban tool. It’s an attention fragmenter for the AI coding era — and the 5,900 GitHub stars make more sense the longer you use it.
What Makes Backlog.md Different
But most project management tools (Linear, Plane, Notion) are designed for human task tracking. They assume you write the spec, you write the code, you move the card.
Backlog.md flips that. It’s designed for AI agents to write code, and for you to review it in manageable chunks. I covered a similar workflow in the Claude Mem review — same idea of checkpointing agent output before merging.
The core idea is the 3-checkpoint workflow:
- Spec — Agent writes the specification. You review and approve before any code is written.
- Plan — Agent writes the implementation plan. You check for architectural issues.
- Code — Agent generates the code. Each checkpoint produces its own PR with a digestible diff.
This turns a 15,000-line firehose into three 5,000-line reviews — each with a clear checkpoint where you can say “wrong direction, try again” before the agent burns tokens writing code you’ll never use.
Quick Start: 2 Minutes to Running
And installation is dead simple — one npm command:
npm install -g backlog.md
Took about 22 seconds on my machine. Then initialize a project:
cd your-project
backlog init "My Project" --defaults
And you’re in business. Tasks are plain markdown files in backlog/tasks/:
backlog task create "Add JWT auth" -d "Implement token-based auth for the API" --priority high
Your terminal becomes a Kanban board:
backlog board
And full-text search works out of the box:
backlog search jwt
Here’s what surprised me: tasks are stored as .md files with YAML frontmatter. That means they’re natively git-friendly — every task creation is a diff, every status change is a commit. No database, no API, no cloud sync. Just files.
How It Compares
| Feature | Backlog.md | Linear | Plane | Notion |
|---|---|---|---|---|
| Open Source (MIT) | ✅ | ❌ | ✅ | ❌ |
| Local-first (files) | ✅ | ❌ Cloud | ❌ Cloud | ❌ Cloud |
| AI Agent 3-checkpoint workflow | ✅ | ❌ | ❌ | ❌ |
| Tasks as Markdown files | ✅ | ❌ DB | ❌ DB | ✅ |
| Terminal Kanban (CLI) | ✅ | ❌ Web only | ❌ Web only | ❌ Web only |
| Git-native (diff/PR per task) | ✅ | ❌ | ❌ | ❌ |
Linear is a better team project management tool. Plane is a solid open-source alternative. Notion is a knowledge base with task lists bolted on.
But none of them are designed for what Backlog.md does: segmenting AI agent output into reviewable chunks.
What to Watch Out For
Windows users may hit a snag. The npm package uses optional platform-specific binaries, and on Windows x64 the native binary didn’t auto-install for me. Had to manually install [email protected] alongside the main package. Still, the Linux/macOS experience is reportedly smoother.
This isn’t a Jira replacement. If you need sprint planning, burndown charts, or multi-team roadmaps, Backlog.md is not your tool. It’s laser-focused on the AI agent workflow — spec → plan → code — and doesn’t try to compete with full-featured project management suites.
The MCP setup is cool but requires buy-in. Connecting Claude Code or Codex via the MCP connector is straightforward, but it assumes your team is already using AI coding agents heavily. If you’re still in the “occasionally ask ChatGPT” phase, the 3-checkpoint overhead might feel like overkill.
Bottom Line
Backlog.md is the first tool I’ve seen that treats AI coding agents as first-class citizens in the development workflow — not as a sidecar that you occasionally delegate tedious tasks to. Its 3-checkpoint system is exactly what every team using Claude Code or Codex needs to maintain quality control without becoming the bottleneck.
If your team generates more AI code than you can comfortably review, give it a try. It’s free, it’s MIT, and it takes about 2 minutes to see if the workflow clicks for you.