Best task managers for AI coding agents, tested

This is a roundup from someone who ships with these agents every day and makes one of the tools on the list. I’ve tried to score every tool the same way, including Hatchlog, and to say where each one loses. Prices and facts were checked against each product’s own site or repo on the date in the byline.

If you want X, pick Y

• Free and inside your repo, cross-platform: Task Master or Backlog.md. Pick Backlog.md if you also want a board.
• A visual board on a Mac that the agent keeps current, with per-task token cost: Hatchlog.
• A full team issue tracker in the cloud: Linear.
• Planning next to code you already host on GitHub: GitHub Projects.
• A weekend project where anything heavier is overkill: a plain TODO.md.
• Driving several agents in parallel from one board: Vibe Kanban — but it’s sunsetting, so weigh the maintenance risk first.

The criteria

Every tool is scored on the same seven things: its price model, the platforms it runs on, whether it’s local or cloud, which agents it supports, whether it has a board UI, whether it tracks per-task token cost, and how much setup it takes. The per-task cost column is the one most of these tools skip, and it’s the reason Hatchlog exists, so I’ve kept it honest: only Hatchlog reports it, and only for Claude Code.

Task Master

Free, source-available (MIT with the Commons Clause), and cross-platform. It breaks a request into tasks and hands them to your agent over MCP, with the tasks stored as files in your project. There’s no visual board and no per-task cost, and that’s the point: it’s the lightweight, editor-native option. If you’re on Windows or Linux, or you like living in files, it’s a strong default. Full comparison: Task Master vs Hatchlog.

Backlog.md

Also free and open source (MIT), also stored as Markdown in your Git repo, but with a board the other file-based tools don’t have: backlog board paints a Kanban board in your terminal, and backlog browser serves a local web board with drag-and-drop. Agents connect over MCP or a CLI-instructions file, so Claude Code, Codex, Cursor and Gemini CLI all work. No per-task token cost. If you want a board and want it free and in-repo, this is the one I’d try first. More in Backlog.md vs Hatchlog.

Vibe Kanban

Apache-2.0 and free, run with npx vibe-kanban, with a local web board built for running many agents at once: each task gets a workspace with a branch, a terminal and a dev server, and you review diffs in the UI. It switches between 10+ agents. The catch is a big one: bloop, the company behind it, is shutting down (announced 10 April 2026) and the project is continuing community-maintained, so I wouldn’t start a long-term workflow on it without accepting that. See Vibe Kanban is shutting down for how to move, or what to keep using if parallel-agent orchestration is your thing.

Linear

The most polished tool here, and the only one really built for a team. Free tier, paid plans from $10 per user a month billed yearly, and a hosted MCP server so agents can find, create and update issues. Two honest caveats for this audience: it’s cloud-hosted, so your issues live on Linear’s servers, and it’s an issue tracker first, which is more process than a solo developer driving one agent usually wants. If you already run a team on Linear, adding the agent connection is the obvious move. See Linear vs Hatchlog for the solo-developer angle.

GitHub Projects

Free with any GitHub account, with table, board and roadmap views built right on your issues and pull requests. If your work already lives on GitHub, planning next to it with nothing to install is hard to beat. Agents reach it through GitHub’s official MCP server or the API, mostly at the issue and pull-request level, so it’s less of an agent-native board than Backlog.md or Hatchlog, and there’s no per-task token cost. For a team on GitHub, it’s the free default.

A plain TODO.md

The honest baseline. A TODO.md is a Markdown file with checkboxes; it’s free, portable, offline, and versioned with your code. Any agent that can read and write files can use it with no setup. It has no board and tracks no cost, and you’re the one keeping it current, which is exactly where it breaks down once a project grows. For a weekend project, it’s genuinely enough, and I’d reach for it before installing anything.

Hatchlog

My tool, scored the same way. Hatchlog is a native Mac app with a board the agent reads and updates itself, and it shows the Claude Code tokens each task cost and what that would be at API prices. It connects Claude Code, Cursor, Codex and Claude Desktop in one click. Where it loses: it’s Mac-only, it’s paid (a one-time purchase, from $6), the token cost is measured for Claude Code only, and it’s not a team tracker the way Linear is. If you’re a solo developer on a Mac who wants to see the board and know what each feature cost, that’s the trade it’s built for.

If your real question is cost, not tasks, the tool to compare is a usage tracker, not a task manager: see ccusage and per-task cost. And if the problem is really an agent that forgets the plan between sessions, start with the free ways in keeping Claude Code on track across sessions before you pay for anything.