The best task trackers for AI agents

We make Hamloy, so read this knowing that. Every fact about the other tools comes from their own sites and docs, read on 10 October 2026, and each has a full comparison where we wrote one.

What matters when agents do the work

A tracker for people asks: is it easy to write, easy to find, easy to report on? With agents doing the work, other questions come first:

1. Can every agent reach it? Claude Code, Codex, Cursor, an app in the browser. A command line, an MCP server, or a file.
2. Do two agents ever take the same task? With two sessions at once, they will, unless something says "taken".
3. Where do the decisions go? An agent that needs you should leave a question you can answer later, not a line in a terminal you closed.
4. What does done mean? Ticked, or merged and pushed.
5. Where do you look? A file, a terminal, a board, a phone.

At a glance

The eight

Hamloy: for several agents and you on one list

One project's tasks, requests and questions, shared by every agent on it through 12 MCP tools. A task is claimed before it's started, across Claude Code, Codex, Cursor and the rest. Decisions come to you as questions with options and the one the agent recommends; what only you can do comes as a request with the steps. Two agents editing the same file are warned, with who to message; no file is locked. The hamloy program follows each task through git, and you see it all on a live board (claims, questions, git).

Best for two or more agents at once, across tools, with a person who isn't always at the terminal. Not for you if you want it open source, offline or in the repo. Free for one project; Pro is €8 a month per person (pricing).

A TODO.md: for one agent on one branch

A checklist in the repo. Free, offline, reviewed with the code, readable by any agent. It has no claims, and each branch has its own copy, so two sessions drift apart. Best for one agent at a time. Hamloy vs a TODO.md.

Claude Code's task list: for the steps of one task

Claude Code's built-in TaskCreate, TaskList and TaskUpdate keep a checklist on your computer, shared between Claude Code sessions that set the same list ID. On newer models the tools are off unless you turn them on (tools reference). Experimental agent teams claim tasks with file locking inside one session. Best for keeping one session on track; it fits inside any of the others. Hamloy vs Claude Code's own to-do list.

Beads: for an open-source issue graph in the repo

A "distributed graph issue tracker for AI agents, powered by Dolt", driven by the bd command line: atomic claims, many kinds of dependency, bd ready for what's unblocked, and issues labelled for a human (Beads). No board of its own. Best for people who want everything local, open and scriptable. Hamloy vs Beads.

Task Master: for turning a spec into a plan

Reads a PRD and writes tasks with dependencies and subtasks, using a model it calls with your key or through Claude Code or Codex signed in; keeps them in .taskmaster/tasks/tasks.json (Task Master). Best for starting a project from a written spec with one agent working through it. Hamloy vs Task Master.

Backlog.md: for Markdown tasks with a board in the terminal

Every task is a plain Markdown file in the repo. backlog board draws a board in the terminal and backlog browser serves one on your own machine; agents use it through the command line or MCP, and its setup can configure Claude Code, Codex, Gemini CLI, Kiro or Cursor. MIT (Backlog.md). Its README advises one task per agent session rather than describing claims. Best for people who want a board but want the tasks to stay files in git.

GitHub Issues: for a repo's public record

Free with every repository, with sub-issues, dependencies and Projects, and closed by the pull request that fixes it. Assign an issue to Copilot, Claude or Codex and it comes back as a draft pull request, on paid Copilot plans (GitHub Issues). Best for open source, bug reports and anything contributors should see. Hamloy vs GitHub Issues.

Linear: for a team planning a product

"The product development system for teams and agents": issues, projects, cycles and triage, an MCP server, and agents such as Codex, Cursor and Devin you delegate an issue to (Linear). Free up to 250 issues. Best for teams whose plan of record lives in a tracker. Hamloy vs Linear.

Which to pick

• One agent, one branch, you watching: a TODO.md or Claude Code's own list.
• You want it open source and in the repo: Beads, or Backlog.md for plain Markdown files.
• You start from a spec: Task Master.
• A team's plan, or a public project: Linear or GitHub Issues.
• Several agents at once, in different tools, and decisions that wait for you: Hamloy, often beside Linear or GitHub Issues rather than instead of them.