A task brain for AI coding tools, not another editor
Task Master (the npm package is task-master-ai) solves a problem that the coding agents around it do not: keeping a long build organized. The agents in Cursor, Windsurf, and Claude Code are good at writing the next file, but they lose the thread across a multi-day feature. Task Master takes a product requirements document, parses it into a numbered list of tasks and subtasks with explicit dependencies, and then tracks the state of each as work proceeds. It does not write your application code. It decides what to do next and remembers what is done.
The project is young, created in 2025, and already at 27,548 stars as of 2026-06, which tells you how much demand there is for structure on top of autonomous agents.
What it does
You point it at a PRD, run parse-prd, and it produces tasks.json: a structured task graph with parent and child tasks, priorities, and dependency links so the tool knows what is unblocked. From there you ask for the next task, expand a task into subtasks, mark things done, and move tasks between named tags (backlog, in-progress, and any you define) that act as separate lanes. The point is that this state is plain JSON in your repo, committed to git, so the plan is auditable and survives across sessions rather than living in an editor’s private memory.
Install
The common path is as an MCP server, so your editor’s AI can drive it from chat. For Claude Code:
claude mcp add taskmaster-ai -- npx -y task-master-aiFor Cursor, Windsurf, and VS Code, you add a task-master-ai entry to the editor’s MCP config (~/.cursor/mcp.json, the Windsurf config, or .vscode/mcp.json) with the command npx -y task-master-ai and your API keys in the env block. There is also a direct CLI:
npm install -g task-master-ai
task-master init
task-master parse-prd prd.txt
task-master nextThe multi-model setup, and the API keys you need
Task Master splits work across model roles: a main model for execution, a research model for gathering background (often a search-grounded model), and a fallback model when the main one fails. You configure which provider fills each role in .taskmaster/config.json. The practical consequence is that it expects API keys for whatever providers you assign, and the README documents support for around a dozen of them. You need at least one (Anthropic or OpenAI), but a research-heavy setup pulls in more. Budget for the fact that parse-prd, expand, and research each make model calls, so a large PRD can stack up cost quickly, and there is no built-in spend cap.
Where it fits, and where it doesn’t
Reach for Task Master when a feature is large enough that “just ask the agent” loses coherence, when you want the plan tracked in git rather than in an editor, or when several agents or sessions need a shared task list. It is complementary to the coding agents, not a replacement.
It is overkill for a small change you can describe in one prompt, and it adds a layer of configuration and cost that only pays off on longer work. One licensing point matters before you build a business on it: the code is MIT with the Commons Clause, so you can use and modify it freely but you may not sell it, including offering it as a hosted or paid service whose value comes mainly from its functionality. GitHub therefore classifies the license as non-standard rather than open source.
task-master versus spec-driven and editor-native planning
| Task Master | github/spec-kit | editor-native planning | |
|---|---|---|---|
| Stars | 27,548 | 111,962 | n/a |
| Shape | PRD to tracked task graph | spec-driven workflow | ad hoc, in-chat |
| State | tasks.json in git | spec files plus issues/PRs | editor memory |
| Multi-model | main, research, fallback roles | per your agent | per your agent |
| License | MIT with Commons Clause | MIT | n/a |
Counts are from GitHub as of June 2026. spec-kit is the spec-driven approach: you write a specification and drive implementation from it, which overlaps but emphasizes the spec rather than a live task graph with dependencies and state. Editor-native planning means just asking the agent to keep a plan, which works until the feature outgrows a single context window. Task Master’s niche is the persistent, dependency-aware task graph in between.
The star curve
For a 2025 project, the star curve rose fast, which is the signature of a tool that hit a real pain point as autonomous coding agents went mainstream and people needed a way to keep them on track. Treat the steep early slope as adoption interest rather than proof of maturity; the project is still in its 0.x line.
What the issue tracker warns you about
A few real issues are worth knowing. PRD parsing supports .txt and .md only; feeding it a .docx sends the raw binary to the model and produces garbage, so convert to plain text first. The tracker has reported data-integrity bugs where concurrent operations across tags could corrupt parent-task metadata, and where a status write could change a task id’s type and inject a bad parentId, so keep tasks.json under git and commit often so you can recover. The MCP server is also memory-heavy per process because it loads many provider SDKs at startup, which adds up if you run several editors at once. None of these are dealbreakers, but they are the difference between a smooth setup and a confusing afternoon.
Related
For the spec-driven alternative, see spec-kit. For the coding agents Task Master orchestrates, see cline, Aider, and OpenHands. For the skills and plugins ecosystem around Claude Code, see anthropics/skills. For what is trending, see LLM tooling, the daily digest, and the weekly report.
FAQ
Is Task Master free and open source? It is free to use and the source is public, but the license is MIT with the Commons Clause, which forbids selling it (including as a paid hosted service). That makes it source-available rather than OSI open source, so check the license before any commercial offering built on it.
How do I set up Task Master as an MCP server? Add a task-master-ai entry to your editor’s MCP config running npx -y task-master-ai, with your model API keys in the env block. For Claude Code, claude mcp add taskmaster-ai -- npx -y task-master-ai.
How many API keys does Task Master need? At least one (Anthropic or OpenAI). It assigns separate model roles for main, research, and fallback, so if you use a dedicated research model you will configure additional providers. The README documents around a dozen supported providers.
Task Master vs spec-kit, what is the difference? spec-kit drives implementation from a written specification. Task Master maintains a live, dependency-aware task graph with state tracked in tasks.json. They overlap on planning but emphasize different artifacts, and many teams use a coding agent alongside either.
Can Task Master parse a Word document as a PRD? No. It supports .txt and .md only. A .docx will be sent as raw binary and parsed incorrectly, so export your PRD to plain text or Markdown first.