Turns product requirements documents into dependency-aware task lists that AI coding agents can execute step-by-step.
Also known as stephennmiller/taskgraph-ai
- Stars
- 4
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 15, 2026
- License
- —
What is taskgraph-ai MCP?
taskgraph-ai turns product requirements documents into dependency-aware task lists that AI coding agents can execute step-by-step.
Summary generated from the project's README
About taskgraph-ai MCP server
Converting product requirements documents into structured, dependency-aware task lists is what TaskGraph enables for developers using Claude Code. It automates the transition from high-level planning to granular execution by generating tasks, subtasks, priorities, and test strategies directly from your documentation. The system tracks progress and dependencies within your repository, ensuring that coding agents always know which task is unblocked and ready for implementation.
This tool is designed to work locally within your existing development environment. It requires Claude Code to be installed and signed in, but it does not require any additional API keys because all AI calls run through your authenticated session. You can manage the workflow using the taskgraph CLI, slash commands like /taskgraph:next, or by adding the taskgraph-ai MCP server to give Claude direct access to its toolset. It is particularly useful for long-running builds that need to be paused and resumed, as the task statuses and progress notes are stored in your project files.
What you can do with it
- 1Generate a development planTransform a detailed product requirements document into a sequence of dependent tasks and subtasks.
- 2Resume interrupted coding sessionsPick up exactly where a previous development session ended by reading the stored task statuses.
- 3Identify the next priorityDetermine which specific task is currently unblocked and ready to be implemented by the AI.
Written from the project's README. Every use case is checked against a quote from it.
How to install taskgraph-ai MCP
Formats verified 10 Nov 2026Add to CursorAdd to GooseInstall taskgraph-ai MCP in Claude Code
$ claude mcp add taskgraph-ai -- npx -y taskgraph-aiInstall taskgraph-ai MCP in Cursor
~/.cursor/mcp.json
{
"mcpServers": {
"taskgraph-ai": {
"command": "npx",
"args": [
"-y",
"taskgraph-ai"
]
}
}
}Install taskgraph-ai MCP in VS Code
.vscode/mcp.json
{
"servers": {
"taskgraph-ai": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"taskgraph-ai"
]
}
}
}Install taskgraph-ai MCP in VS Code CLI
$ code --add-mcp '{"name":"taskgraph-ai","type":"stdio","command":"npx","args":["-y","taskgraph-ai"]}'Install taskgraph-ai MCP in Claude Desktop
claude_desktop_config.json
{
"mcpServers": {
"taskgraph-ai": {
"command": "npx",
"args": [
"-y",
"taskgraph-ai"
]
}
}
}Install taskgraph-ai MCP in Codex
$ codex mcp add taskgraph-ai -- npx -y taskgraph-aiInstall taskgraph-ai MCP in Codex config.toml
~/.codex/config.toml
[mcp_servers.taskgraph-ai] command = "npx" args = ["-y", "taskgraph-ai"]
Install taskgraph-ai MCP in Gemini CLI
$ gemini mcp add -s user taskgraph-ai npx -- -y taskgraph-aiInstall taskgraph-ai MCP in Gemini settings.json
~/.gemini/settings.json
{
"mcpServers": {
"taskgraph-ai": {
"command": "npx",
"args": [
"-y",
"taskgraph-ai"
]
}
}
}Automated checks
Ran Nov 14, 2026Full report →Passed our automated checks. That is not a security audit: read the code and give the server the narrowest access it needs. How checks work · What to fix for more points
- Source repository linked and reachablestephennmiller/taskgraph-ai is public and reachablegithub.com/stephennmiller/taskgraph-ai ↗Pass
- Package names this repository as its sourcetaskgraph-ai names stephennmiller/taskgraph-ai as sourcenpmjs.com/package/taskgraph-ai ↗Pass
- Build provenance (Sigstore attestation)Sigstore attestation for taskgraph-ainpmjs.com/package/taskgraph-ai/v/0.63.48#provenance ↗Pass
- Open source license (OSI)No licenseFinding
- Commit or release in the last yearLast commit or release 2026-11-13github.com/stephennmiller/taskgraph-ai ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNo popular package one edit away from taskgraph-aiPass
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubTaskGraph turns a product requirements document (PRD) into a dependency-aware task list, then walks Claude Code through it one task at a time. You describe what to build, TaskGraph plans the work, and Claude Code implements it, with the plan and progress stored in your repo.
- Plan from a PRD: generate tasks and subtasks, with dependencies, priorities and test strategies.
- Always know what's next: ask "What should I work on next?" and get the first unblocked task.
- Just ask Claude: with the plugin installed, say what you want in plain
language. Slash commands (
/taskgraph:next) and thetaskgraphCLI do the same things when you'd rather type them. - No API keys: every AI call runs through your signed-in Claude Code.
- Resumable long builds: stop a build partway and
taskgraph looppicks up from the task statuses. See Pros and cons for what it costs.
Attribution: TaskGraph is based on Task Master by Eyal Toledano (@eyaltoledano) and Ralph Khreish (@RalphEcom). It is named
taskgraph-aiwithtaskgraphandtaskgraph-mcpbinaries, so it can sit next to the original, and it keeps the original license and copyright notices (see Licensing). Coming from Task Master, or already using TaskGraph's earlier slash commands? See Upgrading.
Pros and cons
On our benchmarks, long builds scored by hidden tests against
a real tool, TaskGraph ties or trails plain Claude Code
on score and costs more: about 60% more with one task per session. What it
buys is a plan you can read and review in the repo, and a clean resume: after
an interruption taskgraph loop comes close to its uninterrupted score, for
14% more than an uninterrupted run, against almost three times as much with
plain claude --continue.
Use TaskGraph when a build is long enough that you will stop and resume it, or when you want the plan and its progress kept in the repo. For a build that fits in one sitting, plain Claude Code is cheaper and just as good.
Contents
Pros and cons · Set up in Claude Code · How to use it · Reference · Upgrading · Privacy · Docs · Licensing
Set up in Claude Code
You need Claude Code installed and signed in, and Node.js 22.19 or later. To sign in, run Claude Code once:
claude
1. Install
npm install -g taskgraph-ai
This gives you the taskgraph command (the CLI) and taskgraph-mcp, the
optional MCP server. To try it without installing, run any command through npx:
npx -y -p taskgraph-ai taskgraph list
To build from source instead, see Working on TaskGraph itself.
2. Initialize your project
From your own project folder:
taskgraph init
This creates .taskgraph/ (tasks, docs, config) and a CLAUDE.md import with
TaskGraph's workflow guide (--no-claude skips the import). It does not install
slash commands: step 3 does. Don't run it twice; it only re-adds the same files.
3. Install the plugin
The plugin lets Claude act on plain requests about your tasks. It also adds 16
slash commands (as /taskgraph:<name>) and two agents (task-executor,
task-checker). In Claude Code, run:
/plugin marketplace add stephennmiller/taskgraph-ai
/plugin install taskgraph@taskgraph
/plugin install taskgraph-hooks@taskgraph
Then check the setup from your project folder. taskgraph doctor checks Node,
Claude Code and its sign-in, the project, its data files and tasks, the plugin
and the MCP server, and prints the fix under each one that failed
(taskgraph doctor --fix repairs the task problems that have one safe answer):
taskgraph doctor
4. (Optional) Add the MCP server
Most people can skip this. The plugin and the CLI cover everything; the MCP server gives Claude TaskGraph's tools directly instead of through the CLI.
claude mcp add taskgraph-ai --scope user \
--env TASKGRAPH_TOOLS=standard \
-- npx -y taskgraph-ai
Restart Claude Code and run /mcp to check that taskgraph-ai is connected.
standard loads 16 tools, enough for everything in
How to use it. The installation guide has
the shared .mcp.json setup, and MCP tool tiers the others.
How to use it
With the plugin installed, ask Claude in plain language. The quick reference lists the matching slash command and CLI command for each request.
1. Write a PRD
Create .taskgraph/docs/prd.md describing what you want to build. The more
detail you give, the better the tasks. Skipping the PRD is fine for small jobs:
ask Claude to add a task instead.
2. Generate tasks
Turn .taskgraph/docs/prd.md into tasks
Then, optionally, check how big each task is and break the big ones down:
Analyze the complexity of the tasks and expand the complex ones into subtasks
3. Work through them
Start each session by asking what's next:
What should I work on next?
Then have Claude do it:
Start task 3
When it's finished and checked:
Mark task 3 done
The loop is simple: next, implement, done, next. You can also ask why a task is blocked, to break a task into subtasks, to add or update a task, or to research something for the project.
Or let it build unattended
To plan a PRD and have Claude Code work through every task on its own, run:
taskgraph go .taskgraph/docs/prd.md
It runs parse-prd, then taskgraph loop, which works the tasks in order and
commits after each. If it stops partway, run taskgraph loop again and it
picks up from the task statuses. Add --trust to run with Claude Code's
permission prompts off.
Reference
On the docs site:
- Commands: every slash command and CLI command, a cheat sheet, tags, dependencies and statuses.
- MCP tool tiers: what
TASKGRAPH_TOOLSloads. - Models: the main, research and fallback roles.
- Troubleshooting:
doctor, sign-in, MCP connection problems and building from source. - Benchmarks: the full results behind Pros and cons.
Upgrading
Slash commands now come from the plugin, and Task Master projects import with
taskgraph init --from-taskmaster --yes. The upgrade guide has
the steps for every existing project, and Migrating from Task
Master what moves.
Privacy
TaskGraph has no account, server or telemetry. Tasks stay in .taskgraph/ in
your project. AI commands go through your own Claude Code to Anthropic; the
privacy page lists everything that leaves your machine.
Documentation
Tutorial · Command reference · Configuration · Task structure · Examples · Models · Upgrading · Plugin · Contributing · Changelog
Contributors
Licensing
TaskGraph is licensed under the MIT License with Commons Clause, the same license as Task Master. The Commons Clause does not grant the right to Sell the software.
Allowed: use it for any purpose (personal, commercial, academic), modify it, distribute copies, and sell products built with it when that is not Selling the software.
Not allowed: Selling the software. "Sell" means practicing any or all of the rights granted under the license to provide the software to third parties, for a fee or other consideration (including without limitation fees for hosting or consulting/support services related to the software), as part of a product or service whose value derives, entirely or substantially, from the functionality of the software.
See LICENSE and licensing details.
taskgraph-ai MCP questions
How do I add taskgraph-ai MCP to Claude Code?
Run claude mcp add taskgraph-ai -- npx -y taskgraph-ai in your terminal. Add --scope user to make it available in every project.
How do I add taskgraph-ai MCP to Cursor?
Open ~/.cursor/mcp.json and add the server under mcpServers, using the Cursor config shown in the install section above.
Does TaskGraph require API keys?
No, every AI call runs through your existing signed-in Claude Code session.
Can I resume a long build if it gets interrupted?
Yes, you can stop a build partway and use the loop command to pick up from the current task statuses.
How does TaskGraph handle privacy?
TaskGraph has no account, server, or telemetry, and all tasks are stored locally within your project's directory.