Project Manager
An AI project manager you work with through conversation.
Brief Project Manager as you would a human colleague: explain the outcome, constraints, authority, and what is changing in the real world. It takes responsibility for the working plan—decomposing the outcome, coordinating dependencies and owners, maintaining schedules and risks, tracking evidence, and following changes through.
You set direction and make the business decisions that require your authority. Project Manager manages the project. You do not have to convert reality into card movements, field edits, statuses, dependencies, or reports.
Four-panel overview of working with Project Manager as an AI project manager
“The vendor API will not be available until September 15. The launch date cannot move. Show me the credible options.”
Project Manager traces the affected work, tests the plan against the constraint, updates what the facts support, and brings back the real tradeoff. One conversation can change tasks, dependencies, schedules, risks, decisions, evidence, and reporting together.
project reality → reasoning → coordinated change
What you get
- An AI project manager that builds and maintains a verifiable plan tied to your outcome and success criteria.
- A current view of progress: what changed, what can move, what is blocked, what threatens the outcome, and which decision is needed.
- Connected updates across tasks, dependencies, schedules, risks, and decisions when reality changes.
- Completion backed by evidence, with missing information called out instead of invented.
- Status reports shaped for operators, project managers, executives, or boards without changing the underlying facts.
- One coherent project record across people, delegated agents, external executors, and optional software workflows such as RPD.
- Delegated execution that is recorded rather than ephemeral: a run's branches, worktrees, and per-attempt cost are part of the project state, so an interrupted run resumes instead of restarting.
Project truth stays in a durable, versionable Markdown folder and is validated before changes are saved. Kanban and Timeline visualize that state; they do not become a second source of truth.
PMI alignment
Project Manager is PMBOK 7 principles-aligned with documented tailoring. It is not certified by PMI, and no tool can be — this describes how the skill is built, not an accreditation.
Tailoring is the load-bearing idea. PMBOK 7 makes tailoring a principle and PMBOK 6 states processes are selected per project, so leaving out an area is compliant only when the omission is a recorded decision. A project therefore declares each of the ten PMBOK 6 knowledge areas as applied or tailored out, and tailoring an area out requires a rationale. Reports name a tailored-out area as tailored out with that rationale — never as zero, absent, or on track.
What that buys you: cost, procurement, or stakeholder management can be genuinely absent from a small project without the plan looking negligent, while an area you do claim to run is enforced. Declaring an area out while configuring its module is a validation failure, so the declaration cannot become fiction.
Applied by default: integration (change control, re-verification), scope (success criteria and traceability), quality (acceptance and evidence), and risk. Available when wanted: assumption log, issue log, stakeholder register, lessons register, and closure records. Cost, Earned Value, and critical-path scheduling are not implemented — declare them tailored out, or manage them elsewhere and say so in the rationale.
User guides
- English user guide — manage through outcomes, events, constraints, evidence, and decisions.
- 中文使用指南 — 通过目标、事件、约束、证据和决策管理项目。
Studio
Kanban shows planned, ready, active, completed, deferred, and cancelled work.
Project Manager Studio Kanban view
Timeline shows schedules, dependencies, blockers, and date conflicts.
Project Manager Studio Timeline view
Studio watches the selected project's durable state and refreshes automatically when the CLI, an agent, or another editor changes it. Automatic refresh waits while a task form or schedule draft is open so local edits are not discarded; the Refresh button remains available as a manual recovery.
MCP App
Studio is a browser window. The MCP App puts the same project facts inside the conversation, so the status you are talking about sits next to the sentence that changed it.
It ships two views. An inline card shows task counts, blocked work, verified success criteria, owner gaps, and the target date. A fullscreen board shows every lane with task detail disclosed in place.
The MCP App is read-only. You change projects by talking to the agent, and by the skill and CLI scripts as before; the app never writes. It runs as a stdio CLI over MCP, reading the same validated project state Studio reads.
Hosts that do not render MCP Apps still get the tools — they receive the compact text summary instead of a view.
Install as an Agent Plugin
The repository root is a self-contained Agent Plugins 1.0 package.
An Agent Plugins client that supports GitHub installation can clone yysun/project-manager and load
the repository root without selecting a generated subdirectory.
project-manager/
├── plugin.json
├── mcp.json
├── skills/project-manager/ # canonical skill
├── bin/project-manager-mcp.mjs # bundled MCP server
└── ui/ # self-contained MCP App views
npm run build:plugin refreshes the committed bin/ and ui/ runtime artifacts in place. Source,
tests, and build tooling may coexist with the portable components; Agent Plugins clients discover
only the fixed root manifest, skills/, and mcp.json locations.
plugin.json is the canonical product release version. Bump it together with the standalone skill
and MCP App runtime through one explicit command:
npm run release:version -- 1.10.1
The command does not publish, tag, edit the changelog, or sync an installed copy. After it succeeds,
update CHANGELOG.md, run npm test, and sync the complete affected plugin or skill installation.
Install in Claude Desktop
Claude Desktop does not read Agent Plugins packages, so add the server to
claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/):
{
"mcpServers": {
"project-manager": {
"command": "node",
"args": ["/absolute/path/to/project-manager/bin/project-manager-mcp.mjs"]
}
}
}
Selecting a project
No projects path is configured. When you provide a folder, the agent passes that folder to commands
such as project status <folder> and to the MCP App. When you omit the folder, the agent searches
only below the selected workspace root: one valid project is selected automatically, multiple valid
projects require a choice, and no valid project requires an explicit folder. A folder is accepted
whether or not anything was configured at launch.
Configuration is optional and does two things when you use it. --projects-root <folder> (or
PROJECT_MANAGER_PROJECTS_ROOT) lets projects be selected by ID or name instead of path, and
confines selection to that root — a project outside it is refused. --project <folder> pins a
single project. Without either, the server starts with nothing configured and waits to be told.
Worth knowing: with no projects root configured, the server can read any Project Manager project on the machine — the same reach the CLI scripts already give the agent. It only reads folders that parse as a project, never arbitrary files. Set a projects root if you want that confined.
Install
Choose the installation that matches what you want.
For the complete Agent Plugin — skill, MCP server, and MCP App — ask a client that supports GitHub Agent Plugin installation:
Install the Project Manager plugin from GitHub
yysun/project-manager.
For the standalone skill only, ask Codex:
Install the Project Manager skill from GitHub
yysun/project-manager.
Codex inspects the repository and installs skills/project-manager/ into its skills directory. This
does not install root mcp.json, bin/, or ui/, so the MCP tools and embedded App are unavailable.
Installers that do not infer nested skill paths may require the explicit path
skills/project-manager.
Development
npm ci
npm test
npm run pm-studio:dev
The development server generates a fresh disposable demo project on every start, so you can run it with no setup. To open a specific project instead:
npm run pm-studio:dev -- --project /absolute/path/to/project
Because a Task Contract binds an absolute project root, a demo is only valid for the checkout that generated it, so the demo is generated rather than committed. Create a persistent one — useful when you want Studio edits to survive a restart — with:
npm run demo
That writes demo/pm-studio-demo/ (gitignored), which you can then pass with
npm run pm-studio:dev -- --project demo/pm-studio-demo.
The installable skill is in skills/project-manager/ and Studio source is in
src/project-manager-studio/. The MCP server is isolated in src/mcp-app/; the MCP App adapter and views
live beside the shared Studio code in src/project-manager-studio/mcp-app/. Portable manifests stay
at the repository root.