pm is a human-agent project management suite for coding agents. Your agents claim the work, keep its context as records and raise what they need from you. You steer all of them from one site.
$ curl -fsSL https://github.com/Yeeef/pm/releases/latest/download/install.sh | sh
$ pm init # in your repo
pm keeps a project in two layers, a work store and Markdown records, and the pm cli is the only way into the work store. A service per clone turns both into the site you read and answer.
Projects, sprints and tasks in one work store that every session shares. An agent claims a task before it works on it, and a task has one holder. Dependencies say what waits on what, and pm task ready lists what can start now. That is what lets many agents work one repo without colliding.
Goals, sprint frames, decisions, findings and design pages, as Markdown on your repo's records branch. Each record points at its project or sprint, so the why stays next to the work. Agents read them with pm show; you read them on the site.
A site per clone, served by the pm service, shows every project, sprint and record with live status. The decisions and actions your agents need from you wait at the top. You answer on the page, and the reply reaches the session that asked.
A skill pack is a set of prompt files that ask an agent to behave. Nothing stops it from skipping a step, or from losing the rule an hour into a session. pm puts the agent's decision and action space in software: the pm cli, the work store and the session hooks. A step out of bounds is refused, and the refusal says what to do instead.
Guidance the agent reads, then weighs against everything else in its context. Whether it follows the plan is up to the model, every time.
claude CLI, so it must be installed and logged in, Codex setups included.$ pm task claim acme-q1.3.2 error: not claiming acme-q1.3.2 here: /home/me/acme is the main checkout; agents change code only in a worktree of their own. Make one and work there: … $ pm task close acme-q1.3.4 --reason "Parser done" error: acme-q1.3.4 is held by live session 5f0c2d1e-8a7b-4c3d-9e2f-1a6b7c8d9e0f (claimed 12m ago; its transcript was written in the last 30 minutes); leave it, or ask that session to close it, or to release it with pm task release acme-q1.3.4 $ pm sprint close acme-q1.3 error: open tasks in the sprint: acme-q1.3.5 (in_progress); close each with pm task close or move it with pm task move (a PR review closes with pm need close <id> --reason "merged as <sha>" once the PR is on main) $ pm decision need --title "Cache layer" --sprint acme-q1.3 --question '…' --fact '…' --option redis '…' --cost redis '…' --default redis '…' error: give at least two --option flags; a decision needs a choice; a decision need takes its parts as flags, one line each: … Stop hook · pm hook owner-request · blocked Your reply asks the owner for something that no open request of this session covers: "Can you confirm we drop the v1 API?". Chat requests never reach the owner's site, so the owner may never see them. …
Ids, paths and quoted asks are examples; the wording is pm's. "…" marks where a refusal goes on to give the exact command or form to use.
Every session starts with pm's rules and the project's state. The task, its sprint's frame and the decisions behind it are in the work store and the records, so a fresh agent picks up where the last one stopped. A task whose holder has been idle for 30 minutes on the same machine can be claimed by the next agent.
Plans, decisions and findings have a home: checked records on the records branch, apart from your code branches. pm's rules send context there, so your code tree stays code.
Agents raise the decisions and actions they need from you as requests, and each one lands in the same view. Every sprint has a goal, a scope and a done-when before any work runs: pm sprint open refuses a frame without them.
pm init adds no line to your AGENTS.md or CLAUDE.md. Session hooks for Claude Code and Codex inject pm's context, and each command's --help holds the rest.
The work store syncs through your own git remote under refs/pm/work, and records live on the records branch. On a new machine: clone, pm init, and your agents continue the work. No hosted service, no account.
You do not read transcripts. You read the site: the state of every project, the goal and progress of every sprint, and the requests that wait on you, rendered from the records and the work store as they change.
pm reply add; the agent that asked records your answer.
Hooks give each session its context; the cli is the only way to change the work store. pm writes two layers, and one service per clone renders them for you and carries your replies back.
refs/pm/work, and Markdown on the records branch.The script checks the binary against the release's SHA256SUMS and installs the latest release to ~/.local/bin/pm.
$ curl -fsSL https://github.com/Yeeef/pm/releases/latest/download/install.sh | sh
Pins the pm version, adds the hooks for Claude Code and Codex, creates the records branch and the work store, and starts the pm service.
$ pm init
Open a Claude Code or Codex session and describe the outcome you want. It arrives with pm's rules and the project's state, frames the work as a project and sprints, and asks you on the site for what only you can decide.
pm is open source under the MIT license. Read the code, file an issue, or install it and point your agents at a goal.
pm init --import-bd still imports a Beads export.
DoltThe work store: a SQL database with git-style versioning, embedded in pm and synced through your git remote.