You don't work in one repo — you've got eight, each with failing checks, open PRs, half-finished agent sessions. Commandem is the desktop command center that shows what needs you across all of them, and dispatches the work back.
go install github.com/Scarcrux/commandem/cmd/commandem@latest
The interface
Every tool is great inside one project and blind across them. Commandem lives above all of them — one surface for status, review, planning, and dispatch.
Plan
A cross-project kanban board. Drag a card to restatus and it writes straight back to that project's roadmap, comment-preserving.
Review
Every PR awaiting you, across repos, in one inbox. Open the diff, get an AI review, and post it back to GitHub as a real review — approve, comment, or request changes.
Context
claude, codex, and gemini sessions across all your repos, drawn as a single graph. Jump into any project's chat, code, activity, and roadmap from one window.
Insight
Commits, finished agent runs, check transitions, and deploys — an activity feed and rollups across projects, plus a weekly digest. No timesheet.
Why it's different
Because they live inside one project — and this lives above all of them.
One attention-bucketed view across every repo: failing checks, PRs to review, agents awaiting approval. Answers "what should I do right now?", not "what's happening."
claude, codex, gemini, copilot, aider — per project or per task. Commandem coordinates them; it isn't tied to any one vendor's agent.
Autopilot works in a throwaway git worktree (never your tree), refuses "done" until your tests pass, and is seeded with how you solved it in your other repos. Guardrails before autonomy.
A single Go binary. No accounts, no hosted backend, no telemetry. Bring your own Claude subscription. On-device and local models are first-class for cost + privacy.
Under the hood
The app owns no business logic — it's a shell over one local binary. Everything the window does is also a keyboard-first command, so it drops straight into your shell, pipes, and cron.
$ commandem status # "what needs me right now?" across every repo $ commandem open backend --agent codex # route the best agent per project # Run it unattended, safely: isolated git worktree, gated on tests, # seeded with how you fixed this in your other repos. $ commandem autopilot api "fix the retry race" --worktree --verify "go test ./..." $ commandem recall "flaky retry" # how did I solve this before? (any repo) $ commandem status --json | jq # --json on every verb; ~80 of them
Try it now
Two demos that run 100% client-side. No server, no tokens spent.
Pick a recipe, type a brief, and watch the prompt render against project context — exactly like commandem recipe run.
Paste a task, see its complexity tier and which model Commandem routes it to — a faithful port of the real, free, deterministic scorer.
Open the explainer →Install it, point it at ~/Code, and see what needs you.
commandem scan && commandem status