Overboard

Overboard is a Claude Code plugin that watches every Claude Code session across all your repos and reports what shipped — on the Claude subscription you already have. One /overboard session becomes your assistant: it keeps a browser dashboard current and flags the few things that need you. It shows commit activity, architecture and DB-schema visualizations (Mermaid), your projects' real LLM prompts, Recent work cards distilled from the actual diffs, and — live — the key "needs review" tidbits from your working Claudes. Key-free by design: Python standard library only (no venv, no pip install), and all the AI runs on your Claude Code subscription — no ANTHROPIC_API_KEY.

The dashboard has a thin panel strip on the far left — Projects · Scheduler · Plugins, Settings pinned at the bottom (keys 1 2 3; the URL hash deep-links every panel: #/projects/<name>, #/scheduler, #/scheduler/run/<id>, #/plugins, #/settings/<page>).

Projects is the three-pane board: a condensed left sidebar lists every project with a calendar-style activity grid (each row a week, Mon→Sun, newest week on top — the last ~5 weeks) and any "⚑ N to review" flags; the center becomes the selected project's detail — summary, your assistant's report, Recent work review cards, recent activity, and per-repo architecture / prompts / data-shape analysis; and a collapsible right sidebar is where you set the project's launch/milestone (type, action, target date, goals — with push-back history) and vision/direction. The /overboard assistant reads that context to sharpen its reports and flag slipping launches. Scheduler and Plugins take the full width (below).

Install the plugin in Claude Code

From the Productive Mark marketplace:

/plugin marketplace add mrieck/claude-plugins
/plugin install overboard@productive-mark

Or from a local clone of this repo (it bundles its own marketplace):

/plugin marketplace add /path/to/overboard-plugin
/plugin install overboard@overboard-marketplace

Restart Claude Code (a full quit/relaunch, not just /reload-plugins) so it launches the MCP server with the current config. Check /mcp shows overboard ✔ connected. Requires python3 3.9+ on your PATH (stock macOS and any recent Linux qualify).

Run it

From Claude Code, /overboard opens the dashboard and runs update passes until the backlog is drained, then stops — run it again whenever you want a fresh sweep. To run the dashboard directly:

python3 -m overboard.app          # serve dashboard, open browser (native window if pywebview present)
python3 -m overboard.app --once   # headless refresh, print, exit

It serves a browser dashboard at http://localhost:8787. On first launch a short onboarding wizard picks your sources; local clones are then discovered and analyzed automatically. The good stuff (summaries, Recent work cards, real prompts, architecture write-ups) fills in once /overboard has run. Refresh re-pulls commits; Rescan re-discovers local clones.

Scheduler

The Scheduler panel runs recurring, unattended Claude Code sessions — "nightly triage at 2am", "every 2h between 9–18", "once, tonight at 11" — without the Mac app. It is laid out like the Mac app's: a 10-day agenda on the left (every upcoming fire, click one to open its agent), Agents | History in the middle, and a permanent editor on the right (or the picked run's detail). Each agent (a slot) has a name, a command/prompt, a working folder (or a linked task workspace), a simple schedule (daily / weekly / every N hours / once — deliberately simpler than cron), a timeout and an enabled switch. The prompt field completes /slash-commands (yours, the project's, every installed plugin's commands and skills) and ~/ ./ / paths, and warns when the command's plugin isn't enabled in that folder — with an Enable button.

Runs launch inside Herdr (macOS and Linux) as claude --permission-mode auto, one workspace per project (named after the project folder) with one tab per run named after the task, so you can watch or take over any run in Herdr's own UI. Each run is identified by its own Claude session (captured from the SessionStart hook), so a session of yours in the same folder is never mistaken for it; when the run's Stop hook fires — and no subagent it spawned is still working — the scheduler captures the closing message and transcript, then closes the pane. A fresh claude in a never-trusted folder blocks on the folder-trust dialog; the scheduler answers that one dialog (once per run) and nothing else — any other prompt shows as waiting for input. Take over focuses the session in Herdr and stops managing it (no auto-/exit, the pane stays yours). History shows what each run reported — the closing message as prose — with the raw terminal capture behind a disclosure; hide, remove or clear runs. A once agent disables itself after it fires. Concurrent runs (1–8) and the default timeout are in Settings ▸ Misc; an agent never overlaps its own next firing. Queued runs survive a dashboard restart (a run that waited past its 30-minute window is skipped, not run stale). State lives in ~/.cache/overboard/scheduler/.

Dispatcher

Top-left of the Scheduler panel: type what should get done ("add a changelog entry to project X for this week's commits") and press Dispatch. A short-lived /overboard:dispatch session reads the request, works out the project (via the plugin's list_projects), the prompt and the time, and answers with a task — the scheduler runs it (now, or as a once agent for later), and the result (the worker writes a result.json into a task folder granted via --add-dir: summary, files, links) lands in the Recent dispatches feed, each with a #id. The dispatcher may also reply from the task ledger ("what's running?") or reject a non-task. Follow-ups ("make it shorter") rebuild the task's context into a fresh session. It needs an Overboard run directory (Settings ▸ Misc) — your workspace root, where new project folders may be created. State lives in ~/.cache/overboard/dispatch/.

Dispatching from your phone (Telegram, Slack next) is Overboard for Mac's feature; the dashboard's dispatcher is browser-only by design.

Tests: python3 -m unittest discover -s tests (stdlib only). To run the dashboard against a copy of someone else's state without touching your own, point it at another folder: OVERBOARD_STATE_DIR=/path python3 -m overboard.app.

Slots fire only while the dashboard server is running. On an always-on Linux box, run it as a systemd user service:

# ~/.config/systemd/user/overboard.service
[Unit]
Description=Overboard dashboard + scheduler

[Service]
ExecStart=/usr/bin/python3 -m overboard.app --serve --no-open
WorkingDirectory=%h/.claude/plugins/overboard   # wherever the plugin checkout lives
Restart=on-failure

[Install]
WantedBy=default.target

systemctl --user enable --now overboard
loginctl enable-linger $USER   # keep it running when you're logged out

Configuration

Out of the box, Local-only mode needs no credentials at all — it reads commits straight from the git clones on your machine with git log. Add a GitHub token (classic with repo scope, or fine-grained with repository contents + metadata, read) or a Bitbucket API token (workspace + Atlassian email) in Settings ▸ Sources to also include commits you pushed from other machines — every source merges into the same board.

Tokens are saved to ~/.cache/overboard/credentials.json (mode 0600, machine-local) and never sent back to the browser. The commit window (how many days of inactivity before a repo drops off the board) is adjustable in Settings too. Non-secret app defaults are in overboard/projects.json — refresh_interval_minutes, commit_window_days, and optional local_roots (defaults include ~/projects, ~/code, ~/Developer, …) for where local clones are discovered. Other files under ~/.cache/overboard/: state.json (commits, analysis, activity), ai.json (the assistant's summaries and write-ups), context.json (your launches + vision), events.jsonl (append-only activity log). Delete any to reset that piece.

License

MIT