Away-from-keyboard telemetry. Run afk start on a machine you are about to walk away from and get a shareable dashboard URL that shows whether everything is still fine.

curl -fsSL https://afk.osv.im/install | sh
export PATH="$HOME/.local/bin:$PATH"   # this terminal only; the installer says how to make it stick
afk start                              # prints a dashboard URL for your phone
afk run -- <command>                   # wraps one command and reports how it went

macOS only for now. The installer puts one bash script in ~/.local/bin/afk (AFK_INSTALL_DIR overrides). That directory is not on a stock Mac's PATH, so the installer ends with the line to add to ~/.zshrc or ~/.bash_profile (in fish, fish_add_path ~/.local/bin) and, until then, the full path to run. Read it first if you like: it is short, and so is the client. A self-hosted server serves the same installer at its own /install, and a client installed from it defaults to that server; set AFK_SERVER to point an existing client elsewhere.

• Keep the Mac awake while a session runs: sleep pauses the client, and the server ends a session that goes quiet for ten minutes. Lid open and plugged in, or caffeinate -s ./cli/afk start.
• A session lasts an hour, then continues in a fresh one with a new link. The old dashboard stays readable and links to its successor.
• Sessions are kept 7 days after they end, then deleted. Anyone with the link can read or delete one; there are no accounts.
• macOS only, and alpha: afk.osv.im is one person's server, run best effort.

Everything below and nothing else: no file contents, no environment variables, no process arguments, no agent transcripts, and no output of a command that succeeds. The wire format is spelled out field by field in docs/PROTOCOL.md.

• At the start: host name, macOS version, core count, memory size, and the afk client version.
• Once a second: cpu percent, load averages, the memory pressure level, and memory numbers (free, active, inactive, wired, compressed, swap used and total).
• Every five seconds: the 10 busiest processes, each as pid, parent pid, cpu and memory percent, resident size, and the executable's full path (never its arguments), plus how many processes there were.
• Every five seconds: how many Claude Code and Codex sessions are running, working, waiting on you, or idle, and how many subagents are working. Counts only: no session names, directories, or transcript contents. AFK_NO_AGENTS=1 turns this off.
• For afk run: the command line as typed, with a URL's user:password, the value of a KEY=value argument named like a secret (TOKEN, SECRET, PASSWORD, KEY, AUTH), and the value after --password, --token and the like replaced by *** before it is sent; then its pid, elapsed time, cpu and memory, how many bytes it wrote to stdout and stderr, and its exit code.
• When a wrapped command fails: the last 20 lines of its stdout and stderr (200 characters each), so the dashboard can say why. A command that exits 0 sends no output. AFK_RUN_TAIL_LINES=0 sends none at all.

Where it goes:

• To the server the client points at. afk.osv.im is one person's server in AWS us-west-2, run best effort with no SLA; a self-hosted server is whoever runs it, and AFK_SERVER points the client at one.
• Kept for 7 days after the session ends (AFK_RETENTION_DAYS on your own server), then deleted.
• Readable by anyone holding the link. There are no accounts: the session id is the secret, and the client's write token is never in the URL.
• Deletable by anyone holding the link, at once and with everything it recorded: the Delete control on the session page, or afk delete on the machine.
• Questions and takedowns: open an issue.

Status: early prototype. See docs/ARCHITECTURE.md, docs/PROTOCOL.md, docs/CONFIGURATION.md (every server environment variable), docs/EXTENDING.md, and BACKLOG.md.

• cli/afk – the client. One bash script, macOS only for now.
• packages/shared – Zod schemas for the wire protocol; types are inferred from them.
• packages/server – Node + Hono ingest and dashboard API.
• packages/web – Vite + React + TanStack dashboard.
• infra – OpenTofu config and deploy script for the afk.osv.im deployment.
• docs – architecture, protocol, and extension guides.
• .github/workflows – CI (typecheck/lint/format/test/build, plus infra/ validation) on every PR and push to main, and a deploy pipeline to afk.osv.im on push to main or manual dispatch; see the "CI and deploys" section of infra/README.md.

pnpm install
pnpm build                                          # dashboard into packages/web/dist, server and shared into their dist/
pnpm dev:server                                     # http://localhost:4141 under tsx watch, serves the built dashboard
AFK_SERVER=http://localhost:4141 ./cli/afk start    # in another terminal; open the URL it prints

pnpm dev:server runs the TypeScript source through tsx; the Docker image runs the compiled output instead (pnpm --filter @afk/server start, i.e. node --conditions=afk-compiled dist/index.js, is the same entrypoint locally).

The dashboard follows an active session live over server-sent events and shows the whole trace once it ends. To iterate on the UI without rebuilding, run pnpm dev:web (Vite on http://localhost:5173, proxying /api to the server) and start the client with AFK_PUBLIC_BASE_URL=http://localhost:5173 so the printed URL opens there.

Watch a session's frames arrive as server-sent events (works in a browser tab or curl):

curl -N http://localhost:4141/api/sessions/<sessionId>/stream

Reconnect where you left off with -H 'Last-Event-ID: <index>' or ?after=<index>. GET /api/sessions/<sessionId>/frames?after=<index> returns the same data as one JSON document.