Your Claude has been running for 40 minutes. Doing what?

A single local binary. Run `caprock up`, open localhost:4173, and every Claude Code session you start — in any terminal — shows up live: what it is doing, what it has cost, a live git diff, and a banner if it starts repeating itself. It reads OpenCode sessions too. You can also spawn and control sessions from the dashboard, and run queued tasks unattended behind a test gate.

Effectively no. The daemon binds to 127.0.0.1 — nothing on your network can reach it unless you start it with `--lan`, and then only a device that pairs with a code you read off your own screen. There is no Caprock service in the path, no telemetry, no account; all your data lives in a SQLite file on your disk, and nothing about you is ever sent to us. Three things can reach the network, each off until you turn it on: a check for new releases, Gemini answering questions on your own key, and a weekly report to your own Telegram bot. It's Apache-2.0 on GitHub; read it before you run it.

Two ways. A tiny hook shim registered in ~/.claude/settings.json (non-destructive, backed up, cleanly removable) forwards Claude Code's lifecycle events; and it tails the transcript files Claude Code already writes. You keep talking to Claude in your terminal — Caprock is a window, not a new runtime.

The loop detector flags a session that repeats the same tool with near-identical input several times in a few minutes, and raises a banner. It watches the tools that change something — edits, writes, commands — and deliberately ignores repeated reads, globs and greps: re-reading a file after an edit is ordinary work, and on a replay of 64,733 real tool calls that noise was over half of every alert raised. The tool call itself is not what costs; the turns a stuck session keeps taking are, which is why the banner is worth interrupting you for. For sessions Caprock spawned, it can auto-pause them.

An unattended runner with a test gate, off until you turn it on. You write a task with `done_criteria` — plain commands: tests, typecheck, lint. Caprock runs one Claude Code session per task in its own git worktree, and when that session says it is done, Caprock runs the commands itself. Only green moves a task to Done; a red result bounces the failing output back, and after three rounds it stops and asks you.

Those commands run in a real checkout of your repository, with your privileges, unsandboxed, on a five-minute ceiling each — treat them as commands you are running yourself, because `go build ./...` leaves a binary behind exactly as it would in your own tree. Workers run with permission prompts skipped, which is why it confirms before starting.

No. It never writes to them: a finished task shows you the diff, the checks that passed, and the git command that takes the work. Give it independent tasks only — nothing here merges branches, and nothing notices two workers editing the same file.

It's free and open-source, and it runs on the Claude subscription (or API key, or Bedrock) you already pay for — Caprock adds no service and no fee. Costs shown on the dashboard are what your usage would cost at Anthropic list prices; on a flat plan that's an equivalent, not a charge.