Caprock

What does this agent do, and when should you use it?

Caprock is a local-first desktop app (Tauri, macOS/Windows/Linux) wrapped around a Go daemon that also serves a browser dashboard. It puts Claude Code, Codex, OpenCode, Gemini CLI, DeepSeek Harness and Hermes Agent sessions into terminal tabs and split panes, organized by project and git worktree, with a plain shell alongside. It reads what each agent already writes — transcripts, per-agent databases, OpenTelemetry — prices every turn at API list rates, and breaks spend down by repository, model, service and tool call, side by side with Claude's 5-hour and weekly plan windows. It also shows the repository's pull requests with a per-job, per-OS CI check matrix and can hand a red run, conflict or stale branch back to an agent. All state lives in one local SQLite file on loopback with no account or telemetry; a phone can pair through an end-to-end encrypted relay or the LAN, and caprock mcp exposes a read-only query surface to your agents.

Caprock runs a local daemon (caprock up, default localhost:22776) under a Tauri shell. It observes Claude Code through hooks plus transcripts — the hook and status-line entries in ~/.claude/settings.json are offered on first run, backed up, and skippable — and collects other agents from their own stores: OpenCode's database opened read-only (v1 and v2, deduplicated after migration), Codex's per-session transcripts under $CODEX_HOME or ~/.codex including archived ones, DeepSeek Harness transcripts, Gemini's OpenTelemetry file, and Hermes Agent's ~/.hermes/state.db. Each turn is priced from vendor list-price tables (OpenCode and Hermes pass through their own figures), feeding the Home activity pulse, Cost and Lifetime reports, a context-tax readout of what the next call costs, and per-tool-call groupings. It launches, types into and stops sessions it started, summarizes the repository's pull requests with CI checks per job and OS, and offers a Changes view to stage, commit and push. CLI surfaces include caprock report, caprock export, caprock doctor, caprock sessions rm and caprock license. The opt-in task runner (caprock up --hive … --repo …) runs each task in its own git worktree and only marks it done when the done_criteria shell commands exit 0.

1. A developer running Claude Code and Codex in parallel who wants both agents' sessions and spend visible in one window instead of scattered terminals.
2. Someone on a Claude Max/Pro plan doing heavy agent work who needs to see list-price usage against the 5-hour and weekly windows before hitting a limit.
3. Teams maintaining several services in a monorepo who want per-repository and per-service cost attribution for agent sessions.
4. Maintainers who want PR checks as a job-by-OS grid, with failed jobs sent straight back to an agent via 'Fix CI with agent'.
5. Remote workers who want to approve permission prompts, check spend and continue sessions from a phone on mobile data or the home Wi-Fi.
6. Teams that want their coding agents to look up past decisions, repository cost and CI status themselves through the read-only caprock mcp server.

How do you install or deploy this agent?

Install the desktop app on macOS with the Homebrew cask:

brew install --cask dspv/tap/caprock-app

Or download directly from Releases: macOS .dmg, the Windows setup .exe, Linux .AppImage/.deb/.rpm. The app is not notarized or code-signed yet, so the first .dmg open needs the one-time 'Open Anyway' step and Windows needs SmartScreen's 'Run anyway' — see docs/install-app.md.

For the daemon and browser dashboard alone:

# macOS / Linux
brew install dspv/tap/caprock

# Windows
scoop bucket add dspv https://github.com/dspv/scoop-bucket
scoop install caprock

# with a Go toolchain
go install github.com/dspv/caprock/cmd/caprock@latest

Then run caprock up, which opens localhost:22776 and asks before adding its hook and status-line entries. First run reads the transcripts Claude Code already wrote, so history appears without migration; if the older PyPI Python caprock is installed, remove it first so the two commands do not shadow each other.

How do you use this agent?

Start the daemon and open the UI:

caprock up

In the app, ⌘T opens a shell, ⇧⌘N an agent, ⌘O adds, creates or clones a project; each project expands to its worktrees and the sessions inside, with whatever waits on you pinned to the top. To start at login:

caprock service install
caprock service status

Produce a publishable report and export the raw record:

caprock report --markdown
caprock export --since 30d > events.tsv
caprock export sessions --format csv --out s.csv

Wire Caprock into your agent as a read-only MCP server:

claude mcp add caprock -- caprock mcp
codex mcp add caprock -- caprock mcp

Optionally, the unattended task runner (off by default; workers run with permission prompts skipped, so set a budget per task):

caprock up --hive ~/caprock-tasks --repo ~/dev/myproject
caprock task create --title "Add a /healthz endpoint" \
  --done-criteria "go test ./..." --done-criteria "go vet ./..." --budget 3

For a phone, press Connect a phone under Settings → Use on your phone and scan the QR code, or from SSH run caprock remote on then caprock pair (view-only). Activate Premium features with caprock license set YOUR-KEY.

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

FAQ