Buster Claw — Security skill for Claude Code

Buster Claw

**An assistant on your Mac that uses your tools, keeps working, and shows you what it did.**

Ask it in the app's own Chat, email it, or drive it from the built-in terminal. It reads and acts inside the browser tab you're actually looking at, your Gmail, Calendar and Drive, and it works a durable queue that survives a restart — one canonical command surface, 214 commands. Everything it changes lands on an auditable feed you can read afterwards, and restricted actions are refused outright for untrusted callers.

**You bring the intelligence.** There is no LLM inside Buster Claw and it needs no API keys: it runs **Claude Code, Codex, or OpenCode** under the agent subscription you already pay for. The app is the hands, the memory, and the receipts.

It's an Elixir/Phoenix + LiveView application wrapped in a Tauri desktop shell — Phoenix at the repository root, the shell in `desktop/tauri`.

---

How it actually works

The agent doesn't call a chat API. It works a **queue**.

Trusted inbound requests land in a durable SQLite **Dispatch queue** — today that means mail from a sender on your trusted-senders list, or anything you (or an agent) enqueue by hand. Buster Claw projects that queue into workspace markdown the agent already reads (`Dispatch.md`). The agent pulls an item, does the work, and writes the result back through the `./buster-claw` CLI. The desktop UI gives you the command surface, the audit feed, and the results.

(Integration webhooks do *not* enqueue work — a verified GitHub event becomes a Library snapshot, not an agent task.)

That indirection is the whole design. It means work survives a crash, an agent can be replaced mid-shift, and nothing the agent does is invisible to you.

Features

• One command surface. 214 commands across documents, browser, Google Workspace, integrations, finance, phone, sound and voice, notes, memory, skills, and orchestration — reachable from the CLI and an HTTP API, with per-caller trust tiers and