Deskwork — AI skill for Claude Code

Deskwork

**A terminal-first console for running several long-lived AI agents, across several vendors, without losing track of what it costs.**

macOS · Swift · early, but it runs

[](https://github.com/anthonyproctor/deskwork/actions/workflows/ci.yml)

---

The idea

Every coding tool picks one of two units of work. VS Code and its descendants treat the **file** as the unit. Zed's agent panel, Copilot and Cursor treat the **thread** as the unit — a conversation starts, does a job, gets thrown away.

Once you have been living in this stuff for a few months, neither fits. You stop having conversations and start keeping **specialists**: one for your infrastructure, one for a particular client, one for your notes. Each has its own system prompt, its own memory on disk, its own pinned model. The session is disposable. The specialist is not — it has been accumulating judgement for months, and you don't re-explain your situation to it.

Deskwork calls one of those a **desk**, and makes the desk the unit.

What it does

**Desks, not threads.** Each desk owns one or more long-lived terminals running the vendor's own CLI. Switching desks swaps which terminal is visible; the rest keep running. Come back an hour later and it is mid-thought where you left it.

**It never reimplements an agent.** Deskwork launches `claude`, `codex`, `gemini` or `copilot` in a real pty. Your agent definitions, hooks, memory files and model pins all apply, because nothing is intercepting them. This is the point: ACP-based editors run a Claude Code bundled inside the Agent SDK rather than the CLI on your machine, which is why your own agents do not exist there.

**Splits, when one terminal is not enough.** `cmd-d` and `cmd-shift-d` give a desk a second pane: the agent in one, a shell in the other to look at what it just did. Pane 0 is the agent and confirms before you close it; the rest are login