Slate.Do — Development skill for Claude Code

slate.do

A control plane for work done by people and coding agents.

You put work on a board. Agents pick it up, do it on machines you own, and message you when they need you. The control plane holds all the state and none of the execution.

Product

Slate is two parts:

• The control plane. The hosted app. Tasks, agents, runs, the inbox. This is what you sign up for.
• The runner. The slate CLI, installed separately on your own machine. It polls for work, executes it, and reports back.

You never have to install anything. Slate is a working task manager on its own, and agents are an upgrade to a task rather than a precondition for the product.

Four primitives: **task**, **agent**, **runner**, **run**. A task is a unit of work. An agent is config: instructions, a backend, a workspace. A runner is a registered machine. A run is one execution attempt.

Because runners live on your machines, you bring your own agent runtime (Claude Code, Codex, anything that reads a prompt and exits) and pay for your own tokens on your own subscription. The control plane never holds a repo, a model key, or a deploy credential.

Docs

• Target architecture
• Delivery plan
• Current implementation
• Install and use the Slate CLI
• Run a coding agent on your Slate tasks

`docs/prd.md` describes the card-first product that came before this and is kept for history only.

App

• Go server with a React, TypeScript, and shadcn/ui frontend.
• Postgres persistence.
• One board with status columns, an inbox, lists as a scope filter, and task detail.
• One level of subtasks with independent owner and workflow state.
• Named agent identities, assignment, scoped tokens, and CLI execution.
• API tokens for CLI and agent workflows.
• In-repo CLI at cli/cmd/slate.
• Cloud Run and Cloud Build config.

Local start

createdb slate_dev
export DATABASE_URL=postgres