Cue

🐙
GitHub is the state store
Labels are the state machine, issue comments carry the plan, draft PRs are the output. Runs are resumable from any machine and fully auditable in the issue thread.
🛡️
Humans gate the irreversible steps
You approve the implementation plan and merge the final PR. Cue never merges, never force-pushes, and never touches the base branch.
🧪
Deterministic quality gates
Real test and lint commands gate every change in the worktree. Code validation is deterministic TypeScript and shell execution — never LLM self-judgment.
🌲
Isolated worktrees & scrubbed secrets
Implementation happens in git worktrees outside the target repo. GitHub tokens are completely scrubbed from agent subprocess environments.
🤖
Multi-engine flexibility
Switch seamlessly between OpenAI Codex, Claude Code, and Google Antigravity. One unified config and workflow across all your projects.
📊
Local zero-config dashboard
cue ui displays active spend, the label board, streaming live logs, and recorded prompts and transcripts. Localhost only, no external backend.

How it works

Cue turns interactive coding agents into an asynchronous PR factory. Instead of watching an agent type in your terminal, you interact entirely through GitHub issues and pull requests.

Orange steps are yours — everything in between runs headless.

Why Cue?

Interactive Terminal Agents	The Cue Pipeline
Blocks your terminal: You babysit streaming output and prompt responses.	Asynchronous: Label an issue `agent:ready`, close your laptop, and check back later.
Unchecked architecture: Agent immediately edits files, sometimes taking the wrong approach.	Plan approval first: Agent drafts a plan (`agent:planned`). Implementation only starts once you approve (`agent:approved`).
LLM self-evaluation: Agent decides subjectively whether its changes work.	Deterministic gates: Real test scripts (`bun test`, `npm test`, `pytest`) must pass before a PR is opened.
Dirty working tree: Agent modifies your active git working copy.	Worktree isolation: Agent works in an isolated worktree outside your project root.
Unsafe git operations: Agents can commit, push, or pollute git history.	Runner ownership: Cue owns commits, pushes, and draft PRs; the agent never gets `GH_TOKEN`.

Interactive Terminal Agents

The Cue Pipeline

Blocks your terminal: You babysit streaming output and prompt responses.

Asynchronous: Label an issue agent:ready, close your laptop, and check back later.

Unchecked architecture: Agent immediately edits files, sometimes taking the wrong approach.

Plan approval first: Agent drafts a plan (agent:planned). Implementation only starts once you approve (agent:approved).

LLM self-evaluation: Agent decides subjectively whether its changes work.

Deterministic gates: Real test scripts (bun test, npm test, pytest) must pass before a PR is opened.

Dirty working tree: Agent modifies your active git working copy.

Worktree isolation: Agent works in an isolated worktree outside your project root.

Unsafe git operations: Agents can commit, push, or pollute git history.

Runner ownership: Cue owns commits, pushes, and draft PRs; the agent never gets GH_TOKEN.

CueHeadless coding agents, a fixed pipeline

GitHub is the state store

Humans gate the irreversible steps

Deterministic quality gates

Isolated worktrees & scrubbed secrets

Multi-engine flexibility

Local zero-config dashboard

How it works ​

Why Cue? ​

How it works

Why Cue?