How CoDev Works
The mental model first: the phases, the two review loops, and who decides what. Fifteen minutes, and the rest follows. Start here →
Welcome to CoDev
You describe what you want, in plain language, to an AI agent. CoDev’s agents inspect the
actual repository, propose a plan, and — once you approve it — do the work: they run the
codev CLI on your behalf so you don’t have to. A developer and AI move through four
steps this way — Understand, Build, Review, Ship — plus two that only show up when the
work actually needs them: Specify for a genuinely new product, Launch for a real
rollout decision.
Choose where to start
Section titled “Choose where to start”Getting Started
Install the bundle into a repository in under a minute — the one part of this that really is a CLI command. Read the guide →
Talking to Your Agent
The core idea: you talk, your agent runs the CLI. What to say, and when it actually matters. Read it first →
Tutorials
Four start-to-finish walkthroughs — a first fix, a design-worthy change, outer-loop review, two developers on one contract. Start with Tutorial 1 →
Roles
Every agent CoDev installs, what it may do, and what it may never do — an agent may write code or review it, never both. See the roles →
Slices and Stacks
Why work is split into small pull requests, and how a stack gets built and kept current. Read about slices →
Evaluating Skills
Measure whether a skill actually helps, empirically, rather than trusting that it does. Run an eval →
Agent Platforms
What differs across OpenCode, Claude Code, Junie, and Antigravity. Compare platforms →
CLI reference
What your agent runs on your behalf, documented in full — for scripting, CI, and debugging. See every command →
Architecture
How the bundle is built, installed, and updated — the install/update algorithm and its invariants. Read the design →
FAQ
Credentials, conflicts, licensing — the questions that come up first. Read the FAQ →
Why CoDev
Section titled “Why CoDev”- One workflow for solo developers and multi-developer teams.
- Repository-grounded plans instead of invented APIs or architecture.
- A bounded builder, an automatic pre-PR audit gate, and five parallel specialist reviewers — not just one read-only pass.
- Human authority over material decisions, merge, deployment, and rollout.
- A general-purpose skill-evaluation harness — measure whether a skill actually helps, empirically, rather than just trusting that it does.
- Versioned, conflict-aware installation across existing repositories.
- No runtime dependency in the software being built.