Coroid is an AI software factory for production systems. Engineering sets the standard once: the project context, the quality policies every change must pass, the review rules and who approves. After that, anyone in your company can ask for a change in plain language, without a CLI or knowing how the code is built. Coroid plans the work, writes the code, tests it, reviews it to that standard, and opens a pull request in your repository for you to approve.
These docs cover the product as it ships today. If a feature is not documented here, it is not available yet.
New here
Work through Get started in order. It takes you from an empty account to a merged pull request: create an organization, connect a repository, create a project, run a task, review the result.
Understand how it works
How a task becomes a pull request is the page everything else refers back to — the full path from your description to a PR, with every stage named.
Then Agent roles for who does what, and the specification-driven development guide, which is the highest-leverage thing you can get good at. The quality of what Coroid builds is bounded by the quality of what it is asked to build.
- 1
SpecificationWhat will be built, and how you will know it worked
ArchitectYou can edit - 2Plan
Ordered steps, reviewed before any code is written
ArchitectYou approve - 3
BuildCode and tests written in an isolated workspace
Developer - 4
VerifyYour test suite, plus checks and quality gates
QA - 5
ReviewAn automated review pass over the diff
Reviewer - 6Pull request
On a branch in your repository, waiting for you
You merge
Failed verification and fixable review findings go back to the developer as rework, before any pull request opens.
Find your way around
| Section | What it covers |
|---|---|
| Get started | Account to first merged pull request |
| Concepts | How the system actually works |
| Work and planning | Starting work, tasks, plans, releases |
| Projects and code | Repositories, build and test setup, context |
| The factory | Running work, automations, reports |
| Quality and review | Gates, findings, evidence, browser tests |
| Globalization | Languages, translations, world-ready code |
| Operations and monitoring | Mission Control, findings, notifications |
| Organization settings | Members, keys, policy, retention |
| Billing and capacity | Slots, credits, usage, invoices |
| Integrations | Source control, chat, MCP, evidence, hooks |
| Reference | Glossary, statuses, limits, troubleshooting |
Three things worth knowing immediately
Agent slots control how many tasks run at once — not how fast one runs. A single task's duration comes from the size of the change and your test suite. See Capacity and agent slots.
NEEDS_HUMAN_REVIEW is not a failure. It means the work hit a decision only
you can make. Answer it and the task continues. See
Task statuses.
Something not behaving as expected
Start at the troubleshooting index, which maps symptoms to answers.