Introduction
Argus
AI code review inside your Azure DevOps pull requests. A conversational,
multi-tenant reviewer that posts findings as inline threads, defends or
concedes them when you push back, and answers @argus questions — statefully,
per thread.
Getting started · Onboard a repo · Architecture
The 30-second pitch
Argus is a long-running, multi-tenant code-review service for Azure DevOps. When a PR is opened or updated against a protected branch, Argus checks out the PR merge commit into an ephemeral read-only workspace, loads the repo's own review standards from a rules registry, and reviews the change the way a strong human reviewer would — cross-variant safety, RBAC, secret hygiene, test conventions — not generic lint advice.
What distinguishes it from a one-shot CI check is statefulness: a reply to a three-day-old finding resumes that thread's conversation rather than re-reviewing the PR from scratch.
Argus watches and reports. It cannot approve, merge, or push — humans keep the gate.
Features
| Feature | What it does |
|---|---|
| Inline review threads | Every finding lands as a PR thread anchored to the exact file and line, with a closed severity (must-fix / note) and category set — never free-text driving behaviour. |
| In-thread conversation | Reply "this is intentional because X" and Argus answers in the same thread with its full history in context; it concedes and resolves, or holds the finding with reasons. |
@argus mentions |
Mention Argus anywhere in a PR, including threads it did not start, to ask about the diff, the codebase, or a convention. It reads the code and answers with file references. |
| Incremental review | Push a fix and Argus reviews only the delta since the last reviewed commit; it never repeats findings already posted, and resolves threads the push addressed. |
| Rules routing, per tenant | Review standards live in a versioned rules registry routed by org/project/repo with path-scoped packs; onboarding a repo is configuration, not a deploy. |
| Safe by construction | The agent comments; humans approve and merge. Credentials never enter model context, mutations flow only through a five-tool MCP surface, and hostile PR content is data, not instructions. |
Where to go next
| You are… | Start here |
|---|---|
| An engineer whose PR just got reviewed | Getting started |
| A team lead onboarding a repo | Onboarding a repo |
| A rule-pack author | Writing rule packs |
| A developer of Argus itself | Running locally & offline testing |
| An operator deploying it | Deployment |
| Wondering about cost, privacy, or how to turn it off | FAQ |
| A contributor | Contributing & project |
| Curious about the name | About the name |