Code Review for Your AI Coding Agents¶
An MCP server that gives your coding agent — Claude Code, Cursor, OpenAI Codex, or any MCP client — a set of pre-commit code-review tools.
CloudAEye reviews the changes made during a coding session. Run a command, and it inspects your uncommitted work for bugs, security problems, task misalignment, and more — returning findings before the code reaches a commit, a PR, or CI, and before it ever leaves your machine.
Supported Coding Agents¶
| Client | Status |
|---|---|
| Claude Code | Ships as a plugin |
| Cursor | Planned |
| OpenAI Codex | Planned |
Building against the server directly, or adding a client we don't list? The Tool Reference has the HTTP endpoints and tool signatures.
What you get¶
| Command | What it does |
|---|---|
| Inspect | Bug pass — logic errors, edge cases, error handling, concurrency, dead imports. |
| Security | Security pass — OWASP-style application security plus the LLM, AI-agent and MCP surfaces, and secrets on the changed lines. |
| Review | Both of the above in one call — the widest pass. |
| Slop Score | How much careful reading a change needs before merge — nine signals of low-quality code, scored into one of four review bands. |
| Control Flow | What a change did to the order of calls — before and after sequence diagrams read from the code graph, every added, removed or moved call marked, and callers in other repositories the change may break. |
| Architecture Review | What a change did to the architecture — three pictures, widest first: your services as their deployment files wire them, the repository's layers, and the change — then layer violations, new cycles, first-time dependencies, and the questions a senior engineer would ask. |
| Describe | A Change Description and Important Changes list, ready for a PR body or commit message. |
| Check Task | Does this coding session actually do what the ticket asked? A DONE / NOT DONE verdict with a per-requirement checklist. |
| Ask | A free-form question about the pending change, answered against the repository's code graph. |
| Check PR | The hygiene checklist over an open pull request — docstring and test coverage, README and dependency freshness, title and description quality, secrets — with the full report posted on the pull request. |
| Add Docs | Docstrings for the undocumented code a pull request touches, posted on the pull request as one-click suggestions. |
| Add Tests | Unit tests for the changed code nothing covers, posted the same way — including whole new test files. |
| Create Issue | A finding from a review, filed as a Jira ticket or a GitHub issue with the evidence attached. |
None of them edit your code. They report; you decide what to fix. Four of them write somewhere other people can see — the pull request, or an issue tracker — and each one confirms with you first.
How each command is invoked depends on the client — in Claude Code they are slash commands like /cloudaeye:inspect. See Claude Code Plugin → Setup to install, then the User Guide for what to run.
Documentation¶
| Page | What it covers |
|---|---|
| Overview | How a review works, what the three passes cover, and what leaves your machine |
| Supported Coding Agents | Which clients are supported today |
| Claude Code Plugin — Setup | Install the plugin and connect it |
| Claude Code Plugin — User Guide | Every command: what to pass, what comes back |
| Skills & Commands | Each command in detail — when to run it, what it returns, how to iterate |
| Custom Context | Code-graph enrichment, the baseline branch, and Jira / GitHub ticket resolution |
| Tool Reference | The HTTP endpoints and MCP tool signatures |
Core ideas¶
- Pre-commit. The unit of review is the complete change you are about to commit — including brand-new files you have not added to the repository yet, not just edits to tracked files.
- Session-based. One developer working on one repository has exactly one review session. The diff is uploaded to the server once and referenced by session id, so running Inspect then Describe in the same turn reuses one upload.
- Agent-native. Findings come back as structured data your agent acts on — it can fix what you ask it to and re-run the pass against the same session.