Jev Review plugin
A plugin with a local MCP server: the agent scores its own code on a set of criteria through Jev and improves it by the numbers
- 247 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Sends the task, a diff and chosen files to the remote Jev API on typesafe.ai, so proprietary code leaves the machine
- Reads JEV_API_KEY from the environment and passes it in the request header to Jev
- Does not change code itself; the primary agent makes all edits
Install
- Works in
Manual install
export JEV_API_KEY="your-key"Get the key in the TypeSafe console. Set it before starting the agent.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Sends the task, a diff and chosen files to the remote Jev API on typesafe.ai, so proprietary code leaves the machine
- Reads JEV_API_KEY from the environment and passes it in the request header to Jev
- Does not change code itself; the primary agent makes all edits
This is third-party code. Review the repository files before installing.
What it does
Jev Review gives an agent one tool, `jev_review`, that sends the task, a diff and chosen files to TypeSafe's Jev service and returns structured quality scores. The agent stays the author of the changes. The plugin supplies only a numeric signal: scores across dimensions such as readability, security, testability, error handling, duplication and project structure, plus confidence levels and coarse rubric hints. Alongside the MCP server the plugin ships a `jev-review` skill that sets the working loop. The agent writes a coherent slice of implementation and runs checks, takes the first score as a baseline, inspects the weak important spots itself and makes the smallest justified improvement. It then scores again, passing `previousEvaluation`, and sees what rose and what fell. There is deliberately no overall figure like "82/100"; the dynamics per dimension are shown instead, for example readability from 6.3 to 8.1. The skill tells the agent to stop once requirements and checks pass and further score-chasing adds little value. Correctness and the user's requirements always outrank the score. The server runs locally over stdio on Node.js, the plugin installs for Claude Code, Codex and Cursor, and OpenCode has a manual setup.
Who it is for. For developers who hand larger changes to Claude Code, Codex, Cursor or OpenCode and want an external numeric quality check on the result.
Good fit when
- The agent has finished a nontrivial change and before handing it over you want an objective read on readability, duplication and error handling
- You want to see the dynamics after a refactor: per-dimension scores before and after instead of a general impression
- You want to catch weak spots in security and testability of fresh code before your colleagues review it
- You use several agents and want one quality yardstick for all of them through a single MCP tool
- You already have a Jev API key and the code must not pass through an intermediary: requests go straight to Jev from your machine
Not a fit when
- The code must not be sent to an external service: the chosen task, diff and files go to the Jev API
- You need a written explanation of why a score is low: Jev returns typed decisions, not a prose essay
- You want a replacement for vulnerability scanners: the plugin complements dedicated security tools and does not replace them
- You have no Jev API key from the TypeSafe console: the tool will not run without it
Example request
Finish the CSV export, run it through jev-review and improve the weak dimensions until the scores stop rising noticeablyLimitations
- Node.js 20 or newer and a Jev API key from the TypeSafe console in the JEV_API_KEY environment variable are required.
- Data from the `task`, `diff`, `files` and `repositoryContext` arguments goes to the remote Jev API. Do not pass secrets or unrelated proprietary code in them.
- The plugin is distributed only from the GitHub repository; there is no npm package.
- OpenCode has no plugin install, only a manual MCP setup.
- Jev does not explain a score in prose: it returns scores, confidence levels and coarse rubric hints.
How to disable. In Claude Code disable or remove the plugin through /plugin, or run `claude mcp remove jev-review` if you connected MCP by hand. For Codex and Cursor delete the jev-review block from `~/.codex/config.toml` or `~/.cursor/mcp.json`. You can revoke the JEV_API_KEY key in the TypeSafe console.
FAQ
How do I install Jev Review in Claude Code?
Set the key with `export JEV_API_KEY="your-key"`, then run `npx plugins add NiazMorshed2007/jev-review --target claude-code`. After a restart, check the connection with /mcp. For Codex and Cursor only the --target value changes.
Where does my code go when jev_review runs?
Only to the Jev API at api.typesafe.ai: the task, diff, files and repositoryContext arguments the agent supplies explicitly. Repository files are not discovered or uploaded automatically. The plugin has no hosted backend or telemetry.
Which quality dimensions does Jev Review score?
Changeability, abstraction and API design, project structure, duplication, maintainability, testability, reliability and error handling, security, consistency and documentation. When evidence is present it adds performance, scalability, compatibility and observability.
Why is there no single overall score?
The author decided that the dynamics of individual dimensions are more useful than a blended figure. For example, readability rose from 6.3 to 8.1 while security stayed at 8.2.
How do I rescore code after a fix?
Call `jev_review` again with the `previousEvaluation` parameter. The earlier evaluation is compared locally and is not included in the request to Jev.
Discussion
No questions or reviews yet.
Related
PluginsThe official CodeScene MCP server scores Code Health locally and tells an agent what to refactor and where tech debt hides
Local review of agent changes before commit: the agent itself reads line comments and replies to them through MCP.
A code-review CLI and plugin for agents: reads a git diff and writes line-level comments with context
A Claude Code plugin that hands a task to a local no_human and reviews a branch with a reviewer that never saw the coder's session
Local code structure graph: builds a repository map with tree-sitter and serves the assistant only the context a change actually touches over MCP
A Codex skill for reviewing AI-written code: explains blocks, names files and lines, checks goals against evidence.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Jev Review (this card) | no | 247 | ||
| brooks-lint | no | 1,561 | code review grounded in classic engineering books with source citations, with no external scoring | |
| code-review-graph | no | not required | 32,612 | a local repository structure graph for review, with no quality scoring through an external service |
README in short
The page opens with a quick start covering requirements and the install command, then shows the evaluation loop and setup for Claude Code, Codex, Cursor and OpenCode. It continues with the list of quality dimensions, the agent's workflow, the repository layout and a section on security and privacy.