Jev Review

Continuous software-quality review for AI coding agents, powered by Jev.

Quick start · Client setup · Quality dimensions · Security

Jev Review runs as a local MCP server and gives Claude Code, Codex, Cursor, and OpenCode structured quality scores while they work. Your coding agent remains responsible for diagnosing weaknesses and changing the code; Jev supplies a fast scalar signal across correctness, complexity, changeability, modularity, tests, security, and other independent quality dimensions.

[!IMPORTANT] Your API key stays on your machine. Jev Review has no hosted backend, database, telemetry service, or author-operated proxy. The only remote request is sent directly to the configured Jev API.

Demo

https://github.com/user-attachments/assets/0ff9f873-0652-4826-af3d-6bb4f42c70b1

At a glance

Quick start

Requirements:

• Node.js 20 or newer
• A Jev API key from the TypeSafe console
• Claude Code, Codex, Cursor, or OpenCode

Set your API key before starting the coding agent:

export JEV_API_KEY="your-key"

Install Jev Review directly from GitHub—no npm publication is required:

npx plugins add NiazMorshed2007/jev-review

Choose your coding client when prompted, restart it, and ask the agent to use jev-review while implementing a nontrivial change.

How it works

flowchart LR
    A[Agent implements] --> B[Focused diff and context]
    B --> C[Jev Review MCP]
    C --> D[Jev evaluation]
    D --> E[Structured quality signals]
    E --> F[Agent improves the code]
    F -. review again .-> B

Jev Review is intended for frequent, focused checkpoints: after a coherent implementation slice, after a score-driven improvement, and before final handoff. The first call establishes a baseline. The agent then inspects its own implementation, forms a hypothesis about weak dimensions, improves the code, validates it, and rescores.

Jev returns typed Score, Choice, and Noul decisions rather than a free-form review essay. It does not generate a prose explanation of why a score is low. Jev Review validates and converts those decisions into metric scores, confidence levels, coarse rubric hints, and comparisons with a previous evaluation. The coding agent—not Jev—must determine the actual cause and appropriate code change.

There is deliberately no synthetic “82/100” overall score. Dimension changes such as Readability 6.3 → 8.1 and Security 8.2 → 8.2 are more useful than a blended percentage.

Client setup

Every client starts the same bundled dist/server.js process locally over stdio.

Claude Code

npx plugins add NiazMorshed2007/jev-review --target claude-code

Restart Claude Code and run /mcp to confirm that jev-review is connected.

To load a local clone while developing:

claude --plugin-dir /absolute/path/to/jev-review

Manual MCP-only setup:

claude mcp add --scope user jev-review -- node /absolute/path/to/jev-review/dist/server.js

Codex

npx plugins add NiazMorshed2007/jev-review --target codex

Restart Codex and run /mcp to verify the connection.

Manual setup in ~/.codex/config.toml:

[mcp_servers.jev-review]
command = "node"
args = ["/absolute/path/to/jev-review/dist/server.js"]
env_vars = ["JEV_API_KEY"]

Cursor

npx plugins add NiazMorshed2007/jev-review --target cursor

Restart Cursor and check Settings → MCP. The bundled skill is named jev-review; invoke it with /jev-review or leave it on Agent Decides.

Manual setup in ~/.cursor/mcp.json:

{
  "mcpServers": {
    "jev-review": {
      "type": "stdio",
      "command": "node",
      "args": ["/absolute/path/to/jev-review/dist/server.js"],
      "env": {
        "JEV_API_KEY": "${env:JEV_API_KEY}"
      }
    }
  }
}

If Cursor is launched from the macOS Dock, it may not inherit variables from your shell profile. Make the already-exported key available to GUI applications before starting Cursor:

launchctl setenv JEV_API_KEY "$JEV_API_KEY"

Verify without printing the key:

test -n "$(launchctl getenv JEV_API_KEY)" && echo "JEV_API_KEY is configured"

OpenCode

OpenCode does not currently appear in the portable plugins installer targets. Point it at the same bundled server instead:

git clone https://github.com/NiazMorshed2007/jev-review.git
cd jev-review
opencode mcp add jev-review --global -- node "$PWD/dist/server.js"

For the full skill and MCP setup, add this to ~/.config/opencode/opencode.json, replacing the absolute path:

{
  "$schema": "https://opencode.ai/config.json",
  "skills": ["/absolute/path/to/jev-review/skills"],
  "mcp": {
    "servers": {
      "jev-review": {
        "type": "local",
        "command": ["node", "/absolute/path/to/jev-review/dist/server.js"],
        "environment": {
          "JEV_API_KEY": "{env:JEV_API_KEY}"
        }
      }
    }
  }
}

Run opencode mcp list to verify the connection. OpenCode may display the tool as jev-review_jev_review; the underlying MCP tool is still jev_review.

MCP tool

Jev Review intentionally starts with one tool: jev_review.

{
  task?: string;
  diff?: string;
  files?: Array<{
    path: string;
    content: string;
  }>;
  repositoryContext?: string;
  previousEvaluation?: Evaluation;
}

At least one current-context field is required. Callers should normally send the task and focused diff, adding complete files only when the surrounding implementation is necessary to understand the change. Jev Review never reads the repository automatically.

Jev Review does not impose an additional character, token, or file-count limit. The Jev API currently enforces its own token ceiling: live jev-latest behavior indicates roughly 32,768 tokens for the submitted state, although this number is not published in the API documentation or OpenAPI schema and may change. When Jev returns max_tokens_exceeded, the server asks the agent to reduce unrelated context or split the change into coherent review slices.

The response contains:

• An independent 1–10 score and 0–1 confidence for each applicable metric
• { "applicable": false } for dimensions unsupported by the supplied context
• Prioritized weak dimensions and coarse predefined rubric hints—not generated root-cause explanations
• Per-metric deltas, improvements, regressions, and unresolved weaknesses when previousEvaluation is supplied

Quality dimensions

Always evaluated when the supplied context is sufficient:

• Correctness and requirement fit
• Cognitive complexity
• Readability and intent
• Modularity and cohesion
• Coupling and dependency quality
• Changeability and change amplification
• Abstraction and API design
• Project and file structure
• Duplication and reuse
• Maintainability
• Testability and test quality
• Reliability and error handling
• Security
• Consistency and conventions
• Documentation and explainability

Evaluated only when relevant evidence is present:

• Performance and resource efficiency
• Scalability and flexibility
• Compatibility and API stability
• Observability and operability

The evaluator judges consequences in context. It does not assume short functions, small files, zero duplication, more layers, more comments, or more tests are automatically better.

Evaluation workflow

The included jev-review skill teaches agents to treat Jev as a repeated scalar feedback loop:

1. Understand the task and inspect the repository.
2. Implement a coherent change and run relevant checks.
3. Call jev_review with focused context to establish a baseline.
4. Inspect the code themselves and form a hypothesis for weak important scores.
5. Make the smallest justified improvement and validate it.
6. Rescore with previousEvaluation, then inspect improvements and regressions.
7. Repeat while another evidence-based improvement remains.
8. Stop when requirements and checks pass and further score-seeking would add little real value.

Correctness and the user's requirements always outrank score improvement. A higher score never justifies speculative architecture, unnecessary abstraction, scope expansion, breaking behavior, meaningless tests, or needless rewrites.

Architecture

jev-review/
├── plugin.json                  # Portable Agent Plugin manifest
├── mcp.json                     # Portable stdio MCP definition
├── .claude-plugin/
│   └── plugin.json              # Claude Code adapter
├── .codex-plugin/
│   └── plugin.json              # Codex metadata
├── skills/
│   └── jev-review/
│       └── SKILL.md             # Age