PR Review CLI

Offline-first Pull Request review CLI tool using local LLM

A command-line tool that uses local Large Language Models (LLMs) to automatically review your code changes. Everything runs offline on your machine—no data leaves your computer, ensuring complete privacy and security.

Description

PR Review CLI is a privacy-focused code review tool that leverages local LLMs to analyze git diffs and provide intelligent code review feedback. It integrates seamlessly with your development workflow, supporting both local usage and CI/CD pipelines.

Key Features

Offline-first: All analysis happens locally by default—no data transmission to external services
AI-Powered Reviews: Uses local LLMs (Ollama, vLLM, llama.cpp) for intelligent code analysis
Detailed Reports: Generates comprehensive reviews with severity levels, categories, and actionable suggestions
Flexible Configuration: Customize LLM settings, review parameters, and output formats
CI/CD Ready: Integrates with GitHub Actions and other CI/CD platforms
Docker Support: Run in containers with pre-configured setups
Multiple Formats: Output in JSON, Markdown, or terminal-friendly text

What it does

Reviews code changes between git branches using a local LLM. It analyzes diffs, finds potential issues (bugs, security vulnerabilities, code quality issues), and provides structured feedback in a format you can use in CI/CD or locally.

Installation

From npm

npm install -g pull-request-reviewer-ai

From source

git clone https://github.com/marcinkantyka/pull-request-reviewer-ai.git
cd pull-request-reviewer-ai
npm install
npm run build

After building, use node dist/cli/index.js instead of pr-review, or run npm link to install it globally.

Quick start

1. Set up a local LLM

You'll need a local LLM server running. Ollama is the easiest option:

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull a code review model
ollama pull deepseek-coder:6.7b

2. Run a review

# Review current branch against main
pr-review review --base main

# Compare two specific branches
pr-review compare feature-branch main

# Save results to a file
pr-review compare feature-branch main --format json --output review.json

That's all you need. The tool connects to your local LLM (defaults to http://localhost:11434) and analyzes your changes.

Example Output

Here's what a typical review looks like:

════════════════════════════════════════════════════════════════════════════════
  Code Review Report
════════════════════════════════════════════════════════════════════════════════

  Generated:     2026-02-01T23:17:28.119Z
  Source Branch: fix/repo_rename
  Target Branch: main
  Model:         qwen2.5-coder:7b
  Duration:      30349ms

  Summary
  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─
  Files Reviewed: 8
  Total Issues:   4

  Issues by Severity:
    Medium: 1
    Info: 3

  Score: 9.3/10

  Issues by File
  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─  ─

  File: .github/workflows/ci.yml (yaml) +3 -3

    [INFO] [maintainability]:88
       The Docker image name has been changed to 'pull-request-reviewer-ai:test' but the associated GitHub Secrets should also be updated for consistency.
       Suggestion: Update ${ secrets.DOCKER_USERNAME } in .github/workflows/release.yml to match the new image name.

  File: QUICKSTART.md (markdown) +2 -2

    [INFO] [style]:14
       The URL has been changed from 'pr-review-cli' to 'pull-request-reviewer-ai'. Ensure this change is intentional and that the repository name and description are updated accordingly.
       Suggestion: Verify that the new repository name and description are accurate and update them if necessary.

  File: examples/ci-integration.yml (yaml) +1 -1

    [INFO] [maintainability]:35
       The tool name has changed from 'pr-review-cli' to 'pull-request-reviewer-ai'. Ensure that the new tool is compatible with the existing workflow.
       Suggestion: Verify compatibility and update any documentation if necessary.

  File: src/cli/index.ts (typescript) +17 -1

    [MEDIUM] [maintainability]:17
       The version is hardcoded and falls back to '1.1.0' if the package.json cannot be read.
       Suggestion: Consider using a default value or log an error message when the version cannot be determined.

════════════════════════════════════════════════════════════════════════════════

The report includes:

Summary: Overview of files reviewed, total issues, and score
Issues by Severity: Breakdown of critical, high, medium, low, and info issues
Issues by File: Detailed findings for each file with line numbers, severity, category, and suggestions

Configuration

This tool reads configuration from (in order):

CLI flag --config /path/to/config.yml
Project root config files: pr-review.config.json / pr-review.config.yaml / pr-review.config.yml / pr-review.config.js / pr-review.config.ts / .pr-reviewrc / .pr-reviewrc.json / .pr-reviewrc.yaml / .pr-reviewrc.yml / package.json (pr-review key)
Environment variables (override file values)

Quick Start (Deterministic + Offline-Safe)

To make runs repeatable and keep everything local:

Set temperature: 0
Set seed to a fixed number
Use changeSummaryMode: deterministic
Keep NETWORK_ALLOWED_HOSTS to localhost or internal Docker hostnames

Example pr-review.config.yml:

llm:
  endpoint: 'http://localhost:11434' # Ollama default
  provider: 'ollama' # Options: ollama, vllm, llamacpp, openai-compatible, mock
  model: 'deepseek-coder:6.7b'
  temperature: 0
  topP: 1
  timeout: 60000 # Milliseconds
  maxTokens: 2048
  apiKey: '' # Optional, for secured endpoints
  seed: 42 # Deterministic outputs
  retries: 3
  retryDelay: 1000

network:
  allowedHosts:
    - 'localhost'
    - '127.0.0.1'
    - '::1'

review:
  maxFiles: 50
  maxLinesPerFile: 1000
  excludePatterns:
    - '*.lock'
    - '*.min.js'
    - '*.min.css'
    - 'node_modules/**'
    - 'dist/**'
    - 'build/**'
    - '.git/**'
  includeAllFiles: false # Set true to ignore exclude patterns and size limits
  changeSummaryMode: deterministic # deterministic | llm
  projectContext: '' # Optional: domain rules or architecture notes to guide reviews

  # Context-aware review options
  contextAware: true # Enable multi-file context review
  groupByDirectory: true # Group files in same directory
  groupByFeature: true # Group files by feature/module
  maxGroupSize: 5 # Maximum files per group
  directoryDepth: 2 # Directory levels for grouping
  concurrency: 3 # Parallel review groups

output:
  defaultFormat: 'text'
  colorize: true
  showDiff: false

git:
  diffContext: 3
  maxDiffSize: 10485760 # 10MB

server:
  host: '127.0.0.1'
  port: 0 # 0 = random free port

Environment Variables

You can also use environment variables instead of a config file:

export LLM_ENDPOINT=http://localhost:11434
export LLM_MODEL=deepseek-coder:6.7b
export LLM_PROVIDER=ollama
export LLM_API_KEY=your-key-here       # Optional
export LLM_TIMEOUT=60000
export LLM_SEED=42                     # Optional
export LLM_TOP_P=1
export LLM_MAX_TOKENS=2048
export REVIEW_PROJECT_CONTEXT="Explain domain constraints or key rules"
export NETWORK_ALLOWED_HOSTS=ollama,localhost,127.0.0.1,::1
export UI_HOST=127.0.0.1
export UI_PORT=47831

Run pr-review config init to generate a default config file with all the available options.

Project Context (Recommended)

review.projectContext (or REVIEW_PROJECT_CONTEXT) lets you inject short, project-specific rules that the model must respect. This reduces false positives and aligns findings with how your system really works.

Use it to document architecture constraints, infrastructure protections, or known tradeoffs. Example:

review:
  projectContext: >
    Do not flag exposed ports as security issues; the service is private behind an ALB.
    Authentication is handled by the gateway, not in this repo.

Or with environment variables:

export REVIEW_PROJECT_CONTEXT="Do not flag exposed ports as security issues; the service is private behind an ALB."

Common Configuration Tasks

Set deterministic output:

llm.temperature: 0
llm.seed: <fixed number>
llm.topP: 1
review.changeSummaryMode: deterministic

Review every file (ignore filters):

review.includeAllFiles: true

Use a different config file:

pr-review review --config /path/to/config.yml --base main

Allow internal Docker hostnames:

Add hostnames to network.allowedHosts or set NETWORK_ALLOWED_HOSTS

Commands

Get help for any command:

pr-review --help              # Show general help
pr-review review --help       # Show help for review command
pr-review compare --help      # Show help for compare command
pr-review config --help       # Show help for config command

`review`

Review the current branch against a base branch:

pr-review review --base main
pr-review review --base develop --format json --output review.json

`compare`

Compare two specific branches:

pr-review compare feature-branch main
pr-review compare feature-branch main --severity high --max-files 20

`config`

Manage configuration:

pr-review config init                    # Create default config file
pr-review config init --output custom.yml # Create config with custom name
pr-review config get llm.endpoint        # Get specific config value
pr-review config list                    # List all configuration

Options

Global options:

--server - Start the local UI server
--host <host> - UI server host (default: 127.0.0.1)
--port <port> - UI server port (default: 47831; set 0 for random)

Common options available for both review and compare commands:

--repo-path <path> - Repository path (default: current working directory)
--format <json|md|text> - Output format (default: text)
--output <file> - Save to file instead of stdout
--config <file> - Path to custom config file
--severity <all|high|critical> - Filter by severity level (default: all)
--files <pattern> - File pattern to review (glob pattern)
--max-files <number> - Limit number of files to review
--timeout <seconds> - LLM timeout in seconds (default: 60)
--verbose - Show detailed logs
--no-color - Disable colored output
--exit-code - Exit with code 1 if issues found (useful for CI)

Review Command Options

--base <branch> - Base branch to compare against (default: main)

Compare Command

Takes two required arguments:

<source-branch> - Source branch to review
<target-branch> - Target branch to compare against

Local UI (optional)

Start a local web UI without changing the default CLI workflow:

pr-review --server

Default port is 47831. If that port is busy, the server falls back to a random free port. You can always force a random port with --port 0. The UI is local-only by default. --server cannot be combined with other commands.

In UI mode, only JSON/YAML config files are accepted.

The UI compares committed branch diffs. Uncommitted working tree changes are not included. If you use Ollama or another local provider, the UI can load available models and detect local servers automatically. You can also generate a full pr-review.config.yml template from the UI (Advanced panel).

UI Quickstart

Run pr-review --server and open the URL printed in the console.
Pick your repository path (or use the current directory).
Choose a base branch (default is main).
(Optional) Load models and pick an LLM model.
Click Run Review to generate results.

Tip: Use Advanced to set project-specific context and other overrides for a single run.

LLM providers

Works with any OpenAI-compatible API. Tested with:

Ollama (recommended) - http://localhost:11434
vLLM - http://localhost:8000
llama.cpp server - http://localhost:8080
LM Studio / LocalAI - http://localhost:1234

Set the provider and endpoint in your config or via environment variables. For OpenAI-compatible servers, use the base URL (the tool will call /v1/models and /v1/chat/completions).

Docker

Quick Start

The easiest way to get started with Docker:

cd docker
./start.sh

The script handles everything for you:

Downloads models if needed (only the first time)
Starts services with secure internal network (no internet access)
Blocks all outbound traffic for security

Manual Docker Usage

# Build
docker build -f docker/Dockerfile -t pull-request-reviewer-ai .

# Run standalone
docker run --rm \
  -v $(pwd):/repos:ro \
  -e LLM_ENDPOINT=http://host.docker.internal:11434 \
  pull-request-reviewer-ai compare feature main

Docker Compose

For a complete setup with Ollama and secure network isolation:

cd docker
docker-compose up
# or
docker compose up

The compose setup uses internal: true network mode, which completely blocks internet access from containers. Models are automatically downloaded on first run via start.sh. Check out docker/README.md for more details.

CI/CD

GitHub Actions

See examples/ci-integration.yml for a complete example. Here's a basic setup:

name: Code Review

on:
  pull_request:
    branches: [main, develop]

jobs:
  review:
    runs-on: ubuntu-latest

    services:
      ollama:
        image: ollama/ollama:latest
        ports:
          - 11434:11434
        options: >-
          --health-cmd "curl -f http://localhost:11434/api/tags || exit 1"
          --health-interval 10s
          --health-timeout 5s
          --health-retries 5

    steps:
      - name: Checkout code
        uses: actions/checkout@v4
        with:
          fetch-depth: 0 # Full history for git operations

      - name: Setup Node.js
        uses: actions/setup-node@v4
        with:
          node-version: '20'
          cache: 'npm'

      - name: Install PR Review CLI
        run: npm install -g pull-request-reviewer-ai

      - name: Wait for Ollama
        run: |
          until curl -f http://localhost:11434/api/tags; do
            echo "Waiting for Ollama..."
            sleep 2
          done

      - name: Pull Ollama model
        run: |
          curl http://localhost:11434/api/pull -d '{"name": "deepseek-coder:1.3b"}'
        timeout-minutes: 10

      - name: Run code review
        run: |
          pr-review compare ${{ github.head_ref }} ${{ github.base_ref }} \
            --format json \
            --output review.json \
            --exit-code
        env:
          LLM_ENDPOINT: http://localhost:11434
          LLM_MODEL: deepseek-coder:1.3b

#### Minimal offline-safe example

This version explicitly restricts outbound connections to the local Ollama service.

```yaml
name: Offline-Safe Review

on:
  pull_request:
    branches: [main]

jobs:
  review:
    runs-on: ubuntu-latest

    services:
      ollama:
        image: ollama/ollama:latest
        ports:
          - 11434:11434

    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0

      - uses: actions/setup-node@v4
        with:
          node-version: '20'
          cache: 'npm'

      - run: npm install -g pull-request-reviewer-ai

      - run: pr-review compare ${{ github.head_ref }} ${{ github.base_ref }} --format md
        env:
          LLM_ENDPOINT: http://localhost:11434
          LLM_MODEL: deepseek-coder:1.3b
          NETWORK_ALLOWED_HOSTS: localhost,127.0.0.1,::1

For a more complete example with PR comments, see examples/ci-integration.yml.

Security

The tool only allows connections to localhost by default. If you try to connect to an external host, it will be blocked unless you explicitly allow it in NETWORK_ALLOWED_HOSTS. Everything runs locally on your machine—your code never leaves your computer unless you configure a non-local endpoint.

The local UI server only accepts loopback connections, even if you bind to 0.0.0.0.

The local UI server only accepts loopback connections. Remote network requests are blocked even if you bind to 0.0.0.0.

Troubleshooting

LLM provider is not available

If you see this error, make sure your LLM server is actually running. For Ollama, try:

ollama serve
curl http://localhost:11434/api/tags

Timeout errors

If reviews are timing out, you can increase the timeout. Either use the CLI flag:

pr-review compare feature main --timeout 120

Or set it in your config file (timeout is in milliseconds):

llm:
  timeout: 120000 # 2 minutes

Network security errors

The tool only allows localhost connections. Make sure your endpoint uses localhost, 127.0.0.1, or ::1. Any other hostname will be rejected.

Development

# Install dependencies
npm install

# Build
npm run build

# Run tests
npm test

# Lint
npm run lint

# Type check
npm run typecheck

Contributing

We welcome contributions! Please see our Contributing Guide for details on how to contribute to this project.

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 78 Commits
.github		.github
docker		docker
examples		examples
src		src
tests		tests
.DS_Store		.DS_Store
.dockerignore		.dockerignore
.editorconfig		.editorconfig
.eslintrc.json		.eslintrc.json
.gitignore		.gitignore
.npmignore		.npmignore
.prettierrc		.prettierrc
LICENSE		LICENSE
QUICKSTART.md		QUICKSTART.md
README.md		README.md
eslint.config.js		eslint.config.js
package-lock.json		package-lock.json
package.json		package.json
pr-review.config.yml		pr-review.config.yml
tsconfig.json		tsconfig.json
tsup.config.ts		tsup.config.ts
vitest.config.ts		vitest.config.ts

License

marcinkantyka/pull-request-reviewer-ai

Folders and files

Latest commit

History

Repository files navigation

PR Review CLI

Description

Key Features

What it does

Contents

Installation

From npm

From source

Quick start

1. Set up a local LLM

2. Run a review

Example Output

Configuration

Quick Start (Deterministic + Offline-Safe)

Environment Variables

Project Context (Recommended)

Common Configuration Tasks

Commands

review

compare

config

Options

Review Command Options

Compare Command

Local UI (optional)

UI Quickstart

LLM providers

Docker

Quick Start

Manual Docker Usage

Docker Compose

CI/CD

GitHub Actions

Security

Troubleshooting

Development

Contributing

License

About

Topics

Resources

License

Code of conduct

Contributing

Security policy

Uh oh!

Stars

Watchers

Forks

Releases 3

Packages 0

Uh oh!

Uh oh!

Contributors

Uh oh!

Languages

`review`

`compare`

`config`

Packages