CodeReview AI

Paste source code, get a streamed review from a local LLM, and read the findings in the UI.

Nx monorepo layout:

apps/web                 React + Vite frontend (port 3000)
apps/api                 Express API (port 4200)
packages/shared-types    Shared TypeScript types, Zod, SSE wire format

Clone and run locally

You need:

• Git
• Node.js 22+ and npm 11+
• Ollama for real reviews, or skip it and use mock mode below

On Windows, use Git Bash or another shell that can run bash. The Ollama helper is a bash script.

git clone https://github.com/kevinsmithwebdev/aicr.git
cd aicr
cp .env.example .env
npm install
npm run ollama
npm run ollama:pull
npm run dev

npm run ollama finds the Ollama binary (including the default Windows install path, which Git Bash often misses), starts the local server on port 11434 if it is not already running, and lists installed models. npm run ollama:pull also downloads OLLAMA_MODEL from .env (default: qwen2.5-coder).

When Vite and Express are ready you should see:

Local:   http://localhost:3000/
[ ready ] http://localhost:4200
[ llm ] provider=ollama model=qwen2.5-coder

Open http://localhost:3000. Paste code on the left, pick a language and review focus, then click Review Code. The right panel streams Markdown as the model writes it. You can stop generation, ask a follow-up about the same snippet, Continue after a stop, Try Again, or Start Over.

Leave those processes running until you stop them with Ctrl+C.

To start the apps in separate terminals:

npm run serve:api
npm run serve:web

• Web: http://localhost:3000
• API health: http://localhost:4200/api/health
• Models: http://localhost:4200/api/models

The Vite dev server proxies /api to the Express app, so the browser can call /api/health and /api/reviews without CORS issues.

Without Ollama

Set LLM_PROVIDER=mock in .env to stream a canned review. You can skip npm run ollama and npm run ollama:pull.

Corporate TLS

If npm fails with UNABLE_TO_VERIFY_LEAF_SIGNATURE (corporate proxy or TLS inspection), retry with:

export NODE_OPTIONS="--use-system-ca"

Scripts

Script	What it does
npm run dev	Serve web + API together
npm run serve:web	Vite on port 3000
npm run serve:api	Express on port 4200
npm run ollama	Start local Ollama if needed
npm run ollama:pull	Download OLLAMA_MODEL

How it works

The browser talks only to the local API. The API talks only to Ollama, or to the mock provider if you set that. Submitted code stays in memory for the current review session and is gone when the API process exits.

A review goes through this path:

sequenceDiagram
    participant React
    participant Express
    participant Ollama

    alt New review
        React->>Express: POST /api/reviews
    else Follow-up
        React->>Express: POST /api/reviews/:id/follow-up
    end
    Express->>Express: Zod + prompt
    Express->>Ollama: stream chat
    Ollama-->>Express: tokens
    Express-->>React: SSE start / token / error / done

Loading

1. The React app posts the snippet, language, focus, optional file name, and optional extra instructions.
2. Express validates the body with Zod (shared schemas in packages/shared-types) and rejects oversized or invalid requests.
3. It builds a structured system prompt (summary, findings by severity, recommended changes) plus a user prompt that wraps the code.
4. It creates an in-memory session and asks the LLM provider to stream tokens. The default provider is Ollama (POST /api/chat). LLM_PROVIDER=mock returns a fixed review instead.
5. The API forwards those tokens to the browser as SSE events: start, token, error, done.
6. The UI appends tokens and renders Markdown as they arrive.

EventSource cannot POST, so the client uses fetch plus ReadableStream and parses the SSE wire format itself. Stop aborts the HTTP request and the Ollama call, and keeps any partial Markdown.

After a review starts, follow-ups reuse the same session (original code plus prior messages). Continue asks the model to pick up from the cutoff. Try Again replays the last request. Start Over clears the panel. Sessions are not persisted. Restarting the API forgets them.

Shared types, limits, and the SSE wire format live in packages/shared-types. LLM adapters live in apps/api/src/llm. The web app does not talk to Ollama directly.

Environment

Copy .env.example to .env next to the repo root. The API loads it at startup.

Variable	Default	Purpose
HOST	localhost	API bind host
PORT	4200	API port
WEB_ORIGIN	http://localhost:3000	CORS origin for the web app
LLM_PROVIDER	ollama	ollama or mock
OLLAMA_BASE_URL	http://localhost:11434	Ollama HTTP API
OLLAMA_MODEL	qwen2.5-coder	Default chat model