guide · local ai code review
Experimental Ollama compatibility with DeepSeek-Coder
A privacy-first compatibility setup for exercising a model endpoint you control. In v1.0.4 this is an experimental provider smoke, not the promoted live-review lane.
Local inference
With the default loopback endpoint, review prompts and model inference stay on the review machine. GitHub posting remains optional and explicit.
Your hardware
Runs on an M-series MacBook, a Linux workstation with a consumer GPU, or a shared server on your internal network.
Compatibility contract
Provider doctor can exercise endpoint readiness. It does not promote Ollama review output, inline comments, or posting support in v1.0.4.
Why local?
Cloud-hosted AI code review tools ingest your diff on vendor infrastructure. That's fine for public code, and a hard non-starter for anything under NDA, in a regulated industry, or that your team considers a product secret. A local pipeline removes the hosted-model inference path while keeping deployment under your control.
Ollama is a small runtime that serves open-weight models behind an OpenAI-compatible HTTP endpoint. DeepSeek-Coder is one of the strongest open code models for review-style tasks — it reasons well about diffs, catches real bugs, and runs on hardware you probably already own.
1. Install Ollama
macOS / Linux — one line:
curl -fsSL https://ollama.com/install.sh | shOn macOS you can also download the .dmg from ollama.com. On Windows, use the official installer.
2. Pull DeepSeek-Coder
Grab a model sized for your hardware. Start with the 16B variant — it fits on 16 GB of unified memory and gives you a practical starting point for review quality:
ollama pull deepseek-coder-v2:16bOn a workstation with 24 GB+ VRAM, pull a higher-quality quant:
ollama pull deepseek-coder-v2:16b-lite-instruct-q6_KOllama serves on http://127.0.0.1:11434 by default. Confirm with ollama list.
3. Smoke-test the endpoint
Confirm the OpenAI-compatible route is live before wiring anything else up:
curl http://127.0.0.1:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-coder-v2:16b",
"messages": [
{ "role": "user", "content": "Review this diff: - return x + 1\n+ return x - 1" }
]
}'4. Point NeonDiff at Ollama
Install NeonDiff and initialize a local config for compatibility testing. Licensing is the same as on every other path: from NeonDiff 1.1.0, public repositories are free through the GitHub Action, neondiff review-pr --ci, and the self-hosted worker (daemon, run-once, and local review-pr). The Mac app still needs a license key for now. Private and internal repositories need an active license; a repository whose visibility can't be verified is not reviewed. This guide uses Ollama for provider doctor compatibility checks, not live review. If the worker did review with a non-Codex runtime such as Ollama, pull requests from outside contributors on public repositories would be skipped, because restricted review needs Codex. For the Mac app, for non-public review, or on npm CLI 1.0.x (which requires activation for every repository), complete API-backed activation before provider or review work in this local Ollama setup. Pass the activation key only from an approved secret store over bounded stdin:
curl -fsSL https://www.neondiff.com/install | sh
neondiff init --config config.local.json
security find-generic-password -s YOUR_APPROVED_SOURCE -w | neondiff license activate --config config.local.json --license-key-stdin true --jsonThen enable the local provider in config.local.json. Ollama on loopback normally uses authMode: "none"; do not invent a fake provider key:
{
"providers": {
"providers": {
"ollama-local": {
"enabled": true,
"adapter": "openai-compatible",
"baseUrl": "http://localhost:11434/v1",
"model": "deepseek-coder-v2:16b",
"authMode": "none",
"structuredOutputMode": "ollama-format-json-schema",
"capabilities": {
"review": true,
"jsonOutput": true,
"local": true,
"streaming": false
}
}
}
}
}
neondiff providers doctor --config config.local.json --provider ollama-local --smoke true --jsonThe Ollama compatibility interface is experimental in v1.0.4; it is not yet a promoted live-review claim. Codex CLI is the current promoted review lane; the ZCode-backed engine remains a fallback while the local adapter completes its promotion and evidence work. Use provider doctor as a compatibility smoke only, keep posting disabled, and follow the current provider-status documentation before relying on Ollama for review output.
5. Run a dry-run review
Keep posting disabled until activation, GitHub App readiness, and the selected provider smoke all pass. Then run one current-head dry run:
neondiff doctor github --config config.local.json --json
neondiff review-pr --config config.local.json --repo owner/project --pr 42 --dry-run trueTuning notes
- Context window. DeepSeek-Coder V2 handles ~128k tokens; keep per-file review chunks under ~8k for latency.
- Quantization. Q4 for laptops, Q6 for GPU workstations, Q8 if you have the VRAM. Higher quants noticeably reduce false-positive nits.
- Warm the model. Ollama unloads models after idle; a nightly cron with a dummy request keeps first-review latency low.
FAQ
Run this pipeline on your next PR
Local-first AI code review. Bring your own model, keep your source code on your machine, review every PR.
Get NeonDiff →