Skip to content
DiffGuardianKnow Your Code
NativemacOS · Windows · Linux

Never merge AI code
you don't understand.

Your agents ship 5,000-line pull requests before lunch. DiffGuardian reads each one against your entire codebase, breaks it into the features it actually ships, and walks you through them — so you keep human oversight of what merges without becoming the bottleneck.

  • Your code never touches our servers
  • Bring your own keys, or stay fully local
The flagship feature

It traces the PR, one feature at a time

Hit Trace It! and a 5,000-line PR becomes a guided tour — broken into the features it actually ships, each explained in plain English. You approve having seen the whole thing, not skimmed a diff.

See it explain your whole codebase
Your models, your rules

Bring your own keys.
Or never leave your laptop.

Every feature flows through one router: Anthropic → OpenAI → Ollama → LM Studio. Pick a model per job — a cheap fast model for hunk insights, a premium one for the full review, a local one for chat — and ground any of them in a private knowledge base of your team's style guide and RFCs.

  • BYOK for Anthropic and OpenAI — you pay the provider, no markup
  • Ollama, LM Studio and any OpenAI-compatible endpoint out of the box
  • Local vector knowledge bases — embeddings never leave your machine
  • Reviews cite your team's RFCs, not the model's training data
  • All credentials live in your OS keychain, encrypted at rest
Routing this review through…
Your machinellama-3.3-70b · local
Anthropicnot contacted
OpenAInot contacted
0bytes left your machine
0tokens in cloud cache
100%keys held in your OS keychain
Understand the entire codebase

Ask anything about the whole codebase.

DiffGuardian indexes your entire repo into a code knowledge graph. Ask a question in plain English — how auth works, where a payment is verified, what a service touches — and it explains the answer in a way that actually makes sense, then backs it with a diagram drawn from the actual call graph.

Diagram · drawn from your codeClientAPI GatewayAuth ServiceVaultrequest · session keyresolve(sessionKey)getCredential()tokeninject auth header200 · response
Ask about this project
How does authentication work across the app?

Every request carries a session key. The API resolves the real token from the vault, injects the provider's auth header, then makes the call — so the client never touches a raw secret.

Auth flow · view on canvas
Where is the Stripe webhook verified?
  1. 01Ask in plain EnglishAny question about any corner of the repo — not just the lines in the diff.
  2. 02Explained so it makes senseClear answers grounded in your actual code, with the real files cited.
  3. 03Backed by a visual aidThe knowledge graph draws the flow as a diagram you can actually follow.
Pricing

Free to start. Upgrade for the full AI suite.

Every plan reviews locally. The cloud models, knowledge bases and Trace It unlock on Pro.

Basic
Free

Forever. No card required.

Download
  • 1 connected public repo
  • 5 AI reviews per month
  • Locally-hosted models only
  • AI reviews & summaries
  • Dependency graphs
  • Stack maps
Team
$8/seat/mo

Same price per seat, 5 seat minimum. Or $80 a seat a year.

Manage team seats
  • Everything in Pro, for everyone
  • Centralised seat billing for the whole team
  • Seat management and invoicing
  • Minimum of 5 seats

Set up by an organisation admin, who buys the seats for the team. Everyone else just downloads the app — individual sign-in happens inside DiffGuardian during onboarding, never here.

Compare plans

Every feature, side by side.

FeaturesBasicFreePro$8/moTeam$8/seat/mo
Repositories & limits
Connected repositories1, public onlyUnlimitedUnlimited
AI reviews5 / monthUnlimitedUnlimited
Models
Locally-hosted models (Ollama, LM Studio)
Cloud providers with your own keys (Anthropic, OpenAI)
Review & understanding
AI reviews & summaries
Dependency graphs
Stack maps
Trace It — guided walkthrough of a diff
Ask — question your project and PRs
Explain & Polish
Analyze Project
Knowledge bases & RAG
MCP connections & MCP server
Voice mode
Teams
Centralised seat billing5 seat minimum
Questions

Your questions, answered.

How is this different from the AI review tools I already have?
Plenty of tools review code with AI now, and several are genuinely good at it. But nearly all of them hand the judgment straight back to a model: it leaves the comments, suggests the fixes, and you click approve on its say-so. DiffGuardian is built for the opposite outcome — its whole job is to make sure the person reviewing actually understands the PR before they approve it. The point isn't to automate the review away; it's to keep the one gate that still matters in human hands.
Is my code private?
Your code never reaches a DiffGuardian server. Reviews run on your own machine, and cloud reviews go directly from your machine to the provider whose key you supplied — under your own agreement with them. Because local models are supported on every plan, you can use the AI features without a single line of code leaving your laptop. See the Security page for the full data-flow breakdown.
Which AI providers are supported?
Any locally-hosted model you run yourself — Ollama, LM Studio, vLLM, llama.cpp, or any OpenAI-compatible endpoint including an internal gateway. Anthropic and OpenAI are available on Pro with your own API key. You pay the provider directly; there is no markup.
Do I need an internet connection?
Not for local models. Point DiffGuardian at Ollama or LM Studio and you can review a diff on a plane. Cloud models obviously need the network, and DiffGuardian falls back to your local model automatically when a cloud provider is unreachable.
Which platforms is DiffGuardian available on?
macOS, Windows and Linux. Each one is a real desktop build against the platform's own UI toolkit rather than a browser shell — that is why it launches instantly and respects the conventions you already know. The same features, the same model routing and the same local-only mode are available on all three.
How does the acceptance-criteria filter actually work?
The AI reads the PR description and any linked ticket, then extracts the criteria as a list. Pick one and DiffGuardian filters the file tree and hunks down to just the code that implements it. Hunk insights re-phrase themselves under that scope, so the same change reads differently when you're checking “login flow” versus “telemetry”.
How does billing work, and can I cancel?
Pro is $8 a month, or $80 a year — two months free. Team is the same price per seat: $8 a seat a month or $80 a seat a year, with a minimum of five seats. Cancel any time from the billing portal — you keep access until the end of the period you already paid for.