Privacy and data handling
Exactly what leaves your machine and where it goes — and why your source code never reaches DiffGuardian's servers.
DiffGuardian is a native desktop app for macOS, Windows and Linux. Your repositories, diffs, and source never pass through DiffGuardian's infrastructure. This page summarises what leaves your machine and where it goes. The full statement lives at Security.
#Four things that are always true
Code never reaches us. Diffs are read from your local git and analysed on your machine. They are never sent to, processed by, or stored on DiffGuardian's servers.
A fully local option always exists. Run reviews against Ollama or LM Studio and your code never leaves your machine — no cloud, no network call, no exception. This is available on every plan.
Your keys, your agreement. With your own Anthropic, OpenAI, Google Gemini or Kimi key, code goes straight from your machine to the provider you already contract with, governed by your terms rather than ours.
Secrets stay encrypted and local. Every API key and code-host token is encrypted in your operating system's credential store (the Keychain on macOS) and used only on your device. They are never transmitted to our servers.
#What leaves your machine
| Data | Destination | Your control |
|---|---|---|
| Code and diffs for AI review | The AI backend you choose — local (stays on device) or your own cloud provider key (direct to Anthropic, OpenAI, Gemini or Kimi). Never our servers. | Pick the backend per feature; choose local to stay on-device. |
| Knowledge-base embeddings | On-device by default. Sent to a cloud provider only if you explicitly select cloud embeddings. | Use the built-in or a local embedding model to keep everything on-device. |
| Knowledge base and MCP server definitions | Nowhere. They are stored on your device, not in your DiffGuardian account. | Configured and deleted entirely in the app. |
| MCP tool calls | The MCP server you configured, called from your machine. Never proxied by us. | Turn MCP tool calls off in Settings → Features. |
| API keys and code-host tokens | Nowhere. Encrypted in your operating system's credential store, used only on your device. | Stored locally; remove them at any time. |
| Account and plan | DiffGuardian's server, to verify your sign-in and your plan. | Required to use a signed-in account. |
| Anonymous product analytics | A product-analytics provider — anonymous events, with personal and code fields stripped before sending. | Opt out in Settings. |
| Feedback you submit | Our error-reporting provider — only the text and optional diagnostics you choose to send. | Sent only when you submit feedback. |
What our servers never receive: your source code, diffs, AI prompts, AI responses, repository names, branch names, pull-request contents, your code standards, or your knowledge base and MCP configuration.
The hosted service also holds no opinion about how you use AI. It reports which plan you are on and nothing else — there is no server-side vendor allowlist, model allowlist, budget or quota for the app to fetch, and no endpoint that would carry one.
#Where reviews run
Every AI feature flows through one router, and you choose the backend per job. That choice alone decides where your code goes:
- Local models — Ollama and LM Studio run on your machine; requests go to
localhostonly. Nothing is transmitted anywhere. - Your own keys — sent from your machine directly to Anthropic, OpenAI, Google Gemini or Kimi, never relayed through DiffGuardian. Check each provider's API data-usage terms for how they treat what you send. For this flow DiffGuardian is not a data processor — the provider is your subprocessor.
- DiffGuardian's servers — used only for sign-in, billing, and service status. They never receive your code, diffs, prompts, or AI responses.
#Subprocessors
We rely on a small set of established providers, each handling only what its function requires, and none of them ever receive your code: identity and sign-in, payments, a managed database, product analytics, and error monitoring. A current list is available on request.
Your own integrations — your code host and your chosen AI provider — run under your agreements, not ours, and your code goes to them directly rather than through us.
#Compliance
DiffGuardian is an early-stage product and is not yet SOC 2 certified. The product is built so that the highest-risk data — your source code — is architecturally kept off our servers entirely, which removes the largest class of exposure rather than merely controlling it. See Security for the current position.