LOCKE FOR OLLAMA
Run models locally with Ollama — and still control what enters the context.
Ollama is the privacy-conscious choice on its face: models run on your own hardware, so prompts never reach a model vendor's servers. That genuinely closes one risk — provider access to your data — and for many use cases it's the right architecture. But it's easy to read "local model" as "problem solved," and that's the trap this page exists to flag.
Self-hosting changes who can see the data, not whether the data can be seen. Sensitive fields still flow into the model context in plaintext; they can be persisted in your own request logs, captured in shell history, or echoed into transcripts that get shared, synced, or backed up. And almost no team runs Ollama in isolation — the same people also use ChatGPT, Claude, or Gemini in the browser, where the provider gap is wide open again.
Locke operates on the data itself, independent of where the model runs. It detects and masks sensitive values before they enter any model context — local or remote — so your Ollama prompts stay clean in your own logs, and the same control applies the moment you switch to a cloud tool.
The risk with Ollama
Running a model locally with Ollama removes the provider's access to your prompts, but the sensitive data still enters the model context in the clear — and can persist in your own logs, shell history, and transcripts, while any cloud AI tools your team also uses remain fully exposed.
How Locke helps
Locke governs what enters the model context, not who hosts the model — masking sensitive fields before they reach Ollama or your logs. Locke — the Sonomos desktop app, coming soon — applies one consistent control across local models and cloud AI tools alike.
Keep using Ollama — without the exposure
Locke runs entirely on your device. Sensitive data is detected and masked before any prompt is sent, so nothing confidential ever leaves your machine. Pricing for Locke, the desktop app, is coming soon; Canary is free and open source.
