Your config never leaves your machine.
Your system prompt, tool schemas, and retrieval setup are intellectual property — often the most sensitive description of how your product works that exists anywhere. This page is written for the engineer who has to say yes to running agentsnap. It's short because the architecture makes it short.
Local-first by architecture, not by policy
agentsnap is a CLI you run on your own machine. There is no server side, no account, no upload path in the code. Snapshots are files in your repo. The strongest privacy guarantee is the absence of the feature that could violate it.
Secrets redacted at capture
Before hashing, before anything is written to disk: values under key-like field names (api_key, token, secret, password, …) and credential-shaped strings (sk-…, AKIA…, ghp_…, bearer tokens) are replaced with a digest.
No telemetry. No network calls. Nothing phoned home.
The tool makes zero network requests — it can run on an air-gapped machine. This website carries the same policy: no analytics, no trackers, no cookies.
You share only the report
The one-page HTML drift report is the only artifact designed to travel — and you review it before it does. Manifests, prompts, and probe transcripts stay where they were created: with you.
A rotated key is still a diff — without the value ever existing.
Redacted values are stored as <redacted>:sha256-…. The digest changes when the secret changes, so a rotation shows up in your change history with proof — while the secret itself is never written, hashed whole, or recoverable.
* [medium] Retrieval setting changed: api_key <redacted>:sha256-ff4dbedd -> <redacted>:sha256-2a94793b # the key was rotated on Wednesday. agentsnap can prove it — # and could not reveal the value even if asked to.
Auditable in an afternoon
agentsnap is stdlib-only Python across a dozen small files — no dependency tree to vet, no compiled blobs, no plugin loader. If your security review process is "an engineer reads the code," that review fits in one sitting. The snapshot + diff layer goes open-core on GitHub in July 2026, which makes that review even easier.
For design partners
- Mutual NDA available before any config is discussed — a template an individual can sign is ready to go.
- Delete-on-request, always. Any report or notes shared during the partnership are destroyed on request, no questions.
- You run the CLI. The default working mode is that your team executes agentsnap inside your environment; what I see is what you choose to forward.
The honest limits
Manifests live on your disk under your controls — treat the snapshots directory like any file containing your prompts: keep it out of public repos (snapshots/ ships gitignored) and inside your normal disk-encryption posture.
And no compliance-badge theater: agentsnap is an early-stage tool, not a SOC 2-certified platform. What it has instead is an architecture where the sensitive data simply never moves. If your review needs specific answers, ask directly — you'll get them from the person who wrote the code.
Have a security reviewer with questions?
Put them in front of the person who wrote every line. Short call, straight answers, code on screen if you want it.