Security and data. What we see, and what we don't.
Wireproof answers questions about third-party APIs and checks integration code. Here is exactly what reaches us in each way you can run it, what we keep, and what we don't offer yet.
Two ways to run it
On your machine
claude mcp add wireproof -e WIREPROOF_API_KEY=… \ -- npx -y wireproof mcp
The facts and rules download once with your key. Every tool call after that, including checking code, is answered on your machine. Your code never reaches us. The wireproof check, go-live and replay commands work the same way.
Our MCP endpoint
https://wireproof.dev/api/mcp
Tool calls are sent to us and answered there. Fine for personal and open-source work; for a company repo, use the local mode or check with your security team first.
What reaches us
| When | What we receive |
|---|---|
| Local mode, CLI | Your key and the CLI version, once per run, to download the facts. Nothing about your code. |
| Hosted: lookups | The API name, topic, version or search words your agent asks about. |
Hosted: check_code, go_live_checklist | The code your agent passes in, usually one file. It's checked in memory to build the answer, then dropped: not stored, not logged. |
wireproof replay | Nothing. It sends recorded webhooks only to the URL you give it, normally your own local handler. |
| The request-access form | Your name, work email, company, the APIs you use and your note. It becomes an issue in our private GitHub repository so we can reply. |
What we keep
- Usage logs: one line per hosted tool call (the tool's name, your key's name, the time), one line per data download (your key's name, the CLI version, the time), and rate-limit events. Never the arguments, never code.
- The logs are our host's (Vercel's) function logs. We don't copy them anywhere else.
- Keys: random 256-bit values. We store only a SHA-256 hash, so a key can't be read back from our side. Each key has its own rate limit and can be revoked at any time.
What Wireproof is made of
- The facts come from the providers' public documentation and from tests against our own sandbox accounts. No customer data goes into them.
- Recorded webhooks for
replayare deliveries to our own sandbox accounts, with only the provider's headers kept and our account ids replaced. - No AI model runs on our side. The server looks answers up in a fixed dataset, and nothing you send is used to train anything.
Where it runs
On Vercel, with the functions in the US East region (Washington, D.C.). The source code is private.
Not offered yet
So you can decide quickly: there's no SOC 2 report, single sign-on, data processing agreement, customer-facing audit log, or self-hosted server yet. The local mode covers the most common concern, code leaving the machine. If you need one of the others to try Wireproof, tell us; it moves up the list.