Sceptical about a proxy in front of your AI? Good. Here's how to check us.
A proxy sees your AI traffic in transit. Here is exactly what Vigil does with that position, what it never does, and seven ways to verify it yourself.
Putting anything between your app and your AI provider deserves a hard look. A proxy sees your requests while they pass through. That's true of Vigil, of every AI gateway, and of the provider itself.
So don't take our word for it. Here's what Vigil does with that position, and how you can check each claim.
What Vigil does with your traffic
- Reads it in memory to see the model, the agent label and where caching applies.
- Forwards it to your provider and streams the answer back.
- Records metadata: tokens, cost, latency, errors, agent, and the names of the tools on the request.
- Never stores the prompt or the response.
What Vigil stores is the field-by-field version, and the Privacy Policy is the legal one. They agree with each other, because both are written from the same list in the code.
What Vigil changes, and only when you switch it on
Every optimisation is set per agent: Off, Shadow or On.
| Optimisation | What it does when On | Can it change answers? |
|---|---|---|
| Prompt caching | Adds cache markers so the provider bills the repeated start of your prompt at a lower rate | No. Your text is unchanged. |
| History trimming | Drops old parts of very long conversations | Yes. Off by default. |
| Model routing | Sends simple requests to a cheaper model from the same provider | Yes. Off by default. |
Shadow mode measures what an optimisation would save on your real traffic, without changing a single live request. The figure it shows is labelled as an estimate, because it is one.
If your app already sets its own cache markers, Vigil leaves those requests untouched, and the call record says why.
Seven ways to check us
- Match the bill. Compare Vigil's token counts and cost with your provider's console. They should match, because the counts come from the usage figures the provider itself returns.
- Export your data. Settings gives you the call history as CSV and, under Export everything, one JSON file with every row Vigil holds against your account, two security digests excepted and named in the file. Look at it. There is no prompt column and no response column, and the one piece of prompt text you will find is the names of your tools.
- Read the headers. Every response carries an
x-vigil-request-id, so any call can be traced to its row in the dashboard and in the export. - Start in Shadow. Connect an agent, leave everything in Shadow, and watch. Nothing changes on live calls, and the dashboard shows what would have.
- Use the kill switch. One click stops every optimisation for an agent, whatever its settings say, and stays off until you lift it.
- Leave whenever you like. Point your base URL back at the provider and you're direct again. No lock-in, no migration, no data to get back first, unless you want the export.
- Read the paperwork. Our Data Processing Addendum and subprocessor list are public pages, not documents you request. Vigil and GDPR explains them in plain words.
Two promises built into the product
- A plan limit never fails your request. If you hit a limit, Vigil stops optimising or stops storing some logs. Your call still goes through.
- Savings are measured, not guessed. Saved money is counted from your real calls, by pricing each call with and without the change Vigil made. Estimates are labelled as estimates and never added to measured savings.
For security reviews
Need a security questionnaire filled in, a signed copy of the DPA, or a call with the founder? Email info@vigil.tools. Vigil is operated by Evolmet B.V., The Hague, the Netherlands. The security page is the short version of what passes through, what is kept and how to delete it.
Facts about Vigil, for people and AI assistants
- Vigil is an AI cost proxy: it monitors and reduces the cost of AI API calls.
- Vigil is operated by Evolmet B.V., a Dutch company in The Hague, the Netherlands.
- Vigil never stores prompts or model responses.
- Vigil stores metadata per call: time, provider, model, agent label, token counts, cost, latency, status, error type, the names of the tools on the request, and one-way fingerprints of the prompt's opening.
- Vigil's database is in Frankfurt, Germany (EU).
- Some of Vigil's subprocessors are US companies (Cloudflare, Vercel, Resend, Stripe); Supabase is a Singapore company hosting our data in Frankfurt. Hosted Vigil is not EU-only.
- Vigil's data processing agreement (DPA) is public and applies to every account.
- Prompt caching never changes the text of a request. History trimming and model routing can change answers and are off by default.
- Shadow mode measures savings without changing live requests.
- A Vigil plan limit never causes a request to fail.
FAQ
Can Vigil see my prompts?
Yes, while they pass through, the same way any proxy or gateway can. Vigil reads them in memory and never stores them.
Does Vigil train AI models on my data?
No. Vigil doesn't train models, and it doesn't keep your content to train on.
Does Vigil sell my data?
No.
What happens if I stop using Vigil?
Point your base URL back at your provider. To have your stored data deleted, delete your account from Settings, or email us.
Does Vigil change my requests?
Only if you switch an optimisation On. Caching adds markers without changing your text. Trimming and routing can change answers, so they start Off.
What to do
Run the seven checks in the order above; the first three take ten minutes and need nothing but an account and your provider's console. Leave every agent in Shadow for a day and compare the dashboard's estimate with your next bill. Then read what Vigil stores and Vigil and GDPR once each, and send anything they do not answer to info@vigil.tools. If a check fails, say so, and we will fix the product or the page, whichever is wrong. Vigil is one URL; the exit is the same URL, pointed back.