A privacy gateway for AI

Your AI sees the work.
Not the names.

Elided sits between your apps and AI providers. Names, companies, tax codes and bank details are replaced with placeholders before anything leaves your network, and put back in the answer.

Request early access See how it works

Self-hosted. Works with the Anthropic and OpenAI APIs. Built for Italian and EU documents.

Your team writes

The AI provider sees

Tra Alfa Logistica S.r.l.[COMPANY_1], P.IVA 01234567897[VAT_ID_1], in persona del legale rappresentante Mario Rossi[PERSON_1] (C.F. RSSMRA80A01F205X[TAX_ID_1]), con sede in Via Garibaldi 12[ADDRESS_1], Milano, e il Cliente. Il corrispettivo annuo di EUR 120.000 è versato sul conto IT60 X054 2811 1010 0000 0123 456[IBAN_1] entro 30 giorni.

6 values replaced. Dates, amounts and roles stay readable.

One setting. Nothing else changes.

Point your app, your assistant or your coding agent at Elided instead of the AI provider. It speaks the same API, streams the same way, and hands your data back intact.

Swap one address

Change the API base URL to Elided, e.g. ANTHROPIC_BASE_URL. Your keys and models stay the same.

Sensitive values are replaced

People, companies, addresses, tax and VAT numbers, IBANs, emails and phone numbers become consistent placeholders. The model still sees who is who, just not who they are.

The answer comes back whole

Placeholders are swapped back as the response streams in, including inside tool calls, so quotes, drafts and letters come back with the real names in place.

Anthropic APIOpenAI APIClaude CodeCodexOpenRouterAny compatible API

Built for documents that matter.

Contracts, HR files, claims, case notes: the documents teams most want AI help with, and are least comfortable sending out.

Your network, your hardware

Elided runs on your own servers. Detection happens locally, and the mapping between real values and placeholders is encrypted and never leaves.

Several detectors, then a second look

Independent detection engines work in parallel, and a local AI model reviews what they leave behind.

Learns your documents

The first contract from a template takes a moment. The next ones are recognised by their structure and handled almost instantly.

Fails closed

If a detector you rely on is unavailable, the request is held back. Elided never quietly sends text it couldn't check.

Italian and EU documents

Codice fiscale, partita IVA, IBAN, REA and PEC are recognised with their check digits, alongside names, companies and addresses in Italian and English.

Readable where it counts

Roles, dates, amounts, durations and courts stay in clear, so the model can still read the contract and work out a deadline.

See exactly what left the building.

Every document and every request can be opened side by side: what your team sent, and what the AI provider actually received.

Before and after, coloured by type, with the reason each value was hidden.
Live monitoring of traffic, speed and each detector's health.
Your rules: words to always hide or never hide, and checking that can be relaxed for conversations that don't need it, always logged.
Right to erasure: forget a person or company and every spelling of it.

The Elided audit view: a contract on the left with names and identifiers highlighted, and the same contract on the right with placeholders

Choose models on quality and price.

When client names and identifiers stay on your side, the question of which AI provider can see them gets much smaller. Teams can pick the model that does the job best for the money, and keep that choice open as models change.

Same answers, real names: placeholders are consistent across a whole document and conversation, so the model can still reason about who did what.
Works with agents: streaming replies and tool calls are restored as they arrive.
Already have documents? Send them through once and every later request that includes them is masked in a single lookup.

Questions

Is this anonymisation?

No, and we won't call it that. Elided pseudonymises: values are replaced, and a mapping kept on your side puts them back. Under the GDPR that is still personal data for you. What changes is what the AI provider receives. You remain the controller, and your provider agreements still apply.

Can it miss something?

Detection is statistical, so yes, it can. That's why it combines several detectors with a second check, lets you add words to always hide, and shows you exactly what was sent. We test it continuously against large sets of labelled documents, and we'll help you measure it on yours before you rely on it.

Will it slow things down?

Replies stream straight through. Text Elided has already seen, including documents processed at upload, is masked almost instantly; new text adds a short pause before the first word of the reply.

Which AI models does it work with?

Anything that speaks the Anthropic or OpenAI API: the providers themselves, model routers, or a model running in your own cloud. Coding agents such as Claude Code and Codex work too.

Where does it run?

On your infrastructure. The detection models run locally, and the original values never leave your network.

Will it be open source?

That's the plan. If you'd like to try it before then, get in touch.