Messvate is your private personal agent that sits between you and every AI. It catches names, cards, keys and secrets before they leave your device — then puts them back in the answer. No setup. No servers. No leaks.
One gateway. Every model you already use.
Every day, people paste passwords, customer emails, card numbers and private keys into AI chats. Once it’s sent, it’s gone. Messvate makes sure it never leaves.
of data employees paste into AI tools is confidential
median added latency with on-device detection
entity types detected — from IBANs to internal codenames
of your raw sensitive data stored on our servers
How it works
Scroll to watch a real prompt travel through Messvate.
Use ChatGPT, Claude, Gemini, your IDE or your own app. Messvate quietly listens on your device.
A lightweight on-device model flags names, emails, cards, keys and IDs in milliseconds — offline.
Claude reasons over the already-redacted skeleton to catch context leaks — codenames, deal sizes, internal hosts.
Only placeholders reach the provider. The reply is re-filled from your local vault, so you read the real thing.
Draft a reply to Sarah Lim ([email protected]) confirming the refund to card 4532 7781 0045 2291. Mention that Project Falcon budget stays at S$2.4M.
Hi [PERSON_1], your refund to the card ending [CARD_1] is confirmed. [CONTEXT_1] remains fully funded at [MONEY_1].
Features
Everything a security team wants. Nothing a normal person has to think about.
Change one base URL — or flip on the browser extension. No Docker, no proxies, no YAML. Your existing SDKs keep working.
On-device model for speed and privacy. Claude core for context. Neither sees your raw secrets leave.
Reversible tokens are kept encrypted on your device. Answers come back with the real values restored.
Claude, OpenAI, Gemini, Mistral, Llama, DeepSeek, OpenRouter, Azure, Bedrock — or your own endpoint.
“Never send client names to any model.” Done. Claude turns it into a policy.
A private, on-device activity feed shows what got caught, where it was headed, and why. Share policies with your team in one click.
English, Tiếng Việt, 中文, 日本語, Bahasa and 40+ more — names and addresses included.
Median overhead under 30ms. Streaming stays streaming.
Built around PDPA and GDPR principles: data minimisation by default, with exportable audit trails.
Setup
No. The on-device model redacts first. Claude core only reasons over the already-masked skeleton to find context leaks the local pass might miss.
Anthropic Claude, OpenAI, Google Gemini, Mistral, Meta Llama, DeepSeek, xAI, OpenRouter, Azure OpenAI, AWS Bedrock, Ollama and any OpenAI-compatible endpoint.
Barely. Local detection takes single-digit milliseconds; the contextual pass runs in parallel with streaming. Median overhead is under 30ms.
They stay in an encrypted vault on your device and are used only to rebuild the AI’s answer for you. You can wipe the vault any time.
Not at all. Install, choose your AI, keep chatting. Developers get a drop-in API if they want it.
Join the private beta and make every AI conversation safe by default.