Launched this week
Astra
Make AI agents that never see your data
136 followers
Make AI agents that never see your data
136 followers
Your AI agent shouldn't see raw sensitive data to do its job. Most of the time it doesn't need to. Astra tokenizes PHI, PCI, and PII before it reaches your agent. It reasons on safe tokens, acts on real values at execution the raw data never touches the model context. Two lines of code. Works with any agent framework.
Astra
Product Hunt
Astra
@curiouskitty Well we have Three moments, in order:
1. The incident. Someone opens LangSmith and finds a real SSN in a trace. Plain text. Logged. Queryable. That's the Monday morning call.
2. The compliance review. Auditor asks "show me what your LLM receives." Team pulls a sample prompt. Real values sitting right there. Audit fails.
3. The contract requirement. Enterprise client says we'll sign, but we need proof the LLM never sees our data. Current redaction layer can't produce that proof.
What makes them switch: redaction breaks execution. Agent tries to send an email to [REDACTED], fill a form with [REMOVED] pipeline breaks. They're choosing between security and functionality.
Astra removes that tradeoff. Tokens carry enough semantic meaning to reason on, executor resolves at the last mile. Agent works fully, never touches raw values.
That's not better redaction , it's a different architecture entirely.
@obed_mpaka1 the tokenization-before-prompt approach is interesting — what happens when the agent's reasoning output references a token and you need to log or audit that decision? Does the audit trail show the real value or does it stay tokenized end-to-end?
Astra
@jimmypk so the audit log stores tokens, not real values.
[CVT:NAME:A1B2] filled first_name at hospital.com at 14:13:22. Authorized. Uses remaining: 0.
The real value lives in one place, the vault. The reveal log records that a reveal happened, not what was revealed. Those two things are deliberately separate. If they needs to know which patient was affected, they run the token through the executor with proper authorization. The audit trail points to the token. The vault holds the value. Nobody hands them a document full of PHI.
Agent reasoning log : tokens only
Audit trail : tokens, action, timestamp, who triggered the reveal
Vault : real values, access-controlled separately
Reveal log : proof a reveal happened, without storing what was revealed
You can hand that audit log to a regulator as-is. It doesn't become a PHI liability the moment you open it.
That's the point.
how does astra treat this raw sensitive data (which it processes)?
Astra
@shobana_devarajan Love your question , the truth is Astra never stores it.
Raw values go into a vault at interception. The agent gets tokens. When execution happens, the vault resolves the token to the real value in memory, performs the action, and that's it nothing persists. The raw value never sits in a log, a prompt, or an audit trail.
This is a critical problem in financial services. When we build project finance models for renewable energy deals, the data flowing through them — tariff structures, counterparty financials, tax equity terms — is highly sensitive. The idea that AI agents can reason on tokenized data without ever seeing the raw values is exactly what regulated industries need to adopt AI safely. I publish financial model templates on Eloquens (https://www.eloquens.com/channel/samir-asadov-cfa) and data privacy in model distribution is always a concern. Would love to see this applied to financial modeling workflows.
Yo @obed_mpaka1 @Astra quick question.
I’m a researcher for H1Gallery newsletters (you can google us).
We’re featuring Astra in the April 17 H1Gallery issue. H1Gallery highlights excellent homepage headlines, and “AI agents handle sensitive data every day. They shouldn't see it.” really stood out to us. Its clear and compelling.
I wanted to reach out to see if you’d be open to sharing a quick comment on the copywriting strategy behind that headline and the broader messaging. We’d love to include a short note from your team on how you approached it.
Totally optional, of course . The feature is happening regardless either way. Our readers love to hear from the creators behind the headlines tho. And sorry for the late notice!
Thank you so much. Love the product!
Astra
hey @michael_henderson550 I’d love that. Let’s get in touch on LinkedIn. The copywriting strategy was developed by me along with some people on the marketing side at AWS.