AnyBio

Gate your AI

AI governance for patient-facing AI. Put a policy between the model's draft and the patient, get back allow, hold or block with the record already written, and route held messages to a person. REST API and a TypeScript SDK.

Govern is the communication gate, open to AI that is not ours. Your server makes one call before each message goes out, POST /api/v1/govern/evaluate with an application key, and gets back allow, hold or block with the record already written. The same gate, policies, decision records and review queue run under AnyBio's own agents: every patient-facing message a program sends is evaluated, and what it withholds waits for a person.

For JavaScript and TypeScript, the @anybio/govern SDK (0.2.x on npm) wraps the call, with adapters for the Vercel AI SDK and LangChain:

npm install @anybio/govern

Each application runs a policy per channel in one of three modes: off, observe (evaluated and recorded; the caller is always told allow) or enforce (the answer binds). A new application starts in observe.

Pages

Read this first

  • What the gate is: deterministic term and claim matching against an authored policy, plus a disclosure check. It catches what the policy names, and only that. It is not a semantic classifier.
  • What a decision is: a record with the original content, any rewritten content, the rationale, the detected claims, the policy and version, the audience, and the review state. The record is written whether or not the message goes out.
  • What a hold is: the gate refusing to deliver until a person releases or rejects. A patient never receives substituted text; a draft the policy would rewrite for a patient is held.
  • What we do not say: that a message is compliant, that the gate certifies anything, or any percentage. Say evaluate, allow, hold, block, observe, enforce, record.

The response loop the gate sits inside - programs, rules, the symptom agent, hand-offs - is documented under Run a program.

On this page