Developer Infrastructure

One call between your agent
and everything it can break.

Hivenora evaluates proposed actions before they execute. You get a structured decision — allow, require approval, or block — with the context to explain it. No LLM dependencies. No config DSL. Just a function call in the agent's execution path.

Read the Quickstart →Request access

Your application
    │
    ▼
  AI Agent  ──  decides to act
    │
    ▼
hivenora.evaluate()  ──────────────────────→  POST /v1/evaluate
    │                                               │
    │   ◄── decision ───────────────────────────────┘
    │
    ├── is_allowed       → proceed
    ├── requires_approval → wait (human reviews in Control Room)
    └── is_blocked       → stop

The API

Evaluate before you act.

Python

from hivenora import HivenoraClient

hivenora = HivenoraClient()
# reads HIVENORA_API_KEY from env

result = hivenora.evaluate(
    action="crm.delete_contacts",
    intent="Clean Q1 test contacts",
    context={
        "records_affected": 4821,
        "environment": "production",
        "reversibility": "low",
    },
)

if result.is_allowed:
    delete_contacts()
elif result.requires_approval:
    wait_for_approval(result.request_id)
else:
    raise ActionBlocked(result.reasons)

TypeScript

import { Hivenora } from "@hivenora/runtime";

const hivenora = new Hivenora();
// reads HIVENORA_API_KEY from env

const result = await hivenora.evaluate({
  action: "crm.delete_contacts",
  intent: "Clean Q1 test contacts",
  context: {
    recordsAffected: 4821,
    environment: "production",
    reversibility: "low",
  },
});

if (result.isAllowed) {
  await deleteContacts();
} else if (result.requiresApproval) {
  await waitForApproval(result.requestId);
} else {
  throw new ActionBlocked(result.reasons);
}

Private Preview. SDK packages are not yet on npm or PyPI. The REST API is live. Request access to the preview builds.

How it works

The evaluation pipeline.

Intent analysis

Hivenora checks whether the action scope matches the stated intent. A "clean a few contacts" intent that triggers 4,821 deletions is flagged as a mismatch.

Blast radius estimation

Computed deterministically from context — records affected, financial amounts, reversibility, environment. No LLM inference. Same input always produces the same classification.

Policy evaluation

Policies defined in the Control Room are evaluated in priority order. First match wins. If no policy matches, the agent's default decision applies.

Trust model

SDK mode — what it is and what it isn't.

In SDK mode, Hivenora is a checkpoint in your agent's execution path. The enforcement boundary is your application code — not Hivenora's servers. If your code can call a tool without going through evaluate(), that path is not protected.

What SDK mode prevents

Agent logic bugs that cause scope creep
Actions that exceed policy thresholds
Runaway actions within compliant code paths
Actions without human review (via approval flow)

What SDK mode does not prevent

Code that bypasses evaluate() and calls tools directly
A compromised agent binary ignoring the SDK
Application code that ignores the decision result

Stronger enforcement architecture (where Hivenora sits as a proxy with no direct tool path for the agent) is planned. See the Security Model for full detail.

Get started

Integration path.

Register and get an API key

Open the Control Room, create an agent, copy the API key. Set HIVENORA_API_KEY in your environment.

Install the SDK

pip install hivenora or npm install @hivenora/runtime (private preview — contact us for access). Or call the REST API directly with any HTTP client. Docs →

Call evaluate() before consequential actions

Identify the high-impact operations in your agent. Add hivenora.evaluate() before each one. Handle the three decision cases: allow, require_approval, block. Docs →