Ognef

AI agents act across tools and services, but their authorization, limits, decisions, and evidence are scattered. Ognef brings them into one human-readable control plane. A person issues a scoped, expiring, revocable mandate; an objective becomes a typed proposal; and a deterministic, default-deny policy returns allow, confirm, or deny with one explicit reason. Control Room, Command, Trust Atlas, Mandates, and Session views expose who authorized what, which limits still apply, what evidence supports the result, and how to stop or revoke it—without ever treating a model response, voice command, or payment request as authority.

We built Ognef as a polyglot, zero-trust control plane. A Next.js 16/React 19 control room renders mandates, typed proposals, decisions, sessions, and Trust Atlas. A Rust/Axum service owns the deterministic, default-deny policy engine and returns allow, confirm, or deny with explicit reason codes. A Fastify/TypeScript agent runtime converts natural-language objectives into allowlisted, schema-validated proposals—it cannot execute or authorize them—and a FastAPI voice lane uses the same contract. Zod and Protobuf define the canonical data contracts. For partner technologies, we implemented World IDKit/MiniKit proof acquisition and server-side verification logic, a fail-closed 0G-compatible evaluation gateway that hashes minimized inputs and outputs, and a Bun/TypeScript Hedera service using @hashgraph/sdk and Hedera Agent Kit. The Hedera adapters cover HCS hash anchoring, HTS entitlement tokens, capped HBAR transfers, and independent Mirror Node checks, with no Solidity. The hackiest choice is our failure semantics: unavailable identity, compute, or chain services become explicit withheld, unavailable, unresolved, or unanchored states. Ognef never turns a prompt, voice command, credential, or payment request into authority, never puts raw evidence on-chain, and never hides an outage behind a fallback result.