Antaeus

Antaeus is a semantic decision engine that turns human-written policy into versioned artifacts an application can evaluate and inspect. It is meant for bounded questions that need interpretation of language, intent, context or exceptions, while deterministic rules, authentication, authorization and side effects stay in the application's own code. It is aimed at engineers who run meaning-heavy decisions in production and need to know which policy decided a case, what changed, and what happens when evaluation fails.

A policy is written as constrained YAML or JSON against a versioned schema, and Antaeus derives a canonical digest for that exact policy. Named cases bind inputs and expected outcomes to the policy version so wording changes show up as test differences. An evaluator route is chosen separately from the policy meaning; the local quickstart uses a synthetic fixture with no credentials or network calls, and an experimental OpenAI adapter evaluates real conditions with the user's own key.

Every evaluation returns one of four explicit outcomes: allow, review, deny or failure, so an outage is never treated as a denial or a permission. The decision carries the policy, rule results, reason codes and the supporting proof behind it. The hosted API accepts an inline policy or a reference to a published immutable version, uses idempotency keys with a 24-hour replay window, and supports shadow mode. An MCP server lets AI assistants check proposed actions against an approved policy.

The core is open source under the Apache-2.0 license, with Go packages, a command-line workflow, schemas and examples; release v0.6.0 installs with Homebrew, go install or a container image. The hosted beta adds API keys, managed policies, a dashboard and prepaid usage, but runs on shared capacity with no production SLA, and the semantic evaluator is experimental. Request metadata and decisions are kept for 90 days; raw inputs and inline policy text are not stored in that log.

Use cases

• Checking a product listing against selected marketplace policy rules
• Screening user submissions for threats before they reach an application decision
• Turning terms of use into an enforceable content policy
• Testing a policy wording change against named cases before it ships
• Letting an AI assistant verify proposed actions against an approved policy through MCP