Enforce your AI policy at runtime. Prove it later.

SAFi is an open source governance engine for AI agents. It evaluates every turn against your charter and policies before an answer is returned or a tool runs, and writes an audit record anyone can recompute.

Try the live demo

Browse the source

50+ GitHub stars, AGPL-3.0, self-hosted

100% deterministic enforcement, no model calls in the gate

Every turn recorded for audit and recomputable

5 steps every decision follows one governed process

A governed turn in SAFi: a user asks whether to buy Apple stock, the Fiduciary agent answers, the Audit Hub shows the turn, and the value-by-value ledger appears with a confidence figure beside each value — A governed turn: the agent's answer, the Audit Hub, and the value-by-value ledger.

Every decision follows one process

Set the values Your charter and policies compile into the standard the agent is held to, with defined weighting.

Draft the answer The model writes its response, but it has no power to act on its own.

Approve or block A fixed rule evaluates every turn and tool call before anything runs. No model, no judgement, no prompt.

Score against the values An independent pass checks the draft against the values. The judge is never the defendant.

Integrate and record Weights combine into one decision, the outcome is logged, and drift is measured over time.

The gap SAFi closes

Your organization is deploying AI agents, and legal, compliance, and technology leaders are asking hard questions: How do we enforce our AI policies at runtime? How do we know an agent followed them? How do we detect drift from the values and policies we approved? How do we prevent unauthorized tool calls? How do we show an examiner what was produced, which policies applied, and why an action was allowed?

The answer to why a response was allowed has to be more than “a content filter fired.” It has to be “the agent was governed by this policy, with these values, and here is the record.”

What SAFi does instead

SAFi governs AI agents with instruments your organization already understands: a charter, policies, supervision, and a record. An agent can operate under your Organizational Charter, a specific business-unit policy, or both. Charter values are compiled into the value set used to evaluate the agent, with a defined weighting in every evaluation.

Enforcement is deterministic. Every block, approval, redirect, retry and threshold is a fixed rule evaluated in code. The gatekeeper never calls a model.
Reasoning and review are separate. The model may draft a response, but an independent pass scores it against your values before it is allowed. The judge is never the defendant.
Every governed turn is recorded. Anyone holding the audit record can recompute the outcome. That is the determinism claim the product is allowed to make.

Built for regulated industries

SAFi’s architecture was designed for auditability first, which is why it maps onto the strictest AI and record-keeping regimes: hash-chained, tamper-evident records for SEC / FINRA; the EU AI Act transparency tier; a provider allow-list and fail-closed model calls for HIPAA; access export for GDPR. These are platform capabilities designed to support a compliance program, not substitute for one.

Deploy SAFi in your stack.

Self-host it yourself, or let our managed service run it for you.

Try the live demo

Managed service

Prefer to run it yourself? Browse the source.