The guarantee
Every guard rule is backed by 287 automated tests that prove it can't be bypassed, injected, or broken.
102 tests
Murphy's Law: every edge case, every misuse, every attack vector
3 attack scenarios
Pocket OS deletion, OpenAI egg purchase, prompt injection — all blocked
10 profiles verified
Every seed profile tested: schemas load, guards evaluate, states reachable
Guards are code, not prompts. They run before every action. The LLM never sees them.
No prompt injection. No hallucination. No "I decided to do it on my own."
Questions
How do guardrails actually work?
Guards are JSON AST rules evaluated by a deterministic engine — no eval(), no code execution, no prompt dependency. They run BEFORE every FSM transition. If the guard evaluates false, the transition is blocked. The LLM never sees the guard rules. Prompt injection cannot disable them.
Can I really ship to production for free?
Yes. Unlimited guard rules, unlimited FSM schemas, and all 37 MCP tools (14 work offline). Your guards run in your process, evaluate every tool call, and block dangerous actions. Deploy to production today. You only pay when your team needs central control.
Why would I upgrade to Pro?
On Free, your rules live in code. Changing a rule means changing code and redeploying. On Pro, you manage rules from a dashboard or API. The SDK fetches the latest rules at startup. Update a rule at 2am, every agent picks it up automatically, no deploy needed. Pro also gives you an audit dashboard showing what every agent did and what was blocked.
How do managed rules work?
Your SDK fetches the latest rule bundle from Savants on startup and polls for changes every 30 seconds. Guards evaluate locally in your process with zero network latency per check. Events are batched and reported to the dashboard asynchronously. No per-request API calls, no single point of failure, no surprise bills based on traffic.
Does any data leave my machine?
On Free, nothing leaves your machine. On Pro/Team, the SDK reports batched event summaries (action name, result, timestamp) to the dashboard. Your source code, tool arguments, and user data never leave your process. Enterprise customers can self-host the entire policy server.
What if the guardrail itself has a bug?
Every guard is tested by our Murphy's Law suite — 102 tests covering broken schemas, bypass attempts, code injection, prototype pollution, and real-world attack scenarios. The test suite runs on every release. You can also test your own guards with savants guard test before deploying.
How is this different from putting rules in the LLM prompt?
Prompt rules are suggestions — the LLM can ignore, hallucinate around, or be tricked past them via injection. Savants guards are code that runs OUTSIDE the LLM, BEFORE the action executes. The LLM has no way to know the guard exists, let alone bypass it.
Do you store my source code?
No. Savants only stores metadata — function signatures, call graphs, file paths, and line numbers. Guard rules and FSM schemas are stored in your org's isolated Durable Object. Your source code never leaves your machine.
Can I publish my own profiles to the registry?
Yes. Pro and above can publish profiles with savants profile publish. Your schemas and guard rules become available for anyone to install. You keep ownership and can unpublish anytime.