How it works
Register your agents, pick the policies, and every step gets checked.
The engine sits inside each agent and decides every step locally, so your traffic doesn't run through a gateway. The platform is where you manage the policies and review what your agents did.
The workflow
From your first agent to your first block.
One agent, start to finish: the same five steps whether you run one agent on a laptop or fifty in production.
What runs where
The decision happens in your process. The overview lives in the platform.
Nothing sits between your agent and its tools. The platform never sees a step before it runs; it hears about the task afterwards.
- The SDK, in your agent
- One decorator per tool, or an integration for your framework. It registers the agent, opens and closes tasks, and describes each step to the engine.
- pip install kyvvu · Apache 2.0
- The engine, in the same process
- Evaluates every step against the task's history with no network call on the hot path. Policies are fetched once, cached, and refreshed every few minutes.
- p99 under 0.3 ms · no proxy
- The platform, hosted or yours
- Where policies, assignments, logs and incidents come together. Use platform.kyvvu.com, or run it in your own infrastructure on Enterprise.
- hosted · self-hosted · air-gapped
Where it sits
Where Kyvvu sits next to your other controls.
Identity, guardrails and gateways each do a real job, and you should keep them. None of them decides based on what the agent already did.
| Layer | What it decides | Knows the task so far | Data leaves your environment |
|---|---|---|---|
| Identity & access | Who may access what | No | No |
| Guardrails | Whether one message is safe | No | Depends |
| AI gateway | Routing and content policy per call | Limited | Often, if SaaS |
| Credential brokering | Whether to issue a credential | Per task | Credentials only |
| Kyvvu | Whether this action may run now | Yes, the full path | No, it runs in-process |
Your own policies
A policy is a file in your repository.
Start from the open library or write your own. A manifest is a short YAML file, reviewed and versioned like the rest of your code. Connect the repository and assign it from the dashboard or the CLI.
- Standard or custom
- OWASP agentic, AI Act, GDPR, DORA and sector packs are ready to assign. Your own sit next to them.
- Per risk class
- A policy can apply only to high-risk agents, or to every agent.
- No redeploy
- Assignment changes reach running agents within the policy refresh window, five minutes by default.
name: Support agentsdescription: Our own rules, on top of the baselineversion: 1.0.0policies: - name: Block external send after sensitive data read rule_type: tainted_path_block params: taint_kind: step.resource taint_verb: GET taint_property_filter: data.classification: pii target_kinds: [step.message, step.resource] target_verb: POST severity: critical # critical blocks; high, medium and low warn enforcement_point: step_execution - name: Forbid credential read followed by external send rule_type: tainted_path_block params: taint_kind: step.credential taint_verb: GET target_kinds: [step.message] severity: critical enforcement_point: step_executionRollout
Start in shadow mode and turn on blocking when you're ready.
Turn enforcement on one agent at a time, with the same policy file all the way through.
Shadow
Watch what agents do
Kyvvu records every verdict and blocks nothing. You see which rules would fire, on real traffic.
step.message POST smtp/external, after a PII read
ALLOWruns · logged as a would-be block
Suggest
Flag before you block
Risky steps still run but get a WARN, so owners can tune policies without stopping work.
step.message POST smtp/external, after a PII read
WARNruns · flagged to the owner
Enforce
Stop what shouldn't run
Dangerous sequences are blocked before they execute, and every verdict lands in your SIEM.
step.message POST smtp/external, after a PII read
BLOCKstopped before it ran
Questions
How it fits in.
4 questions · ask another
01Is Kyvvu a SaaS?
The platform, where you manage policies and read traces, is a hosted service at platform.kyvvu.com. The engine that makes every decision is not: it's a library inside your agent's own process.
02What reaches the platform?
Policies and assignments go out to your agents. Step traces and incidents come back, so you can see them in the dashboard. If you'd rather keep traces on your own machines, one setting does that.
03Do I have to change my agent?
A decorator on each tool, or the integration for LangChain, LangGraph, CrewAI or Claude Code. The agent's logic stays the same.
04Can I try it without blocking anything?
Yes. Only critical policies block; high, medium and low ones warn, so the step runs and an incident is recorded. Start a new policy as a warning and make it critical once you trust it.
Get started
Your first agent under enforcement, today.
For security and AI teams
See it on your own agents
We run Kyvvu next to one of your agents, write policies from your own rules, and walk your team through the dashboard.
For developers
Five minutes, no account
kyvvu try runs a demo agent against the baseline on your machine. Sign up when you want the dashboard.