For AI and engineering teams
Get your agents through security review.
The demo works. Getting it into production means convincing security, legal and risk that it won't do anything stupid. Kyvvu gives you a control you can point to: one decorator per tool, rules in your repo, checks in your process.
The problem
The agent is ready. The sign-off isn't.
- 01
Stuck in review
Security asks what stops the agent from leaking data or deleting things. 'The prompt says not to' doesn't pass.
go-live blockers
- 02
A control per framework
LangGraph here, CrewAI there, Claude Code for the developers. Building safety into each one separately doesn't scale.
framework sprawl
- 03
No latency to spare
A security proxy in the middle of every tool call adds hops, failure modes and milliseconds you don't have.
latency budget
In practice
The checks you'd otherwise hand-code.
A task that burns through the budget
A research agent goes deep on a question and keeps calling the model. The task has a token budget of 100,000.
Illustrative scenarios. Each one is decided by a single policy from the open policy library, as the engine evaluates it.
How Kyvvu helps
It runs where your agents run.
- 01
One policy, every agent
The same rules for plain Python, LangChain, LangGraph, CrewAI and Claude Code agents, in the cloud, on-prem or air-gapped.
Python/LangChain/LangGraph/CrewAI/Claude Code/any language
- 02
Runs inside your environment
The engine sits in the agent's own process: so there's no proxy or extra network hop, and a check takes under a millisecond. Prompts and data stay with you.
in-process/no proxy
- 03
Rules people can read
Policies are short YAML files in your own git, reviewed and versioned like code. Every verdict in the trace names the rule that decided.
policies as code
- 04
Nothing breaks on day one
Only critical policies block; high, medium and low ones warn. Start a new rule as a warning and make it critical once you trust it.
shadow/suggest/enforce
For developers
Install the SDK, add a decorator, keep the policies in your repo.
The SDK wraps your tools. Every call is checked in-process before it runs, against the policy file you keep next to your code.
- In your process
- No proxy and no extra hop. Prompts and data stay on your machine.
- Under a millisecond
- Policies are fetched once and cached. Every verdict is computed locally.
from kyvvu import Kyvvu, ActionKind, Verbkv = Kyvvu(api_url="https://platform.kyvvu.com", api_key="KvKey-...")@kv.step(ActionKind.step_resource, Verb.GET)def lookup_customer(customer_id: str) -> dict: return crm.get(customer_id)@kv.step(ActionKind.step_message, Verb.POST)def send_email(to: str, body: str) -> str: # Blocked by policy if the task read PII earlier on return smtp.send(to, body)Questions
Common questions.
01Which frameworks are supported?
Plain Python, LangChain, LangGraph, CrewAI and Claude Code, and agents in any language through the local engine. See integrations.
02Do we have to change our agent?
A decorator on each tool, or the integration for your framework. The agent's logic stays the same.
03Can we start without blocking anything?
Yes. Start policies as warnings, see what they would have done on real traffic, and make them critical once you trust them.
Get started
Every agent has to ASK before it acts.
Start with one agent
See it on your own agents
A 30-minute demo with someone who builds Kyvvu: your agent, your rules, and the policy that decides each step.
For developers
Your first block in five minutes
kyvvu try runs a demo agent against the baseline on your machine. No account needed.
Want to see other teams and sectors? All use cases.