An agent's competence should never be its permission.
Karuna is a permission layer for autonomous AI agents. The agent proposes an action. Karuna decides whether it happens. Your team decides what Karuna allows.
Illustrative example of Karuna evaluating proposed agent actions.
The problem
AI agents are starting to run code, move money, and act on the internet with little supervision. Most of them hold their own credentials and police themselves.
How it works
Isolate. Decide. Record.
- 01
Isolate
Agents run in a sandbox with no network access, no keys, and no memory between tasks.
- 02
Decide
Every action goes through a small, deterministic policy engine and a gatekeeper that holds all credentials.
- 03
Record
Every decision is signed and written to a tamper-evident log.
Compatibility
Model-agnostic.
Works with Claude, GPT, Gemini, and open-weight models, and plugs into LangChain, CrewAI, and AutoGPT.
Models
- Claude
- GPT
- Gemini
- Open-weight models
Frameworks
- LangChain
- CrewAI
- AutoGPT
Governance
Policies live in a public ledger called Dharma.
A core set of boundaries can never be removed. Everything else is governed by people. Tightening a rule is fast. Loosening one is slow and public.
- Core boundaries
- Can never be removed
- Tighten a rule
- Fast
- Loosen a rule
- Slow and public
Guarantees
Our goal is a boundary agents can't talk, trick, or compute their way around, built on a core small enough to formally verify.
We publish every guarantee alongside its assumptions.
Pricing
Open core.
Organizations
Coming soon
Shared policies, human approvals, and audit trails across every agent.
Join the waitlist