Sentinel
Sentinel documentation
What Sentinel does, and where to start.
Sentinel is an open-source platform that runs AI agents on Kubernetes. It gives each agent an identity that the platform can verify. It controls which services the agent can reach. It records what the agent did.
Sentinel does not replace your agent framework. Your agent runs on Sentinel without a change to its code, if the agent uses the A2A protocol.
Where to start#
| Your goal | Start here | Time |
|---|---|---|
| See the model calls and tool calls that your agent makes | Quickstart on a laptop | 5 minutes |
| Understand the token counts and the cost of a session | Read the numbers | — |
| Run the platform and deploy an agent | Quickstart on Kubernetes | 20 minutes |
| Put your own agent on the platform | Bring your own agent | — |
| Review the security model | Security | — |
| Find a command or a field | Reference | — |
If you do not know which quickstart to use, read Get started.
The three parts of Sentinel#
Cortex is the data plane. It is a proxy that runs next to each agent. All traffic between the agent and an external service goes through this proxy. The proxy applies the identity checks and the access rules. See Cortex.
Services are the parts that an agent uses to do work. They include tools, skills, sandboxes and memory. See Concepts.
Tools for operators include the web console, the sentinel command-line interface and the
Kubernetes operator. See Install and operate.
What Sentinel gives you#
- An identity for each agent. SPIRE gives each agent a cryptographic identity. The platform identifies the agent before it permits an action.
- Access by delegation. An agent that calls a tool for you receives a token. The token is valid only for that tool. The token gives the agent no more permissions than you have.
- Deployment without a code change. Give Sentinel a container image or a Git repository. Sentinel builds the agent, starts it and adds it to the platform.
- A record of each action. Sentinel collects traces, network data and token costs for each agent.
Maturity#
Some features are ready for production use. Others are experiments. Read Concepts for the status of each feature before you depend on it.