Run your agents

Your favourite agents need no changes and no configuration. Start them as you always have:

opencode
claude

Kleinia recognises a program as an agent when it starts calling models, and from then on:

it gets its own identity and agent:// address, and shows up in agenttop;
its model calls go through the unified model gateway, so it uses the keys and models you set up once, within the quota you allow;
its context is kept by the system, so a new session still knows the project’s conventions and earlier decisions;
every step it takes is recorded in its execution trace;
what it may do follows the tiered permissions.

Run any program as an agent

Any program can be started as a managed agent with a name you choose:

agentctl run build app -- make -j8

It gets an identity and a resource tier, and appears in agentctl ps and agenttop under that name. agentctl kill build stops it.

Call other agents

Apps, services and devices on the system are agents too, and your agent can call them. From a terminal:

ls /agent/local                           # what is there
agentctl schema <agent> <skill>           # what a skill takes and returns
agentctl call <agent> <skill> '{"…": …}'  # call it

See anything, anywhere, all as agent.

Connect an MCP client

Agents that speak MCP can reach every agent on the system through the system’s MCP server, the same way they reach any other MCP server. Add it in your client’s MCP settings; it appears as a set of tools, one for each skill your agent is allowed to call.

Ask for outcomes

Agents work best when you describe the outcome you want rather than the steps, and say what they must not touch:

“Fix the failing login tests and run the full suite; don’t push.”
“Free up 20 GB without touching my Documents folder.”
“Why is the fan so loud?”