Building on Letronna

Each way in is a process your program starts and talks to on stdin and stdout. Letronna's approvals, rules and hooks apply to all of them, as they do in the terminal.

Pick one

• Your program wants an answer. Use letronna -p. It runs one task in the project folder and prints the result, as text or JSON.
• Your program is an editor. Use letronna acp. The editor shows the turn as it runs and answers the approvals itself.
• Your program is another agent. Use letronna mcp serve. It keeps its own model and loop, and calls Letronna's tools when it needs them.

From a script

letronna -p never stops to ask, so choose a mode that allows what the task needs. For an answer your program can parse, ask for JSON:

letronna -p --output json "how many files are in this folder?"

To get the answer itself in a shape you choose, give it a JSON Schema. The answer is printed alone, or as value with --output json:

letronna -p --json-schema schema.json "list the TODOs in src"

To keep one run going and send it more messages, use --input-format stream-json. Every option is on Scripts and one-shot runs.

From an editor

The editor starts this command and speaks the Agent Client Protocol to it:

letronna acp

Each prompt runs as an ordinary Letronna session, in the same list as the others. A call that needs a yes goes to the editor with the terminal's answers: once, always in this session, or no. Every mode but Auto is offered. See Zed and JetBrains.

From another agent

Add this command to the other agent as an MCP server:

letronna mcp serve

It offers these tools:

Calls are answered one at a time, in order; start a second server to run more at once. A tool that needs a yes is refused unless the server was started with --yes, and your deny and ask rules still refuse with it. --allow-turn adds letronna_turn, which runs a whole Letronna turn on a prompt and returns its answer.

The media tools spend your credit: see Hosted models and credit.