Your AI opens a terminal and hands the job to Claude Code.
Chat AIs talk about code. Coding agents change it. The coding-agent tool connects the two - and in interactive mode you watch the work happen in a real terminal window.
You ask your AI to fix a failing test. It can explain the fix. It cannot edit your repo, run the suite and iterate until green. Claude Code and OpenCode can.
The coding-agent tool hands the task to one of them, on the machine where the code lives, and brings the result back into your conversation.
Two modes
Pick fire-and-collect or watch-and-steer.
Batch
Interactive
How the result finds its way back
The terminal session is a separate process. Here is how its answer still lands in the right thread.
Your AI calls the tool
With your prompt and interactive mode switched on.
A tracked task is created
The thread waits for that task - not for the terminal.
A terminal window opens
Terminal on macOS, your $TERMINAL or the first emulator found on Linux, a fresh PowerShell console on Windows.
The agent gets its marching orders
The prompt carries one extra line: when done, call complete-task on this instance's MCP server with this task id.
The result lands in your thread
complete-task delivers the output and the conversation resumes.
The line appended to every interactive Claude Code prompt:
[TASK CONTEXT] When the work is complete, call MCP tool "complete-task"
on MCP server "<your-instance>" with taskId="<task-id>"
and response={"output":"<full result text>"} - pass the complete result.New: it works across instances
Thea in the cloud can now ask Hermes on your machine to run a coding session.
- On the remote instance the session runs as a tracked background task, so nothing blocks while you work in the terminal.
- When the session reports through complete-task, the real result is relayed back to the instance that asked.
- On Windows it opens a real console window. The command is written to a PowerShell script first, so prompts with quotes, braces or newlines survive intact.
- A session launched by the local machine may report on any task its own instance owns.
- No live channel to report back on? Interactive mode falls back to batch, so the caller still gets an answer.
Try it
Then ask in chat - or call it straight from the terminal:
# batch (default): runs headless, returns the output
vibe coding-agent "Fix the failing login test and run the suite"
# interactive: opens a terminal window, reports back when done
vibe coding-agent "Refactor the billing module" --interactiveMode=trueGive your AI hands.
Self-host next-vibe, sign in to Claude Code once, and let your AI ship code changes on its own machine.