Voice-drive a coding agent
ORBIS can launch a terminal coding agent — protoCLI, OpenCode, Claude Code, or Codex — and drive it by voice: "ask proto to add a test to the payments module." The agent reads, edits, and runs code in a directory you choose; ORBIS narrates what it's doing and speaks the result. Under the hood it's the Agent Client Protocol (ACP), wired in as a delegate type. For the schema, see Agent reference → Delegates.
Before you start
- Install the agent and make sure it's on your
PATH. Sign it into its own model — ORBIS drives it, it brings its own brain. - Pick the directory it owns — the repo or folder it should read and edit.
Launch commands (each speaks ACP over stdio):
| Agent | Command | Args |
|---|---|---|
| protoCLI | proto | --acp |
| OpenCode | opencode | acp |
| Claude Code | npx | @zed-industries/claude-code-acp |
| Codex | codex-acp | (none — install zed-industries/codex-acp) |
Add it
In Settings → Agent → Delegates → Add delegate, choose ACP coding agent:
- Name — lowercase, no spaces (e.g.
proto). This is what you say and what the model uses indelegate_to. - Description — be specific about which code it owns; the model uses it to choose. "My coding agent for the ORBIS repo" beats "a coding agent".
- Command / Args — from the table above.
- Workdir — the directory the agent is responsible for.
Or in config/delegates.yaml:
delegates:
- name: proto
description: My coding agent for the ORBIS repo — read, edit, run code there.
type: acp
command: proto
args: ["--acp"]
workdir: ~/dev/ORBISThe workdir is the unit of identity: add the same agent against two repos and you get two scoped delegates — "ask proto in the ORBIS repo…" vs "…in the marketing repo."
Use it
Just talk: "ask proto what's failing in the build," or "have proto add a docstring to the license module." ORBIS launches the agent in its workdir, hands over the task, narrates its tool calls as it works ("Editing app.py", "Running pytest"), and speaks the outcome. Follow-ups continue the same session, so "now also update the changelog" keeps the context.
Good to know
- It edits real files. The agent works directly in its
workdir— changes land on disk. Point it at a repo under version control so you can review and revert. - Permissions. When the agent asks to run a tool, ORBIS currently allows it so the task can proceed. A confirm-by-voice policy for writes and shell commands is on the roadmap.
- It must be installed. ORBIS launches what's on your
PATH; it doesn't bundle the agents. If a launch fails, confirm the command runs in your terminal first (e.g.proto --acp).