> ## Documentation Index
> Fetch the complete documentation index at: https://docs.orchestrahq.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Conductor

> Hand work to an agent in a sentence: Conductor gathers the context, you review it, the agent runs where you said, and the result comes back to the thread

Conductor is the workspace's agent. Ask it anything about your issues, sessions and
Knowledge, or hand it work in a sentence, on the web, in the desktop app, or in the
terminal under `orchestra` (the Conductor view, `ctrl+g` from home). Nothing changes
until you apply it.

## Put an agent on an issue

Say where, with what, and on which model, as you would to a teammate:

```text theme={null}
put an agent on W-77 locally with claude code opus
```

Conductor answers with a card. What you named is what the card opens on, and you can
change any of it before you apply:

* **Where:** this computer, another of your machines, or the cloud. Orchestra Local has no
  cloud, and runs on your own machine with your own login.
* **Repository:** the one the issue's sessions last committed to, or the only one that
  place can use. If several fit, the card asks you to pick.
* **Agent, model and effort:** what that place can really run.
* **Review the plan first:** on by default.
* **Prompt:** the brief the agent starts with, which you can edit. It carries the
  issue, what the workspace has concluded about it, and any work in flight it could
  collide with. It tells the agent to follow the rules in `.orchestra/standards/`.

If something you named cannot run (the machine is offline, the agent is not installed,
the model is not offered), the card says so and what can run instead. **Apply** stays
off, with the reason beside it, until the card can really start.

The first run of a repository and agent on a machine asks you to approve it once, in a
dialog on that machine.

## Choose a permission mode

On a machine, the card has a **Mode** chooser, and each mode says what it does. Left
alone, the run uses the mode set for that repository and agent on the machine. New
repositories start in **Auto** for Claude Code and the Orchestra agent: the agent works
on its own, and a safety check stops risky actions. Other agents list their own modes.
Change a repository's default in the desktop app under **Settings**, **This machine**.

If Auto is not available (an older model, or your organization turned it off), the run
uses **Accept edits** instead and says so on the card and in the session.

**Skip permissions** removes the agent's own safety checks. It sits behind a warning,
it is never a default, and your machine asks you to confirm it for that one run.

Cloud runs stay in their sandbox and have no mode to choose.

## Review the plan

With plan review on, the agent reads the repository and writes a plan first. The plan
appears in the thread and in the session, with **Approve**, **Request changes** (what
you type goes back to the agent, and it plans again) and **Cancel**.

What holds an agent to that depends on the agent:

| Agent | Where | What plan review does |
| - | - | - |
| Claude Code, Orchestra agent | Your machine | Starts in plan mode. Nothing changes until you approve. |
| Codex, OpenCode, Gemini CLI, Cursor | Your machine | Its first turn is asked to write a plan and stop. Nothing enforces that. |
| Any agent | The cloud | Its first turn is asked to write a plan and stop. Nothing enforces that. |

An approved plan continues in the mode you chose.

## Describe new work

Describe it in a sentence and Conductor proposes an issue with **Background**,
**Scope**, **Where** and **Done when**. Apply it, then put an agent on it.

## Ask what is true now

Conductor starts each question already knowing your urgent issues, what is in flight,
what merged and what your team has committed to, all read with your own access.
With a repository connected, it can also read pull requests, CI, releases and code.

It checks what people wrote against what actually shipped. If an issue says In
Progress but its pull request merged, Conductor tells you. If it cannot check
something, it says how old the claim is.

Ask what to do next and it leads with what blocks the goal, then gives an ordered list
with an owner and a reason for each item. These answers use a stronger model, and
Conductor shows the most the answer may cost before it starts.

On a hosted workspace, "connected" means connected to your workspace, so Conductor can
read a connected repository for anyone in it. On Orchestra Local it reads the
repositories you enabled on your machine.

## Coming back

The thread follows the run: its status, its questions, its plan, its final report and
any pull request it opened. Answer a question or tell the agent something from the
thread, or open the session.

A new conversation opens with what needs you, what is in flight, and what changed,
each linked to its session.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.