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

# Director

> The interactive Claude session that spawns and coordinates Murmur agents on your behalf using the MCP server — the human-in-the-loop entry point.

A **director** is an interactive Claude Code session, running the built-in [`murmur-director`](/concepts/agent-personas) [agent persona](/concepts/agent-personas) with the [Murmur MCP server](/mcp-server/overview) attached. It is the recommended way to manage [agents](/concepts/agents) for any task that involves more than one agent or more than one step.

The director itself does not write code. It [spawns](/cli/spawn) autonomous agents, monitors them via [`mcp__murmur__status`](/mcp-server/status) and [`mcp__murmur__ls`](/mcp-server/ls), reviews the pull requests they open, and waits for your call to merge, follow up, or kill.

## Start a director

```bash theme={null}
murmur director
```

This launches Claude Code with the `murmur-director` persona and the MCP server pre-wired. See [`murmur director`](/cli/director) for flags, including `--persona` to override the default.

## What the director does

1. **Plans.** Takes your goal, decides if it's one task or many, picks a persona per piece, sequences dependencies.
2. **Spawns.** Calls [`mcp__murmur__spawn`](/mcp-server/spawn) for each piece. Parent and child relationships are preserved automatically.
3. **Monitors.** [Events](/concepts/events) from agents (PR opens, CI results, child lifecycle, follow-ups) arrive on the director's channel and surface to you in real time.
4. **Reviews.** Runs `gh pr diff` and `gh pr checks` on each PR, summarizes the change, and presents merge/rework recommendations.
5. **Cleans up.** Proposes killing completed agents after their PRs land.

## Behavioral contract

The `murmur-director` persona is constrained to make the director predictable as a coordinator:

* **Questions stay read-only.** If you ask "should we do X?", the director answers. It does not spawn, edit, or merge. Discussion is not authorization.
* **Destructive actions require an explicit imperative.** Merge, kill, and close need a clear "do it" from you. The director will recommend; it will not act.
* **Facts pass through unmodified.** Specific facts you provide (error messages, IDs, config snippets) are forwarded to spawned agents verbatim, never paraphrased or filtered.
* **The human is the quality gate.** The director gates PRs against CI and presents its review. You decide what lands on `main`.

## When not to use a director

A director is overhead for a single, well-scoped task. For one-shot work, use [`murmur spawn`](/cli/spawn) directly, or [`mcp__murmur__spawn`](/mcp-server/spawn) from inside another agent's session. The director's value comes from the conversational loop across multiple agents; if the loop has one step, you don't need it.

## Related

| Type      | Page                                                           |
| --------- | -------------------------------------------------------------- |
| Reference | [`murmur director`](/cli/director)                             |
| Reference | [`mcp__murmur__spawn`](/mcp-server/spawn)                      |
| Reference | [`mcp__murmur__status`](/mcp-server/status)                    |
| Concept   | [Agent Personas](/concepts/agent-personas)                     |
| Concept   | [Agents](/concepts/agents)                                     |
| Concept   | [Events](/concepts/events)                                     |
| Guide     | [Multi-agent orchestration](/guides/multi-agent-orchestration) |