With orchestration, a conductor leads the task: it makes the plan, hands each step to a role, reviews the work and decides what comes next. Each role works with its own model, the one that is enough for it. The conductor itself has no tool to edit a file or run a command. You need Moon CLI 0.8.0 or later (moon update).
Choose orchestrationLink to the Choose orchestration section
In Code (on the web with moon rc, or in Moon Desktop) as in the terminal, choose “Orchestration” in the list of models: “A conductor shares the work between several models.” The entry appears when Moon AI offers it for your account, and disappears if your organization forbids it.
The rolesLink to the The roles section
- Conductor: makes the plan, reviews the work and decides.
- Reading: reads and searches the files.
- Code: writes and edits the code.
- Commands: runs commands and tests, makes small fixes.
- Security tests: prepares an intrusion test on an application of your computer. Off by default (see below).
Who does whatLink to the Who does what section
In the Code page, “Who does what…”, next to Orchestration, shows the model of each role. “Each role uses the model you choose. Your limit pays for them.”
- By default, each role follows a recommendation (“Recommended model” followed by the model), worked out from the models you can use in this workspace, their capabilities and their price: a strong model for the conductor, and for each step a model that is enough for it. The reasons are written under the model: “Knows how to use tools”, “Strong in code”, “About … times cheaper than the conductor”, “About … % of your 5-hour limit per message”.
- Your choice: take another model for a role (“Your choice ·” followed by the model). A model that does not suit a role stays in the list, greyed out, with the reason (“Does not suit this role”). “Back to the recommendation” sets the role back to its default.
- Your choice applies to this workspace, on all your devices.
During a taskLink to the During a task section
- The split is read again from Moon AI at each turn: a changed role, or a rule of your organization, applies from the next message, without restarting Moon CLI. If Moon AI does not answer for a moment, the last split is still used for 10 minutes at most, without the security tests.
- Each step counts in your limit at the price of the model that did it, and goes through Moon Blur like the rest.
- In Code, except in Autonomous and No limits, the conductor first makes the plan, then asks you “Launch this plan?” before editing your files or running a command.
- In Code, the plan of the conductor fits on one line (“Plan · 1 of 4 done”, then “Plan done”) and its progress is in “Advancement”. Each sub-agent has its card: its role, its model, its steps and its duration. The reads and searches that follow each other fit on one line, “6 files read · 1 search”, which opens with a click.
Organizations and teamsLink to the Organizations and teams section
The rules of your organization or your team apply. Its administrators set them in the “Code · Orchestration” section of the rules page:
- “Allow the orchestration”: turned off, the entry Orchestration is not offered to the members.
- “Allow security tests”: off by default in an organization and in a team.
- “Default split of the members”: the starting model of each role, the recommendation or one of the allowed models.
- “Lock this role”: the members see “Set by your organization” and cannot change it.
A model your organization does not allow is never chosen.
Security testsLink to the Security tests section
A specialist prepares requests against an application running on your computer, and Moon sends them. Only test your own applications.
- Turn them on: in the Code page, “Security tests”, for this session only, when Moon AI offers them for your account and your organization allows it. In an organization or a team, an administrator must open them first.
- Your approval for each port: before sending requests to an application, Moon asks for your approval, for each port and each session, in every mode, Autonomous and No limits included. The card gives the port, the name of the program listening on it and its process number (otherwise “Unknown process”), and warns that the requests can change its data. If another program takes this port, the card comes back.
- Only the device that turned the tests on can grant them. Elsewhere (the terminal during
/rc, a page controlling Moon Desktop remotely), you can only refuse; in the terminal, Esc and Ctrl+C refuse too. Turned off from the page, by closing the page or by stopping/rc, nothing more is sent, even in the middle of a series. - Only your computer: the requests only go to 127.0.0.1 or ::1, never to the ports of Moon CLI nor to known control ports (Docker, Chrome and Node debugging, Kubernetes, Redis, Ollama, e-mail servers…). They are limited (40 per task, 120 per hour and per session), with no redirect followed, and each one is recorded in the session.
- The conductor receives a verdict and one line of proof, never the list of requests.