Feature
Swarm Mind: several AI models on one task
Swarm Mind is the part of Milly Lab where several AI models work on one task. Council Mode gives every model the same brief and shows the answers side by side to compare and merge; Project Mode has an orchestrator model split the task among worker models. Swarm Mind is included in the Max ($42/month) and Advanced ($72/month) plans.
Last updated:
What is the difference between Council Mode and Project Mode?
Council Mode duplicates the work on purpose, so you can compare several independent answers to one brief; Project Mode divides the work, so each model does a different part of a larger task.
| Council Mode | Project Mode | |
|---|---|---|
| What the models do | Each reviewer model answers the same brief on its own card, all at the same time | An orchestrator splits the task into subtasks; worker models do them in dependency order |
| What you get | One card per model and a Final card that merges them | One integrated deliverable, plus each subtask's raw output |
| Best for | Decisions, reviews and open questions with more than one defensible answer | Multi-part work: code changes, reports, research in several stages |
| Connected workspace | Reviewers read the files; the Final card can propose changes | Workers read the files and propose changes |
| Changes to your files | Staged until you press Apply | Staged until you press Apply |
How does Council Mode work?
In Council Mode every reviewer model you pick answers the same brief independently and at the same time, and the orchestrator model writes a Final card that combines the answers when you ask it to.
- Write the brief, attach reference material if you have any (it becomes a read-only Original card), and choose an orchestrator model and the reviewer models.
- Each reviewer answers on its own card without seeing the others and ends with a confidence rating: High, Medium or Low.
- Compare and refine: see what a card changed against the Original material, open a card's version history and restore an earlier version, or regenerate one card.
- Ask follow-up questions: chat with the model behind one card, or write to the whole council and @mention models to give each one its own part of the message.
- Start a debate when the cards disagree. It runs in 4 stages: argument, counterargument, evaluation of the evidence and a verdict.
- Merge into the Final card, either all cards at once or one card you choose. A card can be downloaded as a Markdown file.
How does Project Mode work?
In Project Mode an orchestrator model breaks your request into subtasks with dependencies, gives each subtask to the worker model whose strengths fit it, runs them in order and integrates the results into one deliverable.
- Stages: analyze, decompose, execute, review and integrate.
- Each subtask gets a category, such as coding, architecture, reasoning, writing, analysis, research or math, and goes to a worker model that is strong in it. Work is spread so that no worker you picked sits idle.
- During a run you can pause, resume or cancel it, and move a subtask to a different model.
- Your original request, the raw output of every subtask and the final result are all kept and visible. The result can include files to download.
- When the run is done, follow-up messages continue the same project.
- The run keeps going on the server if you close the page; when you open the session again, it replays what you missed.
Which plans include Swarm Mind?
Swarm Mind is included in the Max, Advanced and Enterprise plans; Free, Starter and Pro do not include it. In Swarm Mind you can use the chat models your plan includes.
| Plan | Price | Swarm Mind | Models in Swarm Mind |
|---|---|---|---|
| Free | $0 | No | — |
| Starter | $12 per month | No | — |
| Pro | $25 per month | No | — |
| Max | $42 per month | Yes | GPT-5.6 Terra, GPT-5.6 Sol*, Claude Sonnet 5, DeepSeek V4 Pro, Gemini 3.7 Flash, GPT-5.6 Luna, Claude Haiku 4.5 |
| Advanced | $72 per month | Yes | GPT-5.6 Terra, GPT-5.6 Sol*, Claude Sonnet 5, Claude Opus 5*, DeepSeek V4 Pro, Gemini 3.7 Flash, ChatGPT Codex*, GPT-5.6 Luna, Claude Haiku 4.5 |
| Enterprise | Contact sales | Yes | Set in the contract |
* Offered in Swarm Mind, not in plain chat.
How much does a Swarm Mind run cost?
Every model call in a run counts against that model's monthly allowance, the same as a chat message, so a council of four models costs about four answers, and debate rounds and project subtasks add more calls.
- When a model's monthly allowance is used up, further calls to it are charged to your top-up balance at 2× the provider's cost.
- Files a run reads from a connected workspace are sent to the models as input, so a large repository makes a run cost more.
- Up to 4 Council and Project sessions can run at the same time on one account.
Can Swarm Mind work on my GitHub repository or Google Drive?
Yes. You can attach a GitHub repository, Google Drive, OneDrive or a PostgreSQL database to a Council or Project session; the models read it, and every change they propose waits in a review panel until you press Apply.
How staged changes and Apply work