Skip to main content
Max uses GitHub Copilot as its AI backend, which means any model available in your authenticated Copilot CLI can be used. The active model is set via COPILOT_MODEL in ~/.max/.env.

​
Setting the default model

​
COPILOT_MODEL
string
default:"claude-sonnet-4.6"
The model ID used by the orchestrator session. Must be available in your authenticated Copilot CLI.To list all available models: copilot listModels
The setup wizard presents available models fetched live from your Copilot CLI and writes your choice to ~/.max/.env. If Copilot is not yet authenticated during setup, it falls back to a curated list.

​
Available models

The following models are shown during max setup when Copilot is not yet authenticated:
Model IDNameNotes
claude-sonnet-4.6Claude Sonnet 4.6Default — fast, great for most tasks
gpt-5.1GPT-5.1OpenAI’s fast model
gpt-4.1GPT-4.1Free included model
The actual list of available models depends on your Copilot subscription and which models are enabled for your account. Run copilot listModels to see what’s available in your environment.

​
Switching models

In the terminal UI, use the /model command:
/model gpt-4.1
Or ask Max in plain English:
Switch to gpt-4.1
The change takes effect immediately and is persisted to ~/.max/.env.

​
Model persistence

Whenever you switch models at runtime, Max calls persistModel() to write the new value back to ~/.max/.env:
src/config.ts
export function persistModel(model: string): void {
  persistEnvVar("COPILOT_MODEL", model);
}
This means your model choice survives daemon restarts — the next max start picks up whatever was last written to ~/.max/.env.

​
Automatic model routing

Max includes an optional model router that automatically selects the most appropriate model based on the complexity of each message. It is disabled by default.

​
How it works

When enabled, every incoming message goes through a two-stage selection process:
  1. Keyword overrides — checked first, bypass the cooldown. Certain keywords force a specific model regardless of classification.
  2. LLM classification — the message is classified as fast, standard, or premium by GPT-4.1. Each tier maps to a configured model.
A cooldown mechanism prevents rapid model switching between consecutive messages.

​
Tier-to-model defaults

TierDefault modelUse case
fastgpt-4.1Short answers, quick lookups
standardclaude-sonnet-4.6Most conversational tasks
premiumclaude-opus-4.6Complex reasoning, design tasks

​
Keyword overrides

The router ships with one built-in override rule:
Rule nameKeywordsModel
designdesign, ui, ux, css, layout, styling, visual, mockup, wireframe, frontend design, tailwind, responsiveclaude-opus-4.6
Keyword matching uses word boundaries, so "ui" matches "update the UI" but not "fruit".

​
Router configuration

The router configuration is stored in the max_state SQLite table under the key router_config as a JSON blob: 
{
  "enabled": false,
  "tierModels": {
    "fast": "gpt-4.1",
    "standard": "claude-sonnet-4.6",
    "premium": "claude-opus-4.6"
  },
  "overrides": [
    {
      "name": "design",
      "keywords": ["design", "ui", "ux", "css", "layout", "styling", "visual",
                   "mockup", "wireframe", "frontend design", "tailwind", "responsive"],
      "model": "claude-opus-4.6"
    }
  ],
  "cooldownMessages": 2
}
The cooldownMessages field controls how many messages must pass before the router is allowed to switch models again, preventing thrashing.

​
Follow-up handling

Short follow-up replies (“yes”, “do it”, “go ahead”, “looks good”, etc.) automatically inherit the tier from the previous turn rather than triggering a new classification. Background task completion messages always resolve to standard regardless of content.
When auto-routing is off (the default), Max uses whichever model is set in COPILOT_MODEL for every message. Enable routing only if you want Max to use cheaper models for simple questions and reserve premium models for complex work.

Build docs developers (and LLMs) love

Get started for freeTalk to us