Configuration

Config lives at ~/.config/cmx/config.json (Windows: %AppData%\cmx\config.json). It is created with defaults on first run; edit and run cmx restart to apply.

{
  "gateway":   { "host": "127.0.0.1", "port": 17322 },
  "web":       { "enabled": true, "port": 17323 },
  "providers": {
    "openai":    { "base_url": "https://api.openai.com/v1" },
    "anthropic": { "base_url": "https://api.anthropic.com" },
    "openrouter":{ "base_url": "https://openrouter.ai/api/v1" },
    "groq":      { "base_url": "https://api.groq.com/openai/v1" },
    "deepseek":  { "base_url": "https://api.deepseek.com/v1" },
    "moonshot":  { "base_url": "https://api.moonshot.cn/v1" },
    "zai":       { "base_url": "https://api.z.ai/api/paas/v4" }
  },
  "compression": {
    "enabled": true,
    "min_context_tokens": 60000,
    "keep_recent_tokens": 24000,
    "summarizer_model": "",
    "nudge_frequency": 5,
    "protected_tools": ["task", "skill", "todowrite", "todoread", "write", "edit"]
  },
  "logging": { "enabled": true, "retention_days": 14 }
}

Keys

`gateway`

Key	Default	Meaning
`host`	`127.0.0.1`	bind address — keep it local
`port`	`17322`	gateway port

`web`

Key	Default	Meaning
`enabled`	`true`	serve the web dashboard
`port`	`17323`	dashboard port

`providers`

Map of cmx-provider header values → upstream base_url. The dialect is detected per request from the path (/v1/messages → Anthropic-style, /v1/chat/completions → OpenAI-style), so one provider entry can serve either. If the header is missing, cmx guesses from the Authorization scheme and host.

`compression`

Key	Default	Meaning
`enabled`	`true`	master switch — when off, cmx is a pure transparent proxy
`min_context_tokens`	`60000`	context size above which nudges are appended
`keep_recent_tokens`	`24000`	tail of recent context never compressed or ID-tagged
`summarizer_model`	`""`	model for summary calls; empty = same model as the session. Set a cheap one (e.g. `gpt-4o-mini`) to minimize summarization cost
`nudge_frequency`	`5`	append a compress reminder every N requests while above threshold
`protected_tools`	see above	messages using these tools are never compressed

`logging`

Key	Default	Meaning
`enabled`	`true`	store full request payloads (original bodies) locally in SQLite for stats and debugging
`retention_days`	`14`	payloads older than this are purged hourly; stats aggregates are kept