AI providers

CraftGPT 2.0 uses named provider profiles. You can define as many as you like under providers.profiles, pick one as the server default with providers.default, and give individual mobs a different one with /cg model <profile>.

Provider types

`type`	What it is	Tested live for this beta	Notes
`openai`	OpenAI	Yes (`gpt-6-luna`, with tools)	The default profile
`openai-compatible`	Any OpenAI-style server: vLLM, LM Studio, Ollama’s `/v1`…	Yes (via Ollama `/v1`, with tools)	`base-url` is required, and so is a non-blank `api-key` (any placeholder works if your server ignores it). Sends no reasoning effort, so “thinking” models are slow here
`anthropic`	Anthropic	Yes (`claude-haiku-5-5`, with tools)	Temperature max 1
`google`	Google Gemini	No: request-format tests and a code review
`ollama`	A native Ollama server	Yes (`qwen3:8b`, with tools)	Needs no API key. Use `reasoning-effort: none` for thinking models
`claude-cli`	Experimental. A logged-in Claude Code command-line tool on the server	Earlier build only	For personal or test servers; check the provider’s terms first
`codex-cli`	Experimental. A logged-in Codex command-line tool on the server	Earlier build only	Same as above

API keys

Set api-key to the key itself, or (recommended) to "${ENV_VAR}", which reads it from the server process’s environment. If you use ${ENV_VAR} and the variable isn’t set, the plugin switches itself off at startup. Keys are never logged.

Reasoning effort

reasoning-effort accepts default, none, minimal, low, medium, high or xhigh; which ones work depends on the model.

Keep none for the default model. With gpt-6-luna, anything other than none makes every chat that offers tools fail.

Profile settings

The shipped value column shows the openai profile in the default config.

Key	Shipped value	What it does
`type`	`openai`	One of the types above
`api-key`	`""`	The key, or `"${ENV_VAR}"`. Ignored for CLI types
`model`	`gpt-6-luna`	Model name. Required: a profile without one is a config error
`base-url`	`""`	Blank = the provider’s default. Required for `openai-compatible`
`max-output-tokens`	`1024`	Longest reply, in tokens. Ignored by CLI types
`reasoning-effort`	`none`	See above. A profile that leaves it out gets `default`, which breaks tool chats on `gpt-6-luna`, so always set it
`temperature`	`default`	`default`, or a number from 0 to 2 (0 to 1 for Anthropic). Ignored by CLI types
`timeout-seconds`	`60`	Time limit for one call
`cli-command`	`claude` / `codex`	CLI types only: the program name or full path

Example profiles

The default config ships these as comments, ready to uncomment:

    anthropic:
      type: anthropic
      api-key: "${ANTHROPIC_API_KEY}"
      model: claude-haiku-5-5
      temperature: default        # Claude 5.x models advise against setting temperature
    google:
      type: google
      api-key: "${GOOGLE_API_KEY}"
      model: gemini-3.5-flash-lite
      reasoning-effort: default   # Gemini 3 can't turn thinking off; use default or minimal, not none
    local:
      type: ollama
      base-url: "http://localhost:11434"
      model: qwen3:8b             # about 5 GB, supports tools
      reasoning-effort: none      # otherwise qwen3 thinks before every reply (much slower)

It also includes two commented CLI profiles, claude-subscription (claude-cli, claude-haiku-5-5) and codex-subscription (codex-cli, gpt-6-luna).

CLI providers (experimental)

claude-cli and codex-cli run a command-line AI tool that’s already logged in on the server, instead of calling an API.

The system user that runs the server must be logged in to the tool, with the program on its PATH or given as a full path in cli-command.
They ignore api-key, base-url, temperature and max-output-tokens.
CraftGPT removes API-key environment variables from the tool’s environment so it uses the login.
For Claude, none and minimal effort become low (with a warning); default sends no effort.