Models & Providers

Glue talks to LLMs through two wire protocols: Anthropic native and OpenAI-compatible. The OpenAI-compatible adapter covers OpenAI itself plus everything that speaks its API (Mistral, Ollama, Groq, OpenRouter, Gemini, vLLM, …). The full list of providers and models lives in the bundled catalog (browse it at /models) and can be extended with a ~/.glue/models.yaml overlay.

Set credentials with the standard env vars (e.g. ANTHROPIC_API_KEY, OPENAI_API_KEY) or in ~/.glue/credentials.json. See Configuration for the full list.

Picking the active model

Set it once in ~/.glue/config.yaml:

active_model: anthropic/claude-sonnet-4-6

Override per-invocation with the CLI flag:

glue -m openai/gpt-4.1
glue -m ollama/qwen3-coder:30b

For local Ollama, ollama/qwen3-coder:30b is the current recommended coding model in the bundled catalog.

Or switch interactively:

/model              # open the model picker (browse all models in the catalog)
/model gpt-5        # fuzzy-switch by name

Profiles (model shortcuts)

Profiles are named shortcuts for model refs. They make /model switching faster — they do not assign models to subagents (subagents take a model_ref argument directly when spawned).

Define them in ~/.glue/config.yaml as a flat map of name: provider/model:

profiles:
  fast: openai/gpt-4.1-nano
  deep: anthropic/claude-opus-4-6
  local: ollama/qwen3-coder:30b

Then switch with the @ prefix:

/model @fast
/model @deep

Adding a model the catalog doesn't know about

Drop a models.yaml into ~/.glue/ with your provider/model entry — it merges on top of the bundled catalog. Or, for any OpenAI-compatible endpoint, define a provider with adapter: openai and a custom base_url. See Troubleshooting → "The model I want isn't in the catalog".

Catalog CLI

The catalog is layered: bundled snapshot (compiled into the binary) → remote refresh (cached at ~/.glue/cache/models.yaml) → local overlay (~/.glue/models.yaml). Use glue catalog to inspect and refresh those layers without starting a session:

glue catalog refresh          # fetch the canonical models.yaml into the cache
glue catalog show             # print the active merged catalog (yaml)
glue catalog show --json      # same, as JSON
glue catalog path             # report where each layer is resolved from
glue catalog open             # open the canonical URL in your browser
glue catalog open --print     # print the URL instead of launching a browser
glue catalog edit             # open the cached models.yaml in $EDITOR

refresh writes YAML (not a one-line JSON dump) and preserves comments and key order from the upstream document, so the cached file stays readable in glue catalog edit. The canonical URL is catalog.remote_url in ~/.glue/config.yaml if set, otherwise the bundled default. GLUE_CATALOG_CACHE overrides the cache path used by edit.

See also

• Configuration — credentials and config keys
• LlmFactory — adapter dispatch
• AnthropicClient
• OpenAiClient
• OllamaClient