Configuring an LLM provider

Settings → Advisor is the manual side of getting the Advisor a model: pick Anthropic (bring your own API key) or Ollama (a fully local LLM you run yourself), test the connection, then save. This is separate from the local-agent bridge, which detects a coding agent already on your machine instead of using either of these.

deckgauge · Settings · Advisor
DGDeckgauge Hosted
ANYour own API key
OLLocal Ollama
Connected — claude-haiku-4-5 (420ms)

Choosing a provider, then a successful Test connection.

What it does, and where the data comes from

Two real choices, plus a disabled "Deckgauge Hosted" card in this self-hosted build:

  • Your own API key (provider anthropic) — an API key and a Model (e.g. claude-haiku-4-5). The key is stored server-side and never returned by the API — the settings page only ever tells you whether one is already saved, and re-saving replaces it rather than requiring you to re-enter it to view it.
  • Local Ollama (provider ollama) — a Base URL (e.g. http://localhost:11434) and a Model (e.g. llama3.2). Everything stays on your own infrastructure.

Test connection sends a minimal one-token request through the candidate config and reports the model name and round-trip latency on success, or the failure reason if the provider rejects it or can't be reached. Save persists whichever config you tested (testing isn't required before saving, but it's the only way to know the config actually works).

How to configure it

  1. Open Settings → Advisor (/settings/advisor).
  2. Pick a provider card — "Your own API key" or "Local Ollama".
  3. Fill in its fields — API key + Model for Anthropic, or Base URL + Model for Ollama.
  4. Test connection — confirms the model and reports latency before you commit to it.
  5. Save — the config becomes the board's advisor provider immediately.

Under the hood

A config saved here can also be set declaratively via environment variables (ADVISOR_PROVIDER, ADVISOR_MODEL, plus ADVISOR_ANTHROPIC_API_KEY or ADVISOR_OLLAMA_BASE_URL) — see configuration reference. A row saved through this page always wins over those variables: env is the deployment default, this page is the operator's override. An incomplete or invalid env config (for example, an Ollama URL missing its http:// scheme) is treated as no provider at all rather than half-applied.

If neither a saved row nor a valid env config exists — and the local-agent bridge isn't connected either — asking a question in the panel fails with advisor_not_configured, which the panel renders as a plain-language message linking back to this page.

If it looks wrong

SymptomCauseFix
Test connection failsWrong API key, base URL, or model name — or Ollama isn't reachable at that URLRe-check the field the error names, and confirm Ollama is running there
Clicking "Deckgauge Hosted" does nothingThat option is disabled in this self-hosted buildPick "Your own API key" or "Local Ollama" instead
Save fails with an error bannerThe API rejected the config (e.g. a required field is missing)Read the message, fix the field it names, and save again
Asking a question still says no model is configuredThe config wasn't saved, only testedClick Save after a successful Test connection

Related

Last updated