AI Providers

AI Providers

Jarvis speaks to five backends: Ollama (local, free), Claude, OpenAI, Grok and Gemini. Configure as many as you like; he routes between them and fails over when one is unavailable.

Nothing here needs the config file. /jarvis bell → Admin → AI setup does all of it, and so does /jarvis ai.

---

Setting one up

From the menu. Ring the bell, Admin → AI setup. A row of providers: right-click toggles one on or off, left-click opens its page. There you set the key or the address, choose the model — picked from what your Ollama server actually offers, or typed — and run a test that goes past the routing to the provider itself. A key typed here is captured from chat without being shown or written to the chat log.

From the console.

/jarvis ai                          which providers are on, and their health
/jarvis ai enable claude
/jarvis ai key claude sk-ant-...
/jarvis ai endpoint ollama http://your-box:11434
/jarvis ai models                   what your Ollama server has pulled
/jarvis ai model ollama qwen2.5:7b
/jarvis ai test claude

Both take effect at once. No restart.

---

Tiered routing

Jarvis sorts work into two tiers and sends them different ways:

Tier	What it is	Goes to
Light	Chat intents, banter — frequent, small	Ollama first, cloud as fallback
Heavy	Build planning — rare, large	Cloud first, Ollama as fallback

That is the point of the design: the thing you do a hundred times an hour is free and local, and the thing you do twice an evening gets the better model.

ai:
  provider: auto
  provider-priority: [ollama, claude, openai, grok, gemini]
  light-timeout-seconds: 5      # a slow local box falls through quickly
  heavy-timeout-seconds: 240    # a build runs to thousands of tokens
  # routing:
  #   light: [ollama, claude]
  #   heavy: [claude, ollama]

Setting ai.provider to a single provider name uses that one exclusively.

A provider that fails goes on an exponential cooldown up to five minutes, so a box with no Ollama installed costs the next caller nothing. Providers with no API key are skipped, Ollama excepted.

---

Ollama, local and free

1. Install Ollama on any machine on your network.
2. Pull a model. Small instruction-following models handle chat parsing well:

ollama pull qwen2.5:7b
ollama pull mistral
ollama pull nomic-embed-text     # for experience memory; recommended

3. Point Jarvis at it: Admin → AI setup → Ollama, or /jarvis ai endpoint ollama http://your-box:11434.

ai:
  ollama:
    endpoint: "http://localhost:11434"
    model: mistral
    keep-alive: "30m"
    timeout-seconds: 240

Those are the shipped defaults. keep-alive holds the model in VRAM between requests, which is most of the difference between a snappy butler and a thoughtful one: reloading a 7B model costs seconds. The long timeout is for build plans, which a local 7B can take minutes to write.

Running only Ollama

Fully supported, and Jarvis enters a reduced mode, on every platform: constrained JSON parsing that small models handle reliably, and risky console actions refused unless ai.reduced-mode.allow-risky-actions is on. Build requests are answered from the schematic library; freeform designs are refused until experience memory holds memory.min-successes-for-reduced-mode-builds successful builds (20 by default), and then only through the json planner — the script planner needs a cloud model.

---

The cloud providers

Keys are set from the menu or with /jarvis ai key <provider> <key>.

Claude — console.anthropic.com

ai:
  claude:
    model: claude-haiku-4-5       # the sane default: cheapest, fastest
    # claude-sonnet-5             stronger, mid cost
    # claude-opus-5               most capable, most expensive

Model ids are complete as written; do not append a date.

OpenAI — platform.openai.com

ai:
  openai:
    model: gpt-5.6-terra          # balanced (default)
    # gpt-5.6-luna                cost-optimised, good for busy servers
    # gpt-5.6-sol                 flagship

Grok — console.x.ai · Gemini — aistudio.google.com

Because a server parsing every chat line makes a great many calls, the cheap fast model is the right default everywhere. Reach for the big one only for building.

---

Checking it works

/jarvis ai            provider health, models, cooldowns
/jarvis ai test claude
/jarvis debug         provider, model, memory and subsystem status

/jarvis ai test sends one small request straight to that provider, past the routing, so a failure tells you about the key rather than the fallback chain. It needs jarvis.admin, as do /jarvis ai models and every other setting command; plain /jarvis ai is open to all.

---

Experience memory

He remembers builds that went well and retrieves them when you ask for something similar, matching on what you asked and where you were standing. With nomic-embed-text pulled on your Ollama box the matching is semantic; without it, memory still works on a weaker keyword path.

memory:
  enabled: true
  embedding-model: nomic-embed-text