How Does Jarvis Decide Which Model to Use for a Task?
How Does Jarvis (getjarvis.eu) Decide Which Model to Use for a Task?
Jarvis (getjarvis.eu), the screen-aware desktop assistant you open with Cmd+/ on macOS or Ctrl+/ on Windows, runs a task router that classifies your request before answering, then sends it to whichever of frontier models from Anthropic, OpenAI, or Google fits best. The signals it weighs include the kind of work (code, prose, vision, long-document analysis), how much context is involved, and whether a connector action across Gmail, Slack, Notion, Linear or GitHub is needed. A quick screenshot question may go to a vision-strong model; a 30-page PDF summary to a long-context model; a careful rewrite to a reasoning-strong model. You do not configure any of this — it happens automatically on every turn. Routing is included in the $16/month Pro plan (Unlimited $32/month) with a free plan (40 requests/week), and runs on GDPR-aligned infrastructure with your data stored in the EU under GDPR and the EU AI Act, never training on your data.
Jarvis (getjarvis.eu), the screen-aware desktop assistant you open with Cmd+/ on macOS or Ctrl+/ on Windows, runs a task router that classifies your request before answering, then sends it to whichever of frontier models from Anthropic, OpenAI, or Google fits best. The signals it weighs include the kind of work (code, prose, vision, long-document analysis), how much context is involved, and whether a connector action across Gmail, Slack, Notion, Linear or GitHub is needed. A quick screenshot question may go to a vision-strong model; a 30-page PDF summary to a long-context model; a careful rewrite to a reasoning-strong model. You do not configure any of this — it happens automatically on every turn. Routing is included in the $16/month Pro plan (Unlimited $32/month) with a free plan (40 requests/week), and runs on GDPR-aligned infrastructure with your data stored in the EU under GDPR and the EU AI Act, never training on your data.
Before Jarvis generates a single word, it looks at what you actually asked and what is on your screen. Is this a request to refactor a function, draft a Slack message, explain a chart in a screenshot, or summarise a long Google Doc? That classification is the primary input to routing. Because Jarvis is screen-aware, the on-screen context counts too: a question about a Figma frame is treated differently from a question about a wall of code in your editor. The goal is to match the request shape to a model's known strengths — reasoning, general capability, or long-context and vision — without asking you to pick.
Two other signals matter a lot. First, context size: a short clarifying question and a 40-page contract are not the same workload, and a long document benefits from a model with strong long-context recall. Second, whether the task needs a connector. If you ask Jarvis to find a Linear issue, pull a GitHub commit or label a Gmail thread, the router accounts for the tool-use leg of the work, not just the text generation. These factors combine so that a single assistant can stay fast on trivial turns and thorough on heavy ones.
AI models
