| name | local-model-router |
| description | Applies a three-tier routing policy using only local Ollama models (free/local-first). Trigger on offload easy tasks, latency/cost concerns, offline/privacy-sensitive work, routing policy decisions, and local tier selection. |
| disable-model-invocation | true |
Local model router
Canon lives in %USERPROFILE%\.cursor\OLLAMA_ROUTING.md — follow that file before improvising routing.
Default tiers (local-only)
Easy -> smallest local coder model
Signals: trivial docstrings, renaming, micro-refactors (<30 LOC), summaries of small pasted snippets.
Medium -> larger local coder model
Signals: focused module explanation, bounded tests (<3 files touched), narrowing unknown root-cause hypotheses.
Hard -> strongest local model OR pause and ask
Signals: ambiguous multi-file regressions, security-sensitive audits, roadmap-level architecture spanning services.
If local quality is insufficient, ask for explicit approval before any paid/cloud path.
Privacy overrides
Treat as stay local: secrets, HIPAA/PCI, sensitive customer payloads, unpublished roadmap, embargoed exploits. Fail closed.
MCP + CLI integration
| Surface | Capability |
|---|
Cursor MCP (ollama-local) | Lightweight tool calls into local inference |
Ollama CLI (ollama run) | Manual local fallback without MCP |
Windows reminders
Ensure Ollama service is reachable at OLLAMA_HOST (defaults to http://127.0.0.1:11434). If MCP server fails (npx -y ollama-mcp), rely on manual local CLI.
