Labsco
MCP SERVER

LLM Router

by ypollak2 ↗

Route each prompt to the cheapest model that can still do the job — free or local first, your paid seat only when the task needs it — from inside Claude Code, Codex, Cursor or Gemini CLI.

Model Routing, Multi-Model Consultation & Cost Control
Summary
The secrets rule is the part worth knowing before you route anything real.

A prompt that contains an API key, token or private key is routed to local models only, fail-closed — it cannot leak into an external provider because a cheaper tier happened to win. Around that, the useful habit is calling the classifier before the router: it shows the complexity verdict and the budget pressure that would drive the decision, so you can override rather than discover the choice after the bill.

What it is

A local-first routing layer exposed over MCP. A complexity classifier — regex heuristics for most prompts, a local Ollama or Gemini Flash call for the ambiguous ones — decides how hard a task is, then a free-first router walks up a chain of providers until one can handle it, with circuit breakers skipping providers that are failing or rate-limited. Cost is logged to a local SQLite database. There is no hosted proxy and no account.

What you get
  • Routing you can call directly or inspect: classify a prompt's complexity before committing, route with full parameter control, stream, or hand it off automatically — `llm_classify`, `llm_route`, `llm_auto`, `llm_stream`, `llm_select_agent`, `llm_track_usage`, `llm_reroute`
  • Task-shaped entry points that pick their own routing chain — factual questions, web-grounded research, long-form generation, deep analysis, code and edits — `llm_query`, `llm_research`, `llm_generate`, `llm_analyze`, `llm_code`, `llm_edit`
  • Media and multi-step pipelines — `llm_image`, `llm_video`, `llm_audio`, `llm_orchestrate`, `llm_pipeline_templates`
  • Spend and health visible while you work: usage, savings against baseline, provider health, budget and a dashboard, plus a session-end summary with tier mix, per-provider cost and latency p50/p95/p99 — `llm_usage`, `llm_savings`, `llm_health`, `llm_providers`, `llm_budget`, `llm_dashboard`, `llm_set_profile`, `llm_quality_guard`, `llm_hook_health`
  • A fail-closed secrets rule: a prompt containing an API key, token or private key routes to local models only and cannot reach an external provider
  • Setup, benchmarking and filesystem helpers, and Claude subscription usage tracking — `llm_setup`, `llm_benchmark`, `llm_quality_report`, `llm_save_session`, `llm_fs_find`, `llm_fs_rename`, `llm_fs_edit_many`, `llm_fs_analyze_context`, `llm_check_usage`, `llm_update_usage`, `llm_refresh_claude_usage`
Requirements

Python 3.11+ and `pip install llm-routing` — the PyPI name is `llm-routing` while the command is `llm-router`; the published version is 13.0.2. `llm-router install` registers it with Claude Code by default, and `--host codex`, `--host gemini-cli`, `--host vscode` or `--host cursor` targets the others. No API keys are required on a Claude Pro or Max plan; `LLM_ROUTER_SUBSCRIPTION_PROVIDER` opts into cost-inverted routing around a paid seat. Everything runs locally — no hosted proxy, no account — and cost data stays in a local SQLite database.

Setup effort

One command — pip install llm-routing