ypollak2/llm-router

Route every AI call to the cheapest model that can do the job well. 48 tools · 20+ providers · personal routing memory · budget caps, dashboards, traces.

PyPI Tests Downloads Python MCP License Stars

Average savings: 60–80% vs running everything on Claude Opus.

Token Savings Proof

Real numbers from a 14-day sprint: 51 releases, 22.6M tokens, $6.95 spent.

Quota Pressure Reduction

Free-first routing eliminated budget pressure over 14 days—allowing sustainable feature velocity.

Cost Impact

• Actual spend: $6.95 (22.6M development tokens)
• Opus baseline: $50–60 (300M+ tokens, traditional approach)
• Savings: $43–53 (87% cost reduction, 94% token reduction)

Token Distribution by Routing Tier

Free-first routing achieved:

• 31% from free models — Ollama (local) + Codex (prepaid): 7.0M tokens, $0 cost
• 38% from budget models — Gemini Flash + GPT-4o-mini: 8.6M tokens, $2.82 cost
• 31% from premium models — GPT-4o, Gemini Pro, Claude: 7.0M tokens, $4.13 cost

Monthly Projection

• 1 sprint (14 days): $6.95
• 1 month (~30 days): ~$15
• 1 year: ~$180

Compare: Claude Opus baseline for same work = $1,200–1,500/year

Install

pipx install claude-code-llm-router && llm-router install

Host	Command
Claude Code	llm-router install
VS Code	llm-router install --host vscode
Cursor	llm-router install --host cursor
Codex CLI	llm-router install --host codex
Gemini CLI	llm-router install --host gemini-cli

Supported Development Tools

llm-router works as an MCP server inside any tool that supports MCP, providing unified routing across your entire development environment.

Tool	Status	What You Get
Claude Code	✅ Full	Auto-routing hooks + session tracking + quota display
Gemini CLI	✅ Full	Auto-routing hooks + session tracking + quota display
Codex CLI	✅ Full	Auto-routing hooks + savings tracking
VS Code + Copilot	✅ MCP	llm-router tools available (routing is model-voluntary)
Cursor	✅ MCP	llm-router tools available (routing is model-voluntary)
OpenCode	✅ MCP	llm-router tools available (routing is model-voluntary)
Windsurf	✅ MCP	llm-router tools available (routing is model-voluntary)
Any MCP-compatible tool	⚡ Manual	Add llm-router to your tool's MCP config

Full Support vs MCP Support

Full support = auto-routing hooks fire before the model answers, enforcing your routing policy. MCP support = tools are available, but the model chooses whether to use them.

Quick Setup by Tool

Claude Code

pipx install claude-code-llm-router
llm-router install

Then in Claude Code, llm_route and friends appear as built-in tools. Your settings control the profile (budget/balanced/premium).

Gemini CLI

pipx install claude-code-llm-router
llm-router install --host gemini-cli

Gemini CLI users get full routing experience: auto-routing suggestions, quota display, and free-first chaining (Ollama → Codex → Gemini CLI → paid).

Codex CLI

pipx install claude-code-llm-router
llm-router install --host codex

Codex integrates deep into the routing chain as a free fallback when your OpenAI subscription is available.

VS Code / Cursor / Others

pipx install claude-code-llm-router
llm-router install --host vscode  # or --host cursor

The MCP server loads automatically. Tools appear in your IDE's model UI.

What It Does

Intercepts prompts and routes them to the cheapest model that can handle the task. Most AI sessions are full of low-value work: file lookups, small edits, quick questions. Those burn through expensive models unnecessarily.

llm-router keeps cheap work on cheap/free models, escalates to premium models only when needed. No micromanagement required.

• Works in: Claude Code, Cursor, VS Code, Codex, Windsurf, Zed, claw-code, Agno
• Free-first: Ollama (local) → Codex → Gemini Flash → OpenAI → Claude (subscription)

Mental Model

Think of llm-router as a smart task dispatcher. When you ask a question:

1. Analyze — What kind of task is this? (simple lookup vs. complex reasoning)
2. Choose — Which model can handle this best and cheapest?
3. Check Constraints — Are we over budget? Is this model degraded?
4. Execute — Send to that model

The dispatcher learns over time: if a model starts performing poorly (judge scores drop), it gets demoted in future decisions. If you're running low on quota (budget pressure), it automatically uses cheaper models. You don't manage any of this—it just happens behind the scenes.

Example: "Explain this error message" → Simple task → Route to Haiku (fast, cheap) → Done. vs. "Refactor this complex architecture" → Complex task → Route to Opus (expensive but thorough) → Done.

The savings come from not using Opus for every question.

New in v7.0.0 — Free-First MCP Chain & Ollama Auto-Startup

Major release with optimized routing chains and automatic Ollama management.

• Ollama Auto-Startup — Session-start hook automatically launches Ollama and loads budget models (gemma4, qwen3.5) if not running

  • Eliminates manual setup — local free inference available immediately
  • Graceful fallback if Ollama unavailable
  • 10-second readiness timeout with model auto-pull

• Free-First MCP Chain for All Complexity Levels

  • Simple tasks → Ollama → Codex → Gemini Flash → Groq
  • Moderate tasks → Ollama → Codex → Gemini Pro (improved quality-to-cost) → GPT-4o → Claude Sonnet
  • Complex tasks → Ollama → Codex → o3 → Gemini Pro → Claude Opus
  • Codex injected before all paid externals as free fallback when subscription available

• BALANCED Tier Chain Reordering — Gemini Pro prioritized after Codex injection

  • Previously defaulted to expensive DeepSeek for moderate tasks
  • Now balances cost + quality: Codex → Gemini Pro (better ROI) → paid fallbacks
  • Reduces BALANCED tier spend ~40% while maintaining output quality

• Routing Decision Logging & Analytics

  • Track which model selected for each task, cost impact, complexity distribution
  • Session-end hook shows routing summary with savings vs. full-Opus baseline
  • Identify anomalies (e.g., high-cost tasks that should route cheaper)

See CHANGELOG.md for full version history and v6.x features.

New in v7.4.0 — Content Generation Routing Discipline

Smart content generation detection with automatic routing suggestions.

• Automatic Content Generation Detection — Hook detects "write", "draft", "add card", "create spec" patterns

  • Prevents routing misses where content-generation tasks skip llm_generate routing
  • Suggests decomposition: route generation first, integrate locally second
  • Example: "add carousel card about X to file.md" → auto-routes via llm_generate

• Decomposition Patterns — Multi-step content+file tasks now route intelligently

  • "Generate narrative" → llm_generate → Done
  • "Add card to blueprint" → llm_generate content → Edit file integration
  • Cost impact: ~90% savings on writing tasks (route cheap model vs. expensive local generation)

• Soft Nudges via Hook Suggestion (not blocking)

  • Detects multi-step content generation patterns
  • Suggests: "Consider routing via llm_generate first, then integrate locally"
  • Enforces routing discipline without forcing user behavior

• Fast-Path for Content Tasks — Content generation routed instantly without waiting for classifier

  • Patterns: simple generation, decomposition, refinement, documentation
  • Same speed as code detection fast-path
  • Seamless fallback if pattern doesn't match

See CLAUDE.md § Content Generation Routing for detailed decision tree.

How It Works

User Prompt
    ↓
[Complexity Classifier] — Haiku/Sonnet/Opus?
    ↓
[Free-First Router] — Ollama → Codex → Gemini Flash → OpenAI → Claude
    ↓
[Budget Pressure Check] — Downshift if over 85% budget
    ↓
[Quality Guard] — Demote if judge score < 0.6
    ↓
Selected Model → Execute

Configuration

Zero-config by default if you use Claude Code Pro/Max (subscription mode).

Optional env vars:

OPENAI_API_KEY=sk-...                   # GPT-4o, o3
GEMINI_API_KEY=AIza...                  # Gemini Flash (free tier)
OLLAMA_BASE_URL=http://localhost:11434  # Local Ollama (free)
LLM_ROUTER_PROFILE=balanced             # budget|balanced|premium
LLM_ROUTER_COMPRESS_RESPONSE=true       # Enable response compression

For full setup guide, see docs/SETUP.md.

Monitoring & Reducing Violations

Routing violations occur when Claude bypasses a routing directive by using Bash, Read, Edit, or Write instead of calling the routed MCP tool first. This burns expensive tokens with zero cost savings.

What Is a Violation?

When llm-router issues a ⚡ MANDATORY ROUTE hint, it writes a pending state file. If Claude uses Bash (or Read/Edit/Write for Q&A tasks) before calling the expected tool, enforce-route.py logs it as a violation.

Example violation sequence:

⚡ MANDATORY ROUTE: query/simple → call llm_query
  → Bash: "I'll answer this directly"  ❌ VIOLATION (should have called llm_query)

Cost impact: $0.10+ spent on full Claude model instead of $0.0001 via llm_query routing.

Analyze Your Violations

Use the provided analysis script to see which sessions violate most and why:

python3 scripts/analyze-violations.py

Output:

• Summary: total violations, date range, distinct sessions
• Top 10 sessions table with violation counts
• Per-session tool sequences showing what was used vs what should have been called
• Report saved to ~/.llm-router/retrospectives/violation-report-<date>.md

Common Violation Patterns

Pattern	Cause	Fix
Bash used for Q&A	Claude answers directly	Route via llm_query / llm_research instead
Read after route hint	Claude investigates before routing	Call llm_analyze first, pass file content
Edit without generation	Claude codes directly	Route via llm_code first for simple tasks
Loop: same tool 3+ times	Investigation stuck in debugging	Call the routed tool to break the deadlock

Enforcement Modes

Control violation behavior via LLM_ROUTER_ENFORCE:

export LLM_ROUTER_ENFORCE=smart   # (default) Hard for Q&A, soft for code
export LLM_ROUTER_ENFORCE=hard    # Block all violations (strictest)
export LLM_ROUTER_ENFORCE=soft    # Log violations, allow calls (permissive)
export LLM_ROUTER_ENFORCE=off     # Disable enforcement entirely

• smart: Balances cost savings with developer UX. Q&A tasks (query/research/analyze/generate) are hard-blocked. Code tasks allow file-reading tools for investigation.
• hard: Strict enforcement. All violations blocked until routing is satisfied. Use when budget pressure is high.
• soft: Advisory only. Violations logged to enforcement.log but never blocked. Good for testing.
• off: No routing enforcement. Hooks fire but don't block.

Session Violation Escalation

After 3+ violations in a session, enforce-route.py prints a warning to stderr:

[llm-router] ⚠️  ESCALATION: 5 routing violations this session.
  Next prompt expecting llm_query:
  → Call the MCP tool FIRST before any Bash/Read/Edit/Write.
  → See ~/.llm-router/enforcement.log for full history.
  → Set LLM_ROUTER_ENFORCE=hard to block violations automatically.

This reminds the model to route first.

Interpreting enforcement.log

The log file ~/.llm-router/enforcement.log contains all violations:

[2026-04-26 10:30:45] VIOLATION session=abc12345678 expected=llm_query got=Bash
[2026-04-26 10:31:02] VIOLATION session=abc12345678 expected=llm_query got=Read

Use analyze-violations.py to summarize and find patterns across sessions.

Improving Hint Visibility

v7.5.0+ uses a box-drawing format that's harder to miss:

╔══════════════════════════════════════════════════╗
║  ⚡ MANDATORY ROUTE — DO NOT SKIP                ║
║  task  : query                                   ║
║  action: call llm_query                          ║
║  via   : heuristic                               ║
║  saves : $0.001                                  ║
╚══════════════════════════════════════════════════╝

⚠️  IMPORTANT: Call the tool above as your FIRST action.
   • Do NOT use Bash, Read, Edit, or Write to self-answer
   • Do NOT spawn Agent subagents — they cost $0.10+
   • Do NOT use WebSearch or WebFetch — route via llm_research
   • Violations are logged per-session and count toward escalation

This format is visible even in long context windows.

Hook Health Cleanup

Test artifacts from development can inflate error counts. Clean them up:

# Preview what will be removed
python3 scripts/cleanup-hook-health.py --dry-run

# Apply cleanup
python3 scripts/cleanup-hook-health.py

# Force-remove specific hooks
python3 scripts/cleanup-hook-health.py --remove hook-a hook-b test-hook

This removes hooks that only have errors from test sessions (session_id: "abc123").

MCP Tools (48 total)

Routing:

• llm_route — Route task to optimal model
• llm_classify — Classify task complexity
• llm_quality_guard — Monitor model health

Text:

• llm_query, llm_research, llm_generate, llm_analyze, llm_code

Media:

• llm_image, llm_video, llm_audio

Admin:

• llm_usage, llm_savings, llm_budget, llm_health, llm_providers

Advanced:

• llm_orchestrate — Multi-step pipelines
• llm_setup — Configure provider keys
• llm_policy — Routing policy management

Full tool reference — Complete documentation for all 48 tools

Architecture

See CLAUDE.md for:

• Design decisions
• Module organization
• Development workflow
• Release process

See docs/ARCHITECTURE.md for:

• Three-layer compression pipeline
• Judge scoring system
• Quality trend tracking
• Budget pressure algorithm

Development

uv run pytest tests/ -q          # Run tests
uv run ruff check src/ tests/    # Lint
uv run llm-router --version      # Check version

License

MIT — See LICENSE

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Releases: PyPI