claude-code-llm-router 6.3.1

Multi-LLM router MCP server for Claude Code — smart complexity routing, Claude subscription monitoring, Codex integration, 20+ providers

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, and one-command installs.

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

# One command to start saving
uvx claude-code-llm-router install

# Or: guided 5-minute setup
uvx claude-code-llm-router quickstart

Host	One-line install
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

---

LLM Router is an MCP server and hook set that intercepts prompts and routes them to the cheapest model that can handle the task.

It is built for a common failure mode in AI coding tools: using your best model for everything. In Claude Code, that burns quota on simple explanations, file lookups, small edits, and repetitive prompts. In other MCP clients, it means paying premium-model prices for work that never needed them.

The goal is simple: keep cheap work on cheap or free models, keep hard work on Claude or other premium models, and remove the need to micromanage model selection. Works in Claude Code, Cursor, VS Code, Codex, Windsurf, Zed, claw-code, and Agno.

---

Why

Most sessions contain a lot of low-value turns: quick questions, repo lookups, boilerplate edits, and small follow-ups. Those are exactly the prompts that quietly burn through premium models.

LLM Router offloads that work first, then escalates when the task actually needs more capability.

• Cheap work stays cheap.
• Hard work still gets the best model.
• Your workflow stays the same.

It does not try to replace Claude or force weak models onto hard tasks. It removes the waste around them.

---

Quick Start

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

llm-router install registers the MCP server and installs hooks so prompt routing starts automatically.

If you use Claude Code Pro/Max, you can start with zero API keys. Otherwise add GEMINI_API_KEY for a cheap free-tier fallback.

GEMINI_API_KEY=AIza...              # optional free-tier fallback
LLM_ROUTER_CLAUDE_SUBSCRIPTION=true

---

New in v6.3 — "Three-Layer Compression"

Theme: Every layer of the routing pipeline now compresses tokens — inputs, outputs, and LLM responses.

• Layer 1: RTK Command Output Compression — Bash/shell outputs automatically compressed via smart filters (60–90% reduction). Git, pytest, cargo, docker, npm outputs simplified to essentials
• Layer 2: Model Routing — Existing cost reduction via intelligent model selection (70–90%)
• Layer 3: Token-Savior Response Compression — LLM responses automatically condensed via 4-stage pipeline: filler removal → example consolidation → boilerplate collapse → semantic extraction (60–75% reduction)
• Unified Dashboard — llm_gain shows all three compression layers with token savings per layer
• Optional & Graceful — Compression is off by default; enable via LLM_ROUTER_COMPRESS_RESPONSE=true. Never blocks or crashes — falls back to original on any error
• Telemetry — All compression metrics logged to SQLite for analytics and trend tracking

Combined efficiency: 97x cost reduction vs baseline Opus across all three layers.

---

New in v6.1 — "Memory"

Theme: Your routing decisions are now stateful. The router learns from corrections and applies learned patterns automatically.

• Personal Routing Memory — After 3+ corrections for the same task type, learned routes become hard overrides in future sessions
• Session Memory Injection — See your learned patterns at session start with confidence levels
• Learned Route Application — Auto-route hook applies learned overrides before normal classification
• Community Profile Sharing — Export/import routing profiles via llm_share_profile() and llm_import_profile(url)
• Persistent Storage — Learned routes stored in ~/.llm-router/learned_routes.json across sessions

---

New in v6.0 — "Visible"

Theme: Every routing decision should be legible. v6.0 makes llm-router visible through CLI commands and HUD integration.

• Live Session HUD — Real-time routing indicators show which model handled each decision
• Session Replay — llm-router replay shows formatted transcript of all routed decisions with confidence, cost, and quality scores
• Health Checks — llm-router verify runs end-to-end system diagnostics (config, hooks, providers, recent decisions)
• Terminal Styling System — ANSI color codes, Unicode symbols, and accessible design tokens for routing flow visibility

---

New in v5.9.1

• Linting Cleanup: Removed unused imports and dead code from v5.9.0. All ruff checks now pass.

---

New in v5.9.0

• Caveman Mode — Token-efficient output compression (~75% reduction) with three intensity levels:
  • lite: Professional, minimal filler words
  • full: Standard compression (default)
  • ultra: Telegraphic, maximum brevity
  • Auto-injects system prompt when no user-provided message
  • Safe for all chat models (Claude, GPT-4, Gemini, Groq, etc.)
  • Control via LLM_ROUTER_CAVEMAN_INTENSITY env var
• Model Chain Visibility — Routing logs now show fallback chain (e.g., "claude → gemini → groq") for transparency
• Complexity Downgrade Tracking — New requested_complexity and complexity_downgraded columns in usage database for pressure-aware analytics
• Enforcement Guidance — Task-specific remediation advice in blocklist messages with cost justification

---

New in v5.8.0

• Pressure-Aware Routing Hooks: Hooks now monitor real-time subscription/API budget pressure and automatically downgrade task complexity when models near exhaustion. Prevents stuck sessions when Sonnet hits 100% quota.
• Hook Health Monitoring: New llm_health shows per-hook success/error counts, health status (healthy/degraded/failing), and recent error log. Enables diagnosing hook failures without restarting.
• Budget Burn-Rate Forecasting: forecast.py analyzes 7-day spend history with linear regression + exponential smoothing, projects monthly spend, and shows days-to-limit. Integrated into llm_usage dashboard and auto-route warning when forecast exceeds 90% of monthly budget.
• LLM-as-Judge Quality Evaluation: judge.py evaluates response quality (relevance/completeness/correctness) on 10% sample rate (configurable). Models with avg score < 0.7 over past 7 days are automatically demoted in the routing chain. Fire-and-forget background task — zero latency impact.
• Per-Task Daily Spend Caps: Enforce per-task-type daily limits via .llm-router.yml, independent of per-provider monthly caps. Prevents runaway costs on expensive task types (e.g., max $10/day code generation) while allowing other task types to continue.
• Quality-Based Model Reordering: Router automatically reorders model chain based on historical quality scores. Low-quality models move to the end until they improve, enabling continuous learning without manual intervention.

---

New in v5.5.0

• Security-Friendly YAML Configuration: Enterprises with .env restrictions can now store API keys in ~/.llm-router/config.yaml (permissions: 600). Priority order: .env (project) → config.yaml (user home) → env vars → defaults. Run llm-router init-claude-memory to auto-discover and generate the template.
• Auto-Discovery Initialization: New llm-router init-claude-memory command auto-detects which API keys and local models are configured, generates ~/.llm-router/config.yaml, and persists setup to Claude Code memory (~/.claude/projects/*/memory/) so future sessions don't repeat environment discovery.
• Claude Code Memory Integration: Persistent storage of environment configuration prevents repetitive "what API keys do you have?" questions across sessions. Memory survives session restarts and context compaction.

---

New in v5.3.0

• Sidecar routing service: llm-router now runs as independent FastAPI service (localhost:7337) for classification, eliminating blocking and enabling zero-deadlock operation with Serena, Obsidian, and all other MCP servers.
• Context-aware routing: high-confidence matches only (heuristic patterns) emit routing directives; medium/low confidence prompts allowed unconditionally. Infrastructure tools automatically skipped.
• Observation-only enforcement: enforce-route hook changed to logging-only mode (never blocks), preventing deadlocks and ensuring Serena/MCP tools always work.
• TOCTOU budget enforcement: concurrent routing calls reserve budget atomically before dispatch, preventing multiple calls from slipping under daily/monthly limits. Fixed via _pending_spend global with _budget_lock protection.
• Emergency fallback chain: when the primary profile exhausts all capable models, automatically escalate to BUDGET chain to prevent complete routing failure.
• Dashboard authentication: token-based auth middleware for the optional dashboard UI, with secure token generation and storage.
• Correlation ID tracking: every routing call gets a unique 8-char ID for distributed tracing; logged in both usage and routing_decisions tables.
• Async/event-loop safety: eliminated synchronous filesystem I/O blocking the event loop in budget checks and Claude subscription state reads via asyncio.to_thread().
• Route refactoring: extracted _dispatch_model_loop() to reduce route_and_call() from 960 to 527 lines; added helper functions _resolve_profile() and _build_and_filter_chain().

---

How It Works

1. Intercept the prompt before your default premium model sees it.
2. Classify the task and its complexity.
3. Try the cheapest capable route first.
4. Escalate or fall back when the task needs more capability.

Under the hood, every prompt goes through a UserPromptSubmit hook before your top-tier model sees it:

0. Context inherit      instant, free    "yes/ok/go ahead" reuse prior turn's route
1. Heuristic scoring    instant, free    high-confidence patterns route immediately
2. Ollama local LLM     free, ~1s        catches what heuristics miss
3. Cheap API            ~$0.0001         Gemini Flash / GPT-4o-mini fallback

Prompt	Classified as	Routed to
"What does os.path.join do?"	query/simple	Gemini Flash ($0.000001)
"Fix the bug in auth.py"	code/moderate	Haiku / Sonnet
"Design the full auth system"	code/complex	Sonnet / Opus
"Research latest AI funding"	research	Perplexity Sonar Pro
"Generate a hero image"	image	Flux Pro via fal.ai

Free-first chain (subscription mode): Ollama → Codex (free via OpenAI sub) → paid API

---

CLI Commands (v6.0+)

See What Model Just Handled Your Request

llm-router last                      # Show your last 5 routing decisions
llm-router last --count 10           # Show last 10 decisions

Real-time feedback on every routed decision:

══════════════════════════════════════════════════════════
  5 Most Recent Routing Decisions
══════════════════════════════════════════════════════════

just now   → routed to ollama/gemma4:latest (query/simple) - FREE
2 min ago  → routed to openai/gpt-4o-mini (code/simple) - $0.0010
5 min ago  → routed to openai/gpt-4o (analysis/moderate) - $0.0062

Answer: "What model handled my last question?" → Run llm-router last

---

Session Replay — Review all routing decisions

llm-router replay                    # Show all decisions this session
llm-router replay --limit 20         # Show last 20 decisions only
llm-router replay --session SESSION_ID

Example output:

═══════════════════════════════════════════════════════════
  SESSION REPLAY — 14:30–15:45
═══════════════════════════════════════════════════════════

14:30 → routed to haiku (code/simple)
    ★ Confidence: ★★★★★★★★☆☆ 87%
    🧠 Reasoning: Simple code generation, low risk
    💰 Cost: $0.0001
    ✅ Quality: 97%

14:31 → routed to sonnet (analysis/moderate)
    ★ Confidence: ★★★★★★★★★☆ 92%
    🧠 Reasoning: Architecture analysis requires deeper reasoning
    💰 Cost: $0.0062
    ✅ Quality: 98%

───────────────────────────────────────────────────────────
SUMMARY
───────────────────────────────────────────────────────────
  Total routed: 12 calls
  Cost: $0.186
  Saved: $1.847 (90%)

Health Check — Verify system is healthy

llm-router verify

Example output:

══════════════════════════════════════════════════════════
              llm-router health check
══════════════════════════════════════════════════════════

Configuration
──────────────────────────────────────────────────────────
✅ Configuration loaded from ~/.llm-router/config.yaml
✅ SQLite database: ~/.llm-router/usage.db (45 MB, last write 5 mins ago)

Providers
──────────────────────────────────────────────────────────
✅ Ollama (http://localhost:11434) — 2 models: gemma4:latest | qwen3.5:latest
✅ OpenAI API configured
✅ Gemini API configured

Hooks
──────────────────────────────────────────────────────────
✅ llm-router-auto-route.py (installed + executable)
✅ llm-router-session-end.py (installed + executable)
✅ llm-router-enforce-route.py (installed + executable)

Recent Decisions
──────────────────────────────────────────────────────────
  2 min ago  → haiku (code, simple)      $0.0001
  5 min ago  → sonnet (analysis, mod)    $0.0062
  12 min ago → opus (planning, complex)  $0.062

──────────────────────────────────────────────────────────
✅ No issues detected. You're good! 🚀

Summary of v6.0 Visibility

The three CLI commands make every routing decision visible:

- `→` — routing arrow
- `haiku` — selected model (color: Orchestrate Blue)
- `[87%]` — confidence score
- `(code/simple)` — task type and complexity
- `$0.001` — estimated cost (color: Confidence Green)
- `⚡` — performance indicator

---

### Why teams keep it on

- It cuts premium-model waste without forcing people to think about models on every turn.
- It keeps a real paper trail: routing decisions, spend, budget pressure, and quality signals all stay visible.
- It degrades gracefully: stale hooks, unhealthy providers, exhausted budgets, and partial integrations fall through to safe defaults instead of deadlocking a session.
- It scales from solo use to team rollout with the same primitives: budgets, policies, dashboards, digests, and host-specific installs.

---

## MCP Tools

48 tools across 7 categories:

### Smart Routing
| Tool | What it does |
|------|-------------|
| `llm_route` | Auto-classify prompt → route to best model |
| `llm_auto` | Route + server-side savings tracking — designed for hook-less hosts (Codex CLI, Claude Desktop, Copilot) |
| `llm_classify` | Classify complexity + recommend model |
| `llm_select_agent` | Pick agent CLI (claude_code / codex) + model for a session |
| `llm_stream` | Stream LLM response for long-running tasks |
| `llm_reroute` | Correct a bad routing decision in-session and train the router |

### Text & Code
| Tool | What it does |
|------|-------------|
| `llm_query` | General questions — routed to cheapest capable model |
| `llm_research` | Web-grounded answers via Perplexity Sonar |
| `llm_generate` | Creative writing, summaries, brainstorming |
| `llm_analyze` | Deep reasoning — analysis, debugging, design review |
| `llm_code` | Code generation, refactoring, algorithms |
| `llm_edit` | Route edit reasoning to cheap model → returns `{file, old, new}` patch pairs |

### Filesystem
| Tool | What it does |
|------|-------------|
| `llm_fs_find` | Describe files to find → cheap model returns glob/grep commands |
| `llm_fs_rename` | Describe a rename → returns `mv`/`git mv` commands (dry_run by default) |
| `llm_fs_edit_many` | Bulk edits across files → returns all patch pairs |
| `llm_fs_analyze_context` | Summarise workspace context for smarter routing |

### Media
| Tool | What it does |
|------|-------------|
| `llm_image` | Image generation — Flux, DALL-E, Gemini Imagen |
| `llm_video` | Video generation — Runway, Kling, Veo 2 |
| `llm_audio` | TTS/voice — ElevenLabs, OpenAI |

### Orchestration
| Tool | What it does |
|------|-------------|
| `llm_orchestrate` | Multi-step pipeline across multiple models |
| `llm_pipeline_templates` | List available pipeline templates |

### Monitoring & Admin
| Tool | What it does |
|------|-------------|
| `llm_usage` | Unified dashboard — Claude sub, Codex, APIs, savings, **burn-rate forecast** |
| `llm_savings` | Cross-session savings breakdown by period, host, and task type |
| `llm_check_usage` | Live Claude subscription usage (session %, weekly %) |
| `llm_health` | Provider availability + hook health status (success/error counts, recent errors) |
| `llm_providers` | List all configured providers and models |
| `llm_set_profile` | Switch profile: `budget` / `balanced` / `premium` |
| `llm_setup` | Interactive provider wizard — add keys, validate, install hooks |
| `llm_quality_report` | Routing accuracy, judge quality scores, classifier stats |
| `llm_rate` | Rate last response 👍/👎 — logged for quality tracking |
| `llm_codex` | Route task to local Codex desktop agent (free) |
| `llm_save_session` | Persist session summary for cross-session context |
| `llm_cache_stats` | Cache hit rate, entries, evictions |
| `llm_cache_clear` | Clear classification cache |
| `llm_refresh_claude_usage` | Force-refresh subscription data via OAuth |
| `llm_update_usage` | Feed usage data from claude.ai into the router |
| `llm_track_usage` | Report Claude Code token usage for budget tracking |
| `llm_dashboard` | Open web dashboard at localhost:7337 |
| `llm_team_report` | Team-wide routing savings report |
| `llm_team_push` | Push local savings data to shared team store |
| `llm_policy` | Show active org/repo routing policy + last 10 policy decisions |
| `llm_digest` | Savings digest with spend-spike detection; push to Slack/Discord webhook |
| `llm_benchmark` | Per-task-type routing accuracy from `llm_rate` feedback |
| `llm_session_spend` | Real-time API spend breakdown for the current session |
| `llm_approve_route` | Approve or reject a pending high-cost routing call |
| `llm_budget` | Budget Oracle — real-time spend vs. cap per provider with pressure bars |

### Community Sharing (v6.1+)
| Tool | What it does |
|------|-------------|
| `llm_share_profile` | Export your learned routing profile for community sharing |
| `llm_import_profile` | Import and merge shared routing profiles from URLs |

---

## Routing Profiles

Three profiles — switch anytime with `llm_set_profile`:

| Profile | Use case | Chain |
|---------|----------|-------|
| `budget` | Dev, drafts, exploration | Ollama → Haiku → Gemini Flash |
| `balanced` | Production work *(default)* | Codex → Sonnet → GPT-4o |
| `premium` | Critical tasks, max quality | Codex → Opus → o3 |

Profile is overridden by complexity: simple prompts always use the budget chain, complex ones escalate to premium, regardless of the active profile setting.

---

## Providers

| Provider | Models | Free tier | Best for |
|----------|--------|-----------|----------|
| **Ollama** | Any local model | Yes (forever) | Privacy, zero cost, offline |
| **Google Gemini** | 2.5 Flash, 2.5 Pro | Yes (1M tokens/day) | Generation, long context |
| **Groq** | Llama 3.3, Mixtral | Yes | Ultra-fast inference |
| **OpenAI** | GPT-4o, o3, DALL-E | No | Code, reasoning, images |
| **Perplexity** | Sonar, Sonar Pro | No | Research, current events |
| **Anthropic** | Haiku, Sonnet, Opus | No | Writing, analysis, safety |
| **DeepSeek** | V3, Reasoner | Limited | Cost-effective reasoning |
| **Mistral** | Large, Small | Limited | Multilingual |
| **fal.ai** | Flux, Kling, Veo | No | Images, video, audio |
| **ElevenLabs** | Voice models | Limited | High-quality TTS |
| **Runway** | Gen-3 | No | Professional video |

Full setup guides: [docs/PROVIDERS.md](docs/PROVIDERS.md)

---

## Works With

### Claude Code

Auto-installed by `llm-router install`. Hooks intercept every prompt — you never need to call tools manually unless you want explicit control.

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

Live status bar shows routing stats before every prompt and in the persistent bottom statusline:

📊  CC 13%s · 24%w  │  sub:0 · free:305 · paid:27  │  $1.59 saved (35%)

claw-code

Add to ~/.claw-code/mcp.json:

{
  "mcpServers": {
    "llm-router": { "command": "llm-router", "args": [] }
  }
}

Every API call in claw-code is paid — the free-first chain (Ollama → Codex → Gemini Flash) saves more here than in Claude Code.

Cursor / Windsurf / Zed

Add to your IDE's MCP config:

{
  "mcpServers": {
    "llm-router": { "command": "llm-router", "args": [] }
  }
}

Agno (multi-agent)

Two integration modes:

Option 1 — RouteredModel (v2.0+): use llm-router as a first-class Agno model. Every agent call is automatically routed to the cheapest capable provider.

pip install "claude-code-llm-router[agno]"

from agno.agent import Agent
from llm_router.integrations.agno import RouteredModel, RouteredTeam

# Single agent — routes each call intelligently
coder = Agent(
    model=RouteredModel(task_type="code", profile="balanced"),
    instructions="You are a coding assistant.",
)
coder.print_response("Write a Python quicksort.")

# Multi-agent team with shared $20/month budget cap
# Automatically downshifts to 'budget' profile at 80% spend
team = RouteredTeam(
    members=[coder, researcher],
    monthly_budget_usd=20.0,
    downshift_at=0.80,
)

Option 2 — MCP tools: use llm-router's 45 tools in any Agno agent:

from agno.agent import Agent
from agno.models.anthropic import Claude
from agno.tools.mcp import MCPTools

agent = Agent(
    model=Claude(id="claude-sonnet-4-6"),
    tools=[MCPTools(command="llm-router")],
    instructions="Use llm_research for web searches, llm_code for coding tasks.",
)

Supported Hosts

Host	Install command	Writes files	Hook support
Claude Code	llm-router install	✅	✅ Full auto-route
Codex CLI	llm-router install --host codex	✅	✅ PostToolUse
OpenCode	llm-router install --host opencode	✅	✅ PostToolUse
Gemini CLI	llm-router install --host gemini-cli	✅	✅ Extension hook
GitHub Copilot CLI	llm-router install --host copilot-cli	✅	—
OpenClaw	llm-router install --host openclaw	✅	—
Trae IDE	llm-router install --host trae	✅	—
Factory Droid	llm-router install --host factory	✅ manifest	— (Claude Code compat)
VS Code (MCP native)	llm-router install --host vscode	✅	—
Cursor IDE	llm-router install --host cursor	✅	—
Claude Desktop	llm-router install --host desktop	snippet	—
GitHub Copilot (VS Code)	llm-router install --host copilot	snippet	—

All installs are idempotent — run any command twice safely.

Codex CLI

llm-router install --host codex

Writes ~/.codex/config.yaml, ~/.codex/hooks.json (PostToolUse), and ~/.codex/instructions.md.

codex plugin install llm-router   # or via Codex marketplace

OpenCode

llm-router install --host opencode

Writes ~/.config/opencode/config.json (MCP block), PostToolUse hook, and routing rules.

Gemini CLI

llm-router install --host gemini-cli

Writes ~/.gemini/settings.json, creates the llm-router extension with gemini-extension.json + hooks.json, and appends routing rules.

GitHub Copilot CLI

llm-router install --host copilot-cli

Writes ~/.config/gh/copilot/mcp.json and routing rules.

OpenClaw

llm-router install --host openclaw

Writes ~/.openclaw/mcp.json and routing rules.

Trae IDE

llm-router install --host trae

Writes the platform-appropriate Trae config (~/Library/Application Support/Trae/mcp.json on macOS) and a .rules file in the current directory.

Factory Droid

Factory Droid natively supports Claude Code plugin format (.claude-plugin/) — no extra setup needed:

factory plugin install ypollak2/llm-router
# or via Factory marketplace search: llm-router

The dedicated .factory-plugin/ manifest is included for Factory marketplace discovery.

Claude Desktop

llm-router install --host desktop

Prints the snippet for claude_desktop_config.json. No hooks in Desktop — use llm_auto for savings tracking.

GitHub Copilot (VS Code)

llm-router install --host copilot

Prints the snippet for .vscode/mcp.json and a copilot-instructions.md template.

All at once

llm-router install --host all   # installs/prints all hosts

Docker / CI

RUN pip install claude-code-llm-router && llm-router install --headless
# Pass keys at runtime: docker run -e GEMINI_API_KEY=... your-image

---

Configuration

# API keys — at least one required
GEMINI_API_KEY=AIza...              # free tier at aistudio.google.com
OPENAI_API_KEY=sk-proj-...
PERPLEXITY_API_KEY=pplx-...
ANTHROPIC_API_KEY=sk-ant-...        # skip if using Claude Code subscription
DEEPSEEK_API_KEY=...
GROQ_API_KEY=gsk_...
FAL_KEY=...                         # images, video, audio via fal.ai
ELEVENLABS_API_KEY=...

# Router
LLM_ROUTER_PROFILE=balanced         # budget | balanced | premium
LLM_ROUTER_MONTHLY_BUDGET=0         # USD, 0 = unlimited
LLM_ROUTER_CLAUDE_SUBSCRIPTION=false  # true = Claude Code Pro/Max
LLM_ROUTER_ENFORCE=enforce          # shadow | suggest | enforce (default: enforce)

# Compression
LLM_ROUTER_COMPRESS_RESPONSE=false  # true = enable Token-Savior response compression
LLM_ROUTER_COMPRESS_TARGET=0.6      # target compression ratio (0.0–1.0, default 0.6 = 60%)

# Ollama (local models)
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_BUDGET_MODELS=gemma4:latest,qwen3.5:latest

# Spend limits
LLM_ROUTER_DAILY_SPEND_LIMIT=5.00   # USD, 0 = disabled

# Per-provider monthly budget caps (overridden by llm-router budget set)
LLM_ROUTER_BUDGET_OPENAI=20.0
LLM_ROUTER_BUDGET_GEMINI=5.0
LLM_ROUTER_BUDGET_GROQ=3.0
LLM_ROUTER_BUDGET_DEEPSEEK=3.0

# Quality evaluation (LLM-as-Judge)
LLM_ROUTER_JUDGE_SAMPLE_RATE=0.1    # 10% of calls, 0.0 = disabled

# Enterprise integrations
HELICONE_API_KEY=sk-helicone-...         # enables Helicone routing properties
LLM_ROUTER_HELICONE_PULL=false           # pull spend from Helicone API
LLM_ROUTER_LITELLM_BUDGET_DB=/path/to/litellm.db  # LiteLLM Proxy DB
LLM_ROUTER_LITELLM_USER=my-team          # optional LiteLLM user/team filter
LLM_ROUTER_SPEND_AGGREGATION=max         # max | sum when sources are independent

Enforcement Modes

Choose how strict routing should be. The easiest way is llm-router onboard, which lets you pick a mode interactively.

Mode	Behaviour	Best for
shadow	Observe routing decisions, never blocks	Safest first install
suggest	Show route hints, allow direct answers	Low-friction adoption
enforce	Block routed violations until the route is followed	Maximum savings

LLM_ROUTER_ENFORCE=hard is the strict compatibility alias for enforce. Legacy soft and off values are still supported for direct CLI or env-based control.

Repo-level config (.llm-router.yml)

Commit a routing policy alongside your code — no env vars required:

profile: balanced
enforce: suggest          # shadow | suggest | enforce

block_providers:
  - openai                # never use OpenAI in this repo

routing:
  code:
    model: ollama/qwen3.5:latest   # always use local model for code tasks
  research:
    provider: perplexity           # always use Perplexity for research

task_caps:
  code: 10.00            # Max $10/day for code generation
  research: 5.00         # Max $5/day for research tasks
  analyze: 20.00         # Max $20/day for analysis

daily_caps:
  _total: 2.00           # global $2/day cap

User-level overrides live in ~/.llm-router/routing.yaml (same schema). Repo config wins.

Full reference: .env.example

---

Budget Control

Per-Provider Caps (CLI)

Set monthly spend caps per provider — persisted to ~/.llm-router/budgets.json:

llm-router budget list                  # show all providers with cap, spend, pressure
llm-router budget set openai 20         # $20/month cap for OpenAI
llm-router budget set gemini 5          # $5/month cap for Gemini
llm-router budget remove openai         # remove cap (reverts to env-var or unlimited)

Caps set via CLI take priority over LLM_ROUTER_BUDGET_* env vars. The router automatically routes away from providers approaching their caps.

Via Env Vars

LLM_ROUTER_BUDGET_OPENAI=20.0          # $20/month cap for OpenAI
LLM_ROUTER_BUDGET_GEMINI=5.0           # $5/month cap for Gemini
LLM_ROUTER_BUDGET_GROQ=3.0
LLM_ROUTER_BUDGET_DEEPSEEK=3.0
LLM_ROUTER_BUDGET_TOGETHER=5.0
LLM_ROUTER_BUDGET_PERPLEXITY=10.0
LLM_ROUTER_BUDGET_MISTRAL=5.0
LLM_ROUTER_MONTHLY_BUDGET=50           # global cap across all providers

llm_usage("month")
→ Calls: 142 | Tokens: 320k | Cost: $3.42 | Budget: 6.8% of $50

The router tracks spend in SQLite across all providers and routes away from providers approaching their monthly caps.

Per-Task Daily Caps (Policy-Based)

Enforce spending limits on specific task types within a single day:

# ~/.llm-router/routing.yaml
task_caps:
  code: 10.00         # Max $10/day for code generation
  research: 5.00      # Max $5/day for research tasks
  analyze: 20.00      # Max $20/day for analysis
  query: 2.00         # Max $2/day for simple queries

When a task type's daily spend exceeds its cap, the router raises BudgetExceededError and prevents further calls for that task type until the next day (UTC).

This prevents runaway costs for expensive task types while allowing other task types to continue working.

Burn-Rate Forecasting

The router analyzes 7-day spend history to forecast monthly spend and project days until budget exhaustion:

📈 Burn-Rate Forecast
Current pace: $12.50/month (150 calls this week)
Days to limit: 8 days remaining until $20.00 budget exhausted
Confidence: 85% (7 days of data)

Forecast appears in llm_usage("month") dashboard and auto-route warns when projected spend exceeds 90% of monthly budget.

Quality-Based Model Reordering

The router tracks quality feedback from evaluations and automatically deprioritizes models with consistently low quality:

Judge Evaluation Scores (0–1):
  relevance:    "Does response address the prompt?"
  completeness: "Is response sufficiently thorough?"
  correctness:  "Is content factually accurate?"

Models with average judge score < 0.7 over the past 7 days are moved to the end of the model chain. This allows the router to automatically learn from quality trends and avoid routing to models that produce poor outputs.

Sample rate (default 10%):

export LLM_ROUTER_JUDGE_SAMPLE_RATE=0.1    # Score 10% of calls
export LLM_ROUTER_JUDGE_SAMPLE_RATE=1.0    # Score every call (more expensive)

Quality tracking is opt-in and fire-and-forget — evaluations run asynchronously without blocking the primary task.

---

Hook Health Monitoring

Monitor the health of auto-route and enforce-route hooks:

llm_health
→ auto-route: healthy (1,247 successes, 0 errors)
→ enforce-route: healthy (1,247 successes, 3 errors in last 2 days)

Hook errors are logged to ~/.llm-router/hook_errors.log with context (session_id, task_type). View recent errors:

llm_health  # shows per-hook success/error counts + health status (healthy/degraded/failing)

This enables diagnosing hook failures without restarting Claude Code.

---

Dashboard

llm-router dashboard   # opens localhost:7337

Live view of routing decisions, cost trends, model distribution, subscription pressure, and burn-rate forecast. Auto-refreshes every 30s.

Tabs: Overview · Performance · Config · Logs · 💰 Budget · 📈 Forecast

The dashboard now starts with a tokenized local URL and protects API calls with that token, so local metrics stay private by default.

The Budget tab shows per-provider spend vs. cap with editable cap inputs. Changes are persisted to ~/.llm-router/budgets.json and take effect immediately.

Prometheus metrics

GET http://localhost:7337/metrics

Exposes standard text exposition format, no extra dependencies:

# HELP llm_router_spend_usd Monthly spend per provider in USD
llm_router_spend_usd{provider="openai"} 3.42
llm_router_budget_pressure{provider="openai"} 0.17
llm_router_budget_cap_usd{provider="openai"} 20.0
llm_router_savings_usd_total 12.87

OpenTelemetry tracing

Install the optional tracing extra when you want route-level spans in Tempo, Jaeger, Honeycomb, or any OTLP-compatible backend:

pip install "claude-code-llm-router[tracing]"
OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317
OTEL_SERVICE_NAME=llm-router

Spans cover the routing hot path:

• classify_complexity
• score_all_models
• build_chain
• route_and_call
• provider_call

---

Enterprise Integrations

Helicone

Add observability headers to every routed call and optionally pull spend data back from the Helicone API:

HELICONE_API_KEY=sk-helicone-...      # enables routing properties in every call
LLM_ROUTER_HELICONE_PULL=true         # pull spend per provider from Helicone API

When LLM_ROUTER_HELICONE_PULL=true, Helicone spend is merged with local SQLite spend and LiteLLM spend — budget pressure uses the maximum across all sources.

Headers injected on every call:

Helicone-Auth: Bearer sk-helicone-...
Helicone-Property-Router-Task-Type: code
Helicone-Property-Router-Model: openai/gpt-4o
Helicone-Property-Router-Complexity: moderate
Helicone-Property-Router-Profile: balanced

LiteLLM BudgetManager

Teams running a LiteLLM Proxy with budget management can point llm-router at the proxy's SQLite DB:

LLM_ROUTER_LITELLM_BUDGET_DB=/path/to/litellm.db
LLM_ROUTER_LITELLM_USER=my-team        # optional — filter by user/team key

The router reads spend_logs aggregated by provider (e.g. openai/gpt-4o → openai) and incorporates LiteLLM spend into pressure calculations. This ensures routing pressure reflects traffic flowing through both direct llm-router calls and the LiteLLM proxy simultaneously.

Budget cap lookup from budget_limits table (LiteLLM Proxy ≥ 1.30) is also supported.

If Helicone and LiteLLM are tracking genuinely separate traffic streams, set:

LLM_ROUTER_SPEND_AGGREGATION=sum

The default max mode is safer when multiple systems may be observing the same requests.

---

Session Summary

At session end the router prints a breakdown:

  Free models  305 calls  ·  $0.52 saved  (Ollama / Codex)
  External       27 calls  ·  $0.006       (Gemini Flash, GPT-4o)
  💡 Saved ~$0.53 this session

Share your savings:

llm-router share   # copies savings card to clipboard + opens tweet

---

Roadmap

Positioning: Route every AI coding task to the cheapest capable model. Works across Claude Code, Cursor, VS Code, Codex, Gemini CLI, and more.

Phase 1 — Foundation ✅ Complete

Version	Headline
v1.3–v2.0	Foundation, dashboard, enforcement, Agno adapter
v2.1	Route Simulator — llm-router test "<prompt>" dry-run + llm_savings dashboard
v2.2	Explainable Routing — LLM_ROUTER_EXPLAIN=1, "why not Opus?", per-decision reasoning
v2.3	Zero-Friction Activation — onboarding wizard, shadow/suggest/enforce modes, yearly savings projection

Phase 2 — Smarter Routing ✅ Complete

Version	Headline
v2.4	Repo-Aware YAML Config — .llm-router.yml committed with the codebase, block_providers, model pins
v2.5	Context-Aware Routing — "yes/ok/go ahead" inherits prior turn's route, zero classifier latency
v2.6	Latency + Personalized Routing — p95 latency scoring, per-user acceptance signals

Phase 3 — Team Infrastructure ✅ Complete

Version	Headline
v3.0	Team Dashboard — shared savings across the whole team
v3.1	Multi-Host + Cross-Session Savings — llm_auto, Codex/Desktop/Copilot adapters, persistent savings across sessions
v3.2	Policy Engine — org/project/user routing policy, spend caps, audit log
v3.3	Slack Digests + Codex Plugin — weekly savings digest, spend-spike alerts, Codex marketplace plugin

Phase 4 — Multi-Host Expansion ✅ Complete

Version	Headline
v3.4	Agent-Context Routing — subscription-first chain reordering when Codex or Claude Code is active
v3.5	Multi-Agent CLI Compatibility — OpenCode, Gemini CLI, Copilot CLI, OpenClaw, Factory Droid, Trae
v3.6	VS Code + Cursor IDE Support — native MCP config, routing rules, idempotent install
v4.0	Token Efficiency + Real-Time Spend — tool slim mode, session spend meter, reroute learning, quickstart wizard
v4.1	Playwright DOM Compression + Enforcement Fixes — DOM compression hook, PostToolUse MCP matcher fix, smart enforcement default
v4.2	Quota-Aware Routing + Context-Aware Classification — Ollama-first CC-mode for simple tasks, qwen3.5:32b in BALANCED chains, short code follow-up context inheritance

Phase 5 — Reliability + Observability ✅ Complete

Version	Headline
v5.0	Adaptive Universal Router — budget oracle, model discovery, live benchmarks, scorer, dynamic chain builder
v5.1	Budget Management UX + Enterprise Integrations — persistent caps, dashboard budget tab, Prometheus metrics, Helicone + LiteLLM spend
v5.2	Audit Remediation Release — atomic hooks, stricter CI, structlog, OpenTelemetry spans, classifier evals, release automation
v5.3	Sidecar Service + Zero-Deadlock Hooks — independent FastAPI service, context-aware routing, observation-only enforcement
v5.4	Config Security + Claude Code Memory — YAML config fallback, auto-discovery, memory persistence
v5.5	Security-Friendly YAML Configuration — enterprise .env bypass, auto-discovery init, Claude Code memory integration
v5.6	Hook Enforcement Refactoring + Stuck Pattern Prevention — 6-point hook health system, prevent routing deadlocks, atomic budget checks
v5.7	Advanced Budget Control + Quality Scoring — per-task daily caps, quality-based reordering, judge sample rate
v5.8	Pressure-Aware Hooks + LLM-as-Judge — hooks monitor real-time pressure, auto-downgrade complexity, burn-rate forecasting, quality evaluation

Phase 6 — Visibility & Learning (v6.0–v7.0) 🎯 In Progress

The next 6 months focus on making routing decisions visible, learnable, and shareable.

Version	Headline	Status
v6.0	"Visible" — Live routing HUD, session replay, health checks, design system	🚀 May 2026
v6.1	"Memory" — Personal routing profiles, override learning, community sharing	🚀 Apr 2026
v6.3	"Three-Layer Compression" — RTK input compression, Token-Savior response compression, unified dashboard	🚀 Apr 2026
v6.2	"Quality" — Quality Guard, benchmarks, degradation alerts, accuracy reports	📅 Jul 2026
v6.4	"Local First" — Ollama dashboard, model discovery, auto-recommendations	📅 Aug 2026
v6.5	"Community" — Savings card, README badge, routing config marketplace	📅 Sep 2026
v7.0	"Platform" — Public routing API, agent chains, plugin SDK, marketplace	📅 Oct 2026

Details: docs/ROADMAP_v6.md — competitive positioning, design decisions, monthly product cycle framework.

---

Development

uv sync --extra dev
uv run pytest tests/ -q --ignore=tests/test_agno_integration.py
uv run ruff check src/ tests/
uv run python scripts/eval_classifier.py --limit 20
uv run python scripts/release.py 5.8.0 --dry-run

See CLAUDE.md for architecture and module layout.

---

Contributing

See CONTRIBUTING.md. Key areas: new provider integrations, routing intelligence, MCP client testing.

---

License

MIT

---

_{Built with LiteLLM and MCP}