sr-router

sr-router

Pick the right brain for every job. Stop burning API credits.

---

sr-router intelligently routes LLM requests to the cheapest model that meets quality requirements. Use it as a transparent HTTP proxy (drop-in for the Anthropic API) or as an MCP server.

Quick Start

# Clone and build
git clone https://github.com/jbctechsolutions/sr-router.git
cd sr-router
go build -o sr-router ./cmd/

# Set at least the Anthropic key
export ANTHROPIC_API_KEY=sk-ant-...

# Route a prompt (local only, no API call)
./sr-router route "Write a Go function for rate limiting"

# Start the proxy (point Claude Code at it)
./sr-router proxy --port 8889

For the full walkthrough -- API key setup, Ollama, MCP integration, configuration, and troubleshooting -- see the Setup Guide.

How It Works

                                        +-------------------+
                                        |    Provider APIs   |
                                        |  (Anthropic, etc.) |
                                        +---------+---------+
                                                  ^
                                                  |
Request --> Classify --> Score Models --> Select --> Failover Chain --> Provider --> Response
            |               |              |
            v               v              v
       Route Class     Weighted        Best model
       + Task Type     formula         for the job

sr-router uses a two-layer classification system to pick the right model:

1. Route class -- Is this interactive, background, or compaction? Determined from HTTP headers, content patterns, CLI flags, and environment variables. Each class has its own tier, latency budget, and quality floor.

2. Task type -- What kind of work is this? Code generation, architecture design, summarization, data extraction, etc. Detected via regex pattern matching against the prompt. Each task type specifies required model strengths and a minimum quality threshold.

3. Model scoring -- Every eligible model is scored with a weighted formula:

   score = (cost_weight * cost_score) + (quality_weight * quality_score)
   

   Default weights: cost 40%, quality 60%. Models are filtered by tier membership and required strengths before scoring.

4. Failover -- If the primary model fails (429 rate limit, 5xx server error, or timeout), sr-router cascades to the next model in the tier's failover chain automatically.

Modes

Transparent Proxy

Drop-in replacement for the Anthropic API. Point any client that speaks the Anthropic HTTP protocol at sr-router and it will classify, route, and proxy the request transparently -- including SSE streaming.

sr-router proxy --port 8889

# Then point Claude Code at it:
ANTHROPIC_API_BASE=http://localhost:8889 claude

MCP Server

Run sr-router as an MCP server over stdio for use with Claude Code, Cursor, or any MCP-compatible client. Exposes route, classify, models, and stats as MCP tools.

sr-router mcp

Cost Savings

Task	Without sr-router	With sr-router	Savings
Architecture review	claude-opus ($0.075/1k)	claude-opus ($0.075/1k)	0% (quality required)
Code generation	claude-opus ($0.075/1k)	claude-sonnet ($0.015/1k)	80%
Batch summarization	claude-opus ($0.075/1k)	minimax-m2 ($0.0003/1k)	99.6%
Context compaction	claude-opus ($0.075/1k)	cerebras-glm ($0.0006/1k)	99.2%
Simple code / tests	claude-opus ($0.075/1k)	ollama/codellama ($0.00/1k)	100%

The router never downgrades quality for tasks that need it -- architecture reviews still go to Opus. But summarization, boilerplate, and batch work get routed to models that cost a fraction of a cent.

CLI Reference

Command	Description	Example
route <prompt>	Classify and route a prompt to the best model	sr-router route "Write a Go function for rate limiting"
classify <prompt>	Classify a prompt without routing	sr-router classify "Summarize this document"
models	List all configured models	sr-router models --tier premium
proxy	Start the transparent HTTP proxy	sr-router proxy --port 8889
mcp	Start the MCP server (stdio)	sr-router mcp
stats	Show routing statistics from telemetry	sr-router stats --model claude-sonnet
feedback <id>	Record feedback for a routing event	sr-router feedback abc123 --rating 5
config validate	Validate YAML configuration files	sr-router config validate
config init	Show the resolved config directory	sr-router config init

Global Flags

Flag	Description
--config <dir>	Override the config directory (default: ./config, then ~/.config/sr-router/config)

Route Flags