Claude Code Router
65% Claude Code cost reduction. Zero workflow changes. Deployed over lunch.
The Problem
An AI developer tools startup had standardised their engineering workflow around Claude Code. Fifteen engineers using it daily. The results were good. The bill was not.
$40,000 per month. Nearly all of it on Claude Sonnet — used for everything from quick autocomplete to complex architectural reasoning to running background lint checks. The team knew they were over-spending but didn’t want to change their workflow: Claude Code had become how they worked.
The constraint was simple: whatever the solution, engineers should not change a single keystroke.
What We Built
A transparent proxy server that sits between Claude Code and Anthropic’s API. Engineers point their ANTHROPIC_BASE_URL at http://localhost:3456 and forget about it. The proxy reads each request, classifies the task, routes to the cheapest capable model, and returns a response in the same format Claude Code expects.
No UI. No config file for engineers to manage. No new tools to learn.
How It Works
The Fastify server intercepts all requests on :3456. The routing logic reads the system prompt and messages array from each request to classify the task:
interface RouterConfig { rules: RoutingRule[]; providers: Record<string, ProviderConfig>; } type TaskClass = "background" | "think" | "longContext" | "default"; function classifyRequest(body: AnthropicRequest): TaskClass { const system = body.system ?? ""; const lastMessage = body.messages.at(-1)?.content ?? ""; const totalTokens = estimateTokens(body.messages); // Background tasks: no reasoning needed, use local model if (/lint|format|type.?check|import|whitespace/i.test(system)) { return "background"; } // Explicit reasoning requests if (/think|reason|explain|architect|design|tradeoff/i.test(lastMessage)) { return "think"; } // Large context window needed (>60k tokens estimated) if (totalTokens > 60_000) { return "longContext"; } return "default"; } const ROUTES: Record<TaskClass, string> = { background: "ollama/qwen2.5-coder:7b", // Free, local think: "deepseek/deepseek-r1", // Strong reasoning, cheap longContext: "google/gemini-2.5-pro", // 1M token window default: "anthropic/claude-sonnet-4-5", // Baseline for everything else }; Each provider has a custom transformer that converts the Anthropic request format to the target provider’s API format and normalises the response back:
class OllamaTransformer implements ProviderTransformer { async transform(req: AnthropicRequest): Promise<OllamaRequest> { return { model: req.model.replace("ollama/", ""), messages: req.messages, stream: req.stream ?? false, options: { num_ctx: 32768 }, }; } async normaliseResponse(res: OllamaResponse): Promise<AnthropicResponse> { /* ... format back to Anthropic shape */ } } The Outcome
Monthly Claude Code spend dropped from $40k to ~$14k in the first month. Background routing to local Ollama (running on a Mac Mini in the office) accounted for 40% of the saving alone. Engineers reported zero perceptible quality difference on their daily tasks.
The proxy config now has a staging vs. production mode: in staging, expensive requests are routed to cheaper models by default; in production, the full Anthropic model is used for anything user-facing. This alone reduced their staging environment costs to near-zero.