anthropic-gateway 1.0.0 → 1.1.0

package/lib/server.js

@@ -57,6 +57,18 @@ function makeServer(rawCfg) {
     if (logStream) logStream.write(s + '\n');
   }
 
+  // Token totals for the lifetime of this gateway process (resets on restart).
+  const totals = { requests: 0, input: 0, output: 0, cached: 0 };
+  function recordUsage(input, output, cached) {
+    totals.requests += 1;
+    totals.input += input;
+    totals.output += output;
+    totals.cached += cached;
+  }
+  function usageSummary() {
+    return 'session ' + totals.requests + ' req  in=' + totals.input + ' out=' + totals.output + ' total=' + (totals.input + totals.output);
+  }
+
   const app = fastify({ logger: false });
   const isAzure = cfg.upstreamBaseUrl.includes('.openai.azure.com');
   const openai = new OpenAI({ baseURL: cfg.upstreamBaseUrl, apiKey: cfg.upstreamApiKey });
@@ -98,14 +110,32 @@ function makeServer(rawCfg) {
     try {
       const openaiRequest = convertRequestToOpenAI(anthropicRequest, cfg.upstreamModel, cfg.toolFormat, isAzure);
       if (isStreaming) {
-        const stream = await openai.chat.completions.create({ ...openaiRequest, stream: true });
+        const stream = await openai.chat.completions.create({ ...openaiRequest, stream: true, stream_options: { include_usage: true } });
+        // Capture token usage from the stream without altering what the vendored
+        // converter sees — it just iterates chunks, so a pass-through works.
+        const usage = { input: 0, output: 0, cached: 0 };
+        const wrapped = (async function* () {
+          for await (const chunk of stream) {
+            if (chunk && chunk.usage) {
+              usage.input = chunk.usage.prompt_tokens ?? usage.input;
+              usage.output = chunk.usage.completion_tokens ?? usage.output;
+              usage.cached = chunk.usage.prompt_tokens_details?.cached_tokens ?? usage.cached;
+            }
+            yield chunk;
+          }
+        })();
         reply.hijack();
-        await streamOpenAIToAnthropic(stream, reply, clientModel, cfg.upstreamBaseUrl);
+        await streamOpenAIToAnthropic(wrapped, reply, clientModel, cfg.upstreamBaseUrl);
+        recordUsage(usage.input, usage.output, usage.cached);
+        log('<- ' + clientModel + ' ok  in=' + usage.input + ' out=' + usage.output + ' total=' + (usage.input + usage.output) + '  | ' + usageSummary());
       } else {
         const response = await openai.chat.completions.create({ ...openaiRequest, stream: false });
-        reply.send(convertResponseToAnthropic(response, clientModel));
+        const anthropicResponse = convertResponseToAnthropic(response, clientModel);
+        reply.send(anthropicResponse);
+        const u = anthropicResponse.usage || {};
+        recordUsage(u.input_tokens || 0, u.output_tokens || 0, u.cache_read_input_tokens || 0);
+        log('<- ' + clientModel + ' ok  in=' + (u.input_tokens || 0) + ' out=' + (u.output_tokens || 0) + ' total=' + ((u.input_tokens || 0) + (u.output_tokens || 0)) + '  | ' + usageSummary());
       }
-      log('<- ' + clientModel + ' ok');
     } catch (caught) {
       const error = caught instanceof Error ? caught : new Error(String(caught));
       const rawStatus = caught && typeof caught === 'object' && 'status' in caught ? caught.status : undefined;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "anthropic-gateway",
-  "version": "1.0.0",
+  "version": "1.1.0",
   "description": "Anthropic-compatible local gateway for any OpenAI-compatible model provider — make Claude Code (GUI/CLI) work with kilo, OpenAI, DeepSeek, local models, etc.",
   "license": "MIT",
   "engines": {