Claude Code

Claude Code is a heavy context user - long system prompts, tool schemas, and growing conversation history. Routing it through llmprof shows exactly where its tokens go and how context grows across a session.

Steps

1. Start the proxy (one instance routes Anthropic and OpenAI automatically):

   llmprof up

2. Tell Claude Code to use the proxy via ANTHROPIC_BASE_URL, then run it as usual:

   ANTHROPIC_BASE_URL=http://localhost:4000 claude

   Your ANTHROPIC_API_KEY is sent through to Anthropic unchanged. To make it persistent, export ANTHROPIC_BASE_URL in your shell profile.

3. Open http://localhost:4000. Each Claude Code turn appears as a call. Switch to the timeline view to watch the context grow turn over turn within a session.

What to look for

• The flame graph reveals how much of each turn is tool schemas vs history.
• The timeline makes context creep visible across a coding session - the history band climbing while the system prompt stays flat.
• The waste detector flags an uncached stable prefix, which for an agent that calls the model repeatedly is often the biggest reclaimable lever.

Note

llmprof groups a session’s turns automatically by prefix-chaining the messages, so a Claude Code session shows up as one run with no extra configuration.