agentmap

Map every AI API call in a codebase — offline, no LLM, no API keys. Run one command and get a webpage showing which files call which providers, what models you use, and what each agent is responsible for.

pip install agentmap-scan
cd your-project
agentmap

A browser page opens with four sections:

• Calls by file — every AI API call site, file:line, provider-tagged.
• Providers & models — the distinct providers and model names in use, with counts.
• Agents & responsibilities — each agent's name and what its system prompt says it does.
• Routing — the env vars to send those calls through a gateway.

How it works

Pure regex over your source — it matches provider endpoint hosts (api.openai.com, api.anthropic.com, …) and SDK call signatures (.chat.completions.create, .messages.create, …), so it finds calls in Python, JS/TS, Go, Rust, Ruby, PHP, Java, shell/curl — not just the official SDKs. Agent responsibilities come straight from the system prompts in your code, trimmed to one line. Nothing leaves your machine.

Detects 20 providers: OpenAI, Anthropic, Google Gemini/Vertex, Azure OpenAI, DeepSeek, Groq, xAI, Mistral, Cohere, Bedrock, Together, Fireworks, OpenRouter, Perplexity, Replicate, Hugging Face, Ollama, LiteLLM, LangChain.

Commands

agentmap [PATH]              # scan + open the report (default: current dir)
agentmap --no-open PATH      # scan, write the HTML, don't launch a browser
agentmap install PATH        # write .env.agentmap routing env vars
agentmap install PATH --auto # also rewrite hardcoded provider base_urls in place

Routing → save money

agentmap install writes the env vars that point your OpenAI/Anthropic-compatible calls at a gateway. Point them at Brevitas to compress context losslessly and cut your token bill:

agentmap install . --target http://localhost:4242
source .env.agentmap

Limitations (honest)

Agent detection is a heuristic. It reliably catches raw OpenAI/Anthropic message dicts, system=/system_prompt=/instructions= assignments (incl. X_system, X_PROMPT names), CrewAI Agent(role=…), and .agent("name") calls. It does not yet understand graph frameworks (LangGraph node graphs, AutoGen). PRs adding per-framework matchers welcome.

License

MIT.