I built AgentPulse because I kept getting surprise bills from my AI agents and had no idea which calls were burning money.

The problem: You build an agent, it works great. Then you check your OpenAI bill: $400. Which agent? Which calls? No clue.

AgentPulse is lightweight observability for AI agents:

- Cost tracking per trace (supports GPT-4o, Claude, etc.) - Full span tree showing every LLM call and tool use - Auto-instrumentation for OpenAI/Anthropic (3 lines of code) - Self-hostable with SQLite — your data stays local - Framework-agnostic (works with LangChain, CrewAI, or plain Python)

Quick start:

    from agentpulse import AgentPulse, trace
    
    ap = AgentPulse(endpoint="http://localhost:3000")
    
    @trace(name="my-agent")
    def run():
        # your agent code
        pass

Or one-liner local install:

    curl -fsSL https://raw.githubusercontent.com/nandusmasta/agentpulse/main/install.sh | bash

It's MIT licensed and free forever. I'd love feedback on what's missing or broken.