AgentPulse SDK

Zero-config observability for AI agents. Two lines of code → full LLM call tracking with costs, latency, and traces.

Quick Start

pip install agentpulse

import agentpulse
agentpulse.init()  # That's it — all LLM calls are now tracked

import openai
client = openai.OpenAI()
resp = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}]
)
# → Dashboard shows: model, tokens, cost, latency — automatically

What Gets Captured

With just agentpulse.init(), the SDK automatically captures:

• LLM calls — model, tokens, latency, cost (OpenAI, Anthropic, LiteLLM)
• Errors — failed API calls with error details
• Streaming — full metrics even for streamed responses
• Cost estimates — built-in pricing for 30+ models

Structured Tracing

For richer observability, add sessions, agents, and tasks:

import agentpulse
ap = agentpulse.init()

# Group related calls into sessions
with ap.session("daily-email-check") as s:
    result = process_emails()
    s.log("Processed 5 emails")
    s.set_result("success")

# Decorate agents and tasks
@agentpulse.agent(name="email-processor")
class EmailAgent:
    @agentpulse.task(name="classify")
    def classify(self, email):
        return call_llm(email)

# Monitor cron jobs
with ap.cron("nightly-cleanup") as c:
    do_cleanup()
    # Auto-captures: start, end, duration, success/failure

Manual Events

ap.event("memory_snapshot", {"file": "MEMORY.md", "size_kb": 142})
ap.metric("queue_depth", 23)
ap.alert("Cost spike", severity="warning", details="$5.20 in last hour")

Configuration

All via init() kwargs or environment variables:

agentpulse.init(
    api_key="ap_...",              # or AGENTPULSE_API_KEY
    agent_name="my-agent",         # or AGENTPULSE_AGENT (default: hostname)
    endpoint="https://...",        # or AGENTPULSE_ENDPOINT
    enabled=True,                  # or AGENTPULSE_ENABLED (kill switch)
    capture_messages=False,        # or AGENTPULSE_CAPTURE_MESSAGES (privacy)
    auto_patch=True,               # or AGENTPULSE_AUTO_PATCH
    debug=False,                   # or AGENTPULSE_DEBUG
    flush_interval=5.0,            # seconds between flushes
    max_queue_size=10_000,         # max buffered events
)

CLI

agentpulse status   # Check config and connectivity
agentpulse test     # Send a test event
agentpulse costs    # Print the built-in cost table

Design Principles

• Zero dependencies — stdlib only (urllib, json, threading, contextvars)
• Never blocks — all reporting is fire-and-forget via background thread
• <1ms overhead — just a timestamp + queue append per LLM call
• Privacy by default — prompt/response content not captured unless opted in
• Unpatchable — shutdown() restores all original library methods

License

MIT