Datadog for agent tools

Catch the silent tool failures
your agent can't.

One-line decorator wraps any tool — MCP, function, API. ToolPulse records latency and errors, fingerprints response shape, detects schema drift before your agent acts on bad data, and runs synthetic health checks.

Install. Decorate. Done.

pip install toolpulse

from toolpulse import monitor

@monitor(tool_name="search_web", agent_id="my-agent")
async def search_web(query: str) -> dict:
    # your existing tool, unchanged
    ...

TypeScript users: npm install toolpulse

What you get

  • Schema drift detection. When an API silently changes its response shape, you find out before your agent does.
  • Latency tracking per tool. Spot the tool that's now taking 4× longer than last week.
  • Synthetic health checks. Cron-style probes alert you the moment a tool goes dark.
  • Alerts to Discord, Slack, email, webhooks. No PagerDuty required.

Live demo: our own agent stack

Real latency, real drift events, real uptime — across the LLM tools we monitor for our own agents.

See the live status page →

How we compare

FeatureToolPulseLangfuseHeliconeArize
Tool-level latency
Schema drift detection
Synthetic health checks
MCP-nativepartial
Free tier (calls/mo)100K50K100Ktrial
Self-hostroadmap
See detailed comparisons →