Mission control
for your AI agents.

Know what every agent costs, the moment one breaks, and whether its answers still hold up. Built for the engineers who shipped AI to production and now have to keep it running.

Free to start No card Live in minutes
live · agent activity streaming
this week ↗ 12.4%
£4,870
spent
125,403
runs
94.2%
success
top spenders cost / runs
› content-writer £2,189
› research-agent £1,474
› support-triage £685
› email-classifier £413

Agents fail quietly.

An agent stops firing, quietly burns three times its budget, or starts returning worse answers with the same confidence as the right ones. The first you hear of it is a customer. That is the bug.

A web service crashes loudly. An agent does not: it returns a confident wrong answer and a 200, so the exception tracker never fires and the latency graph stays flat. You put agents in production this year. The tools watching them were built for software that breaks the old way.

001 three things you find out too late
01

The silent failure

Our support-triage agent quietly broke for 11 days. We found out from a customer.

Heartbeat monitoring with expected schedules. Get paged in seconds when a scheduled agent goes silent, with full context on what was supposed to run and what the last successful run produced.

02

The mystery bill

Our token bill jumped from £2,400 to £8,900 last month. Three days to figure out which agent.

Per-agent, per-customer, per-feature cost attribution. Spend alerts the moment any agent breaks its own baseline. Cost-per-successful-run as the metric finance can build budgets around.

03

The quality drift

We shipped a prompt change. Output quality fell 22%. Nobody noticed for two weeks.

Define what good looks like once: a rubric, a JSON schema, an LLM-as-judge prompt. Every run gets scored. Drift surfaces in your dashboard before it surfaces in your support inbox.

002 cost, health, and quality in one place

Meet the agent command center for fast-moving teams.

Every team running agents in production needs the same three things: what they cost, whether they're healthy, and whether they're any good. Most teams bolt that together from a billing dashboard, some logging, and a spreadsheet that's already out of date. AgentPing watches all three on one platform, so you stop stitching the picture together by hand.

Everything is measured per run: one end-to-end execution of an agent, from the moment it starts to the moment it finishes or fails.

[A]

Spend

Stop finding out about the bill at month-end.

  • Per-agent, per-customer, per-feature cost rollups, see exactly which customer is burning tokens
  • Cost-per-successful-run as your headline metric
  • Spend alerts when any agent breaks its own baseline
  • Token-by-token attribution across Anthropic, OpenAI, Bedrock, and Gemini, so a multi-provider agent is one cost line, not four dashboards
  • Server-side pricing that lines up with the provider invoice
See what Spend catches
[B]

Pulse

Find the dead agent before your customer does.

  • Heartbeat monitoring with expected schedules
  • Run-level traces with tool calls, latency, and outputs, step through any run to see what your agent actually did
  • Every run's status, latency, and error captured the moment it lands
  • Alerts to Slack, PagerDuty, Microsoft Teams, email, or webhook
  • Multi-agent trace propagation, follow the trace when one agent kicks off another, no lost context
See what Pulse catches
[C]

Verify

Catch the quality drop before your users feel it.

  • Schema and heuristic checks at zero cost on every run
  • LLM-as-judge with rubrics you write in plain English
  • Below-threshold runs flagged and alerted, per rubric
  • You set the sample rate, you control the spend
  • Drift detection on the live judge-score distribution
See what Verify catches
003 works with what you already use
[A]

Auto-instrumented LLM SDKs

Two lines in your code and every Anthropic, OpenAI, Bedrock, Mistral, Gemini, or Cohere call is captured: tokens, cost, latency, tool use, errors. No manual tracing.

[B]

Native framework integrations

First-class hooks for LangChain, LlamaIndex, Vercel AI, Pydantic AI, CrewAI, and the OpenAI Agents SDK. Trace agent loops, chains, and handoffs without custom plumbing.

[C]

Webhook for everything else

One authenticated HTTP endpoint works from cron, n8n, Make, Zapier, GitHub Actions, Claude Routines, or any shell script. If your agent can curl, it's monitored.

[A] LLM providers auto-instrumented
Anthropic Claude calls captured
OpenAI GPT calls captured
Gemini Gemini calls captured
AWS Bedrock Bedrock calls captured
Mistral Mistral calls captured
Cohere Cohere calls captured
[B] frameworks & sdks native integration
LangChain Chain-level traces
LlamaIndex Pipeline traces
Vercel AI streamText hooks
Pydantic AI agent.run() hooks
CrewAI Crew-level tracking
Agents SDK Agent loop tracking
[C] runtimes & schedulers webhook ping
Claude Routines Webhook on each fire
GitHub Actions Workflow webhook
n8n HTTP node webhook
Make HTTP module webhook
Zapier Webhook action
cron / shell curl in your shell
Anything else?  If your agent can curl, it works. Custom integrations via plain HTTP.  See all integrations →
004 how it works
01

Your agent runs

Python, TypeScript, cron, n8n, Claude Routine, anything. Two lines of code or one HTTP request.

02

AgentPing captures

Tokens, cost, latency, tool calls, outputs. Captured asynchronously and never blocks your agent; if our service is down, yours keeps running.

03

You see everything

Live dashboard, spend and quality alerts to Slack or PagerDuty, drift surfaced the moment scores slip.

Two lines for full instrumentation. One URL for everything else.

Native SDKs for Python, TypeScript, Go, Ruby, and PHP auto-capture every Anthropic, OpenAI, Bedrock, or Gemini call your agents make, with tokens, cost, latency, tool use, and errors.

For agents you can't host an SDK in, a single authenticated webhook works from cron, n8n, GitHub Actions, Claude Routines, or anything that can speak HTTP.

pip install agentping-sdk python 3.9+
import agentping

agentping.init("daily-summary")

# auto-captures every llm call
# anthropic · openai · bedrock · gemini
npm install agentping typescript / node 18+
import { AgentPing } from "agentping";

AgentPing.init({ agent: "daily-summary" });

// works with @anthropic-ai/sdk, openai,
// vercel ai sdk · node, bun, deno
go get github.com/agent-ping/agent-ping-go go 1.21+
import "github.com/agent-ping/agent-ping-go"

agentping.Init("daily-summary")

// wraps anthropic-sdk-go, go-openai
// context-aware, goroutine-safe
gem install agentping ruby 3.0+
require "agentping"

AgentPing.init("daily-summary")

# wraps anthropic-ruby, ruby-openai
# works with sidekiq, rails jobs, rake
composer require agentping/agentping php 8.1+ / laravel
use AgentPing\AgentPing;

AgentPing::init('daily-summary');

// auto-discovered service provider
// hooks scheduled commands + queued jobs
no install · no sdk bash / curl
# a single request, that's it
curl https://api.agentping.io/p/your-agent-id

# report status, duration, cost as params
curl "https://api.agentping.io/p/your-agent-id?status=ok&cost=0.012"
005 who this is for

Built for the teams that ship them.

[I]

The product engineer

"We ship AI features inside our product: a support agent, a triage classifier, a summariser. The token bill needs keeping an eye on."

  • 1 product, 5–20 agents in production
  • Customer-facing AI features
  • Need to know which customer is burning what
  • One-team eng, no dedicated SRE yet

Most start on Team.

[II]

The AI engineering lead

"We run several agents in production. They call other agents. One goes silent and we find out from a Zendesk ticket. We need traces, heartbeats, and an alert path on-call respects."

  • Multi-team, 30–100 agents across products
  • Multi-step loops, agent-to-agent handoffs
  • On-call rotation that needs real signals
  • Audit log + SSO start to matter

Most start on Business.

[III]

The shipping founder

"I shipped one critical agent into the product six weeks ago. I check the logs every morning before opening Linear. I need to sleep through the night."

  • Solo or small team, 1–5 agents
  • One agent does most of the work
  • Cost discipline before quality discipline
  • Slack alert is enough; no PagerDuty yet

Start on Free, grow into Starter.

006 pricing

Start free. Paid plans come with a 14-day trial. No per-seat, no per-span tax.

Free
Free

For developers wiring up their first agents.

  • 1 agent
  • 10k events / month
  • 1-month retention
  • Email alerts
  • Deterministic checks
Start free
Starter
£99/mo

For one team running a handful of agents.

  • 10 agents
  • 100k events / month
  • 3-month retention
  • Email + Slack alerts
  • Deterministic checks
  • Spend anomaly detection
Start free trial
Business
£499/mo

For larger teams with multiple agent fleets.

  • 100 agents
  • 10M events / month
  • 1-year retention
  • SSO / SAML
  • Audit log
  • LLM-as-judge with custom evals
  • Priority support
Start free trial
Running more than 100 agents, or need SOC2 or volume pricing?  Custom enterprise tiers are available.
007 questions you'll have

Will it slow my agents down?

No. The SDK never blocks your agent. Telemetry runs on a separate thread with a hard 2-second timeout, a bounded local queue, and graceful degradation when our service is unreachable. If we go down, your agents run as if we weren't installed.

What's the deployment model?

Fully hosted SaaS. You pick the ingestion region: US or EU.

Can I export my data?

Yes. Full event export to JSON or Parquet. API access on every tier. We don't lock your data in.

008 get started

Stop running your agents blind.

Cost, health, and quality for every agent you run in production. Live in minutes, free to start.

Per-agent cost attribution See exactly which agent is burning your token budget, before the invoice lands.
Live uptime and heartbeats Know the moment a scheduled agent goes silent, not when a customer tells you.
Quality scoring on live traffic Catch prompt drift on your most customer-facing agent before your users feel it.

Free to start. No card. In private beta. The SDK never blocks your agents, and never crashes them.