Zero Dependencies · OSS SDK + Hosted Dashboard

Your agents are running. Do you know what they're spending?

AgentGuard stops runaway AI agents before they drain your budget. Runtime loop detection, hard dollar caps, and cost tracking per LLM call. Three lines of code. Zero dependencies.

Get Started — it's free View on GitHub
Works with LangChain OpenAI Anthropic Any Python agent
106 Tests passing
0 Dependencies
3.9+ Python support
MIT Licensed
Budget enforcement

Set a dollar limit. Agent stops when it's hit. No more surprise invoices.

Loop detection

Catch repeated tool calls before they compound. Automatic circuit breaker.

Cost tracking

See exactly what each agent run costs, per call. Dollar amounts, not guesses.

Zero dependencies

Pure Python stdlib. One package, nothing to audit. No supply chain risk.

Auto-instrument

Patch OpenAI/Anthropic in one line. Costs and traces flow automatically.

Replay tests

Record real runs and replay deterministically for regression tests.

Timeline viewer

Gantt visualization with color-coded spans and click-to-expand details.

How it works

1

Install the SDK

pip install agentguard47

2

Add 3 lines

Wrap your agent with Tracer and HttpSink

3

See traces

Gantt timelines, alerts, and usage in the dashboard

from agentguard import Tracer, BudgetGuard
from agentguard.sinks.http import HttpSink
from agentguard.instrument import patch_openai

sink = HttpSink(url="https://app.agentguard47.com/api/ingest", api_key="ag_...")
tracer = Tracer(sink=sink, service="my-agent")
patch_openai(tracer)  # auto-tracks cost per call

guard = BudgetGuard(max_cost_usd=5.00)  # stop at $5

with tracer.trace("agent.run") as span:
    span.event("reasoning.step", data={"thought": "search docs"})
    with span.span("tool.search"):
        pass  # your tool here

What builders say

"I used AgentGuard's LoopGuard to create a hard circuit-breaker for Ralph's tool-execution loop, which immediately stopped a runaway Gemini model that had been flooding Discord with over 100 duplicate error messages."
"I leveraged AgentGuard's structured tracing to diagnose a hidden failure in Ralph's reasoning, discovering that he was hallucinating missing tools because I had accidentally truncated the system prompt during an update."

Pricing

Free

$0
  • 10,000 events/month
  • 7-day retention
  • 2 API keys
  • 1 user
Start Free

Pro

$39/mo
  • 500,000 events/month
  • 30-day retention
  • 5 API keys
  • 1 user
Get Pro

Team

$149/mo
  • 5,000,000 events/month
  • 90-day retention
  • 20 API keys
  • 10 users
Get Team

Security & Trust

Zero dependencies

No supply chain risk. One pip install, one package to audit.

Open-source SDK

MIT-licensed. Read every line on GitHub. No vendor lock-in.

Encrypted in transit & at rest

TLS 1.3 in transit, AES-256 at rest via Supabase Postgres.

Retention you control

7-90 day auto-cleanup by plan. No indefinite data hoarding.

FAQ

Why zero dependencies?
Every dependency is supply chain risk. AgentGuard uses pure Python stdlib — one audit target, no transitive vulnerabilities, no surprise breaking changes from upstream packages.
What data do you store?
Only what you explicitly send via the SDK. We never auto-capture prompts, responses, or any data unless you pass it in a span or event call.
Can I self-host?
The SDK works fully offline with the built-in JSONL sink — no network calls needed. Dashboard self-hosting is on the roadmap.
Does it work with my framework?
AgentGuard is model-agnostic. LangChain has a built-in callback handler. Everything else works with 3-line manual instrumentation using the Tracer.
Do I need a credit card?
No. The free tier gives you 10,000 events/month with no credit card required.

Get updates

New features, integration guides, and agent reliability tips. No spam.