Runtime guardrails for coding agents

Stop coding agents from looping, retrying forever, and burning budget.

Start local with the free SDK. Add the hosted dashboard when you need alerts, retained incident history, remote kill, and team visibility.

  • Budget enforcement, loop detection, and retry control
  • Free MIT-licensed SDK with zero runtime dependencies
  • Hosted control plane for incident history, alerts, and remote controls

Install with pip install agentguard47. Inspect the SDK on GitHub. OpenAI is the fastest local path today.

Fastest local proof

Run one guarded coding-agent call locally

Keep the first proof simple. One package. One local run. Add the hosted dashboard only after the agent matters to the team.

pip install agentguard47 openai

import agentguard
from openai import OpenAI

agentguard.init(
    service="repo-coder",
    budget_usd=5.00,
    trace_file="traces.jsonl",
    local_only=True,
)

client = OpenAI()
response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[
        {"role": "user", "content": "Review this function and return one bug risk."}
    ],
)

print(response.choices[0].message.content)
See quickstart, doctor, and starter paths
OpenAI LangChain LangGraph CrewAI Anthropic Plain Python
0 Runtime dependencies
MIT SDK license
14-day Dashboard trial
30 Day Pro retention
90 Day Team retention

Illustrative incident

When a coding agent fails, it fails fast.

The buyer problem is not generic visibility. It is stopping retry storms, looped tool calls, and runaway spend before they become a finance or ops problem.

913 retries

A broken tool call keeps firing because the agent never exits the bad path.

$423 overnight

Budget keeps climbing while nobody is watching the run.

0 shared record

The team has no retained incident trail when they come back to debug it.

AgentGuard response

The SDK stops the run locally. The dashboard keeps the incident, alerts the team, and gives you remote control afterward.

Core guardrails

Built for coding-agent failure modes, not generic tracing dashboards.

Hard runtime budgets

Put a dollar boundary on a run before a bad agent burns through budget.

Loop detection

Cut repeated tool calls and runaway agent loops before they compound.

Retry control

Stop broken retry paths from turning one failure into hundreds of useless calls.

Hosted incident operations

Use the dashboard for retained incidents, alerts, service rollups, and remote controls.

Who it is for

For engineering teams shipping coding agents in production.

Platform teams

Need one place to see which services and repos are producing incidents.

AI product teams

Need local guardrails now, then hosted alerts and retained history once usage grows.

Engineering leaders

Need budget boundaries, incident visibility, and remote control before agents reach production scale.

Start local. Add hosted operations when the team needs it.

The SDK is the free acquisition path. The dashboard is the hosted control plane for incidents over time.

Step 1

Prove value locally with the SDK

Use the SDK to verify one guarded coding-agent path before you send anything to a hosted service.

  • Run agentguard doctor
  • Run the quickstart or starter files
  • Inspect local traces and local incident output
Go to quickstart

Step 2

Add the dashboard for team operations

Use the dashboard when incidents need to be retained, shared, alerted on, and operated over time.

  • Retained incident history and alerts
  • Remote kill and hosted controls
  • Service rollups and team workflows
Start dashboard trial

Example incident flow

The SDK stops the incident locally. The dashboard makes the incident operational for the team.

Local stop

Budget cap stops a bad repo-coder run

  • Broken tool path starts retrying
  • Loop builds toward the configured budget
  • SDK stops the run at the runtime boundary
service: repo-coder incident: budget_exceeded reason: retry loop on broken tool action: run stopped locally

Hosted follow-through

The dashboard gives the team the operational trail

  • Retained incident history for review
  • Alert goes to email or webhook
  • Service shows up in shared rollups
  • Remote kill is available if another run goes bad

This is the line between the products: the SDK enforces locally, the dashboard helps the team operate incidents over time.

Trust the install

Zero runtime dependencies

One package to audit. No transitive dependency sprawl in the runtime path.

Open-source SDK

The SDK is MIT-licensed and readable on GitHub.

You choose what to send

The dashboard only sees data you explicitly send through the SDK.

Short retention by default

Hosted plans keep 30 or 90 days of retained incident history, depending on plan.

Simple pricing

The SDK stays free. The hosted dashboard starts with a 14-day trial, then moves to a paid plan.

Free SDK

$0 forever
  • MIT-licensed Python SDK
  • Zero runtime dependencies
  • Local enforcement and local proof
  • Best place to start
Start with the SDK

Pro 14-day dashboard trial

$39/mo
  • 500,000 events per month
  • 30-day retention
  • 5 API keys
  • 1 user
Start dashboard trial

Team

$79/mo
  • 5,000,000 events per month
  • 90-day retention
  • 20 API keys
  • Up to 10 users
Start dashboard trial

Questions people ask first

Do I need the dashboard to start?
No. Start with the free SDK. Add the hosted dashboard when you want retained incident history, alerts, remote kill, and shared team visibility.
What does the dashboard add?
The dashboard adds retained incident history, alerts, service rollups, remote kill, and team workflows. The SDK still owns local enforcement.
Is this generic observability?
No. AgentGuard is for runtime guardrails and incident operations for coding agents. It is not a generic prompt analytics or tracing platform.
What data do you store?
Only what you explicitly send through the SDK. We do not auto-capture prompts, responses, or other data unless you pass it in a span or event call.

Start with the free SDK. Add the dashboard when incidents become a team problem.

Run one guarded coding-agent path locally first. Move into hosted alerts, retained incident history, remote kill, and team visibility when the agent matters.

Get updates

Product updates, new guides, and real coding-agent failure writeups. No spam.