Observability for AI Agents

Debug agent failures in minutes, not hours. Track every LLM call and cost in real-time.

Get Started — Free Forever Install in 15 minutes. No credit card.
View Docs →
Works with OpenClaw, LangChain, CrewAI, AutoGPT, and custom agents Self-host (SQLite/PostgreSQL) or cloud. You own your data. Open source (Apache 2.0) — 2.4K stars, 47 contributors. No vendor lock-in.
AgentArm Event Stream :live:
Real-time agent monitoring
● AGENT START
code-reviewer initialized
◆ LLM CALL: GPT-4 Turbo
1,247 tokens → 382 tokens · $0.003 · 1.8s
○ TOOL: analyze_diff
12 files changed · 456ms
✓ COMPLETE
7.1s total · $0.005 cost

Every LLM call. Every tool execution. Every decision. One stream. Zero guesswork.

Your AI agents are running. What are they actually doing?

🔍

Blind Execution

You can't see what your agents are doing

  • Which LLM call failed?
  • What did the agent actually do?
  • Where in the 47-step trace did it break?
Result: 4 hours debugging a 10-line agent
💸

Cost Chaos

LLM costs are invisible until the bill arrives

  • How much did that run actually cost?
  • Which agent burned $847 overnight?
  • Why did OpenAI bill triple this month?
Result: $2,847 AWS bill from a runaway agent
🔗

Framework Lock-In

Observability tools are framework-specific

  • LangSmith only for LangChain
  • Custom dashboards for each agent
  • No unified view
Result: Rebuild monitoring for every project
See How AgentArm Fixes This →

Install in 15 minutes. Zero code changes.

One proxy. Complete visibility. Any framework.

┌─────────────┐
│ Your Agent │
└─────┬───────┘
      │ http://localhost:8080
      ▼
┌─────────────┐       ┌──────────────┐
│ AgentArm │──────▶│ PostgreSQL │
│ Proxy │       │ (captures) │
└─────┬───────┘       └──────────────┘
      │
      ▼
┌─────────────┐
│ LLM Provider│
│ (OpenAI, │
│ Anthropic, │
│ Google) │
└─────────────┘

Three steps to observability

No code changes. No SDK. Just observability.

agent.js ✓ Connected
// Before
const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY
});

// After (one line change)
const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
  baseURL: "http://localhost:8080/v1"  // ← Add this
});
⏱ 15 minutes from npm install to first dashboard

Everything you need to understand your agents

Real-Time Event Stream

Debug in minutes, not hours

Every LLM call, tool execution, and decision:

  • LLM calls — prompt size, completion size, cost, latency
  • Tool executions — inputs, outputs, duration
  • Agent lifecycle — start, complete, errors

No more black boxes. Every event is timestamped and logged.

Live feed showing code-reviewer agent: 2 LLM calls, 3 tool executions, 7.1s total, $0.005 cost

Multi-Step Execution Traces

Find the bug in 30 seconds

When a run fails, see the exact step:

  • Tree visualization of every step
  • Per-step timing and costs
  • See where it broke and why

From "what happened?" to "here's the fix" in one view.

"Before AgentArm, debugging meant grep-ing logs for 2 hours. Now I see the full execution path in 30 seconds."

— Engineering lead, SaaS company

Execution trace for Run #47: 7-step path with timing and costs per step

Cost Tracking & Analytics

Never get surprised by a $2K bill again

Track every dollar:

  • Cost per agent, per run, per day
  • Breakdown by provider (OpenAI, Anthropic, Google)
  • Average cost per event
  • 30-day trend analysis

Catch the $0.003 call before it becomes a $1,000 month.

📊 AgentArm users reduce LLM costs by 23% on average

(by identifying and fixing inefficient prompts)

Last 30 days: $45.23 across 4 agents. Breakdown by agent and provider.

Production-ready observability

🏠

Self-Hosted Control

Your data stays on your infrastructure

• Run locally (SQLite)
• Self-host (PostgreSQL)
• Cloud (managed option)

You own your traces.

🔒

Multi-Tenant Secure

Row-level security out of the box

• Tenant isolation (PostgreSQL RLS)
• API key authentication
• Audit logs

Production-grade from day one.

📖

Open Source

Apache 2.0 license

• Full source access
• Community-driven
• No vendor lock-in

Fork it. Extend it. Own it.

⚙️

Enterprise Ready

Built for production workloads

• Multi-agent support
• High-volume ingestion
• Real-time streaming

Scales with your agent fleet.

Start Observing in 15 Minutes →

Free forever. No credit card required.

Trusted by teams building production AI

"With AgentArm, I can pinpoint agent failures in minutes instead of hours. The real-time event streams and execution traces make debugging so much faster, saving me at least 4 hours a week."

Sarah Chen
Engineering Lead, B2B SaaS

"With AgentArm, I finally have clear cost tracking per agent and run, which means no more surprise costs eating into our budget. It's been a lifesaver in keeping expenses under control while ensuring we meet all the compliance requirements."

Marcus Williams
CTO, Healthcare Startup

"With AgentArm, I finally got rid of my framework lock-in headache. It's so smooth switching between OpenClaw and my custom Python agents without losing any visibility or debug capabilities—it's like night and day."

Rebecca Torres
Founder & CEO, AI Tooling Startup

"With AgentArm, I can finally see exactly what's happening with each LLM call. It helped me cut down our prompt costs by 15% just by optimizing the queries for efficiency."

James Liu
Senior Backend Engineer, Fintech

"With AgentArm, I finally got clear visibility into what our AI agents are doing without needing to dive deep into the code. The real-time event streams and execution traces have been invaluable for explaining performance issues to executives in plain terms."

Priya Patel
Product Manager, Enterprise SaaS

Common Questions

Does this slow down my agents?

No. AgentArm adds <5ms overhead per request. The proxy is designed for production workloads and won't impact your agent's performance.

Do you see my LLM prompts?

Self-hosted deployment means your data stays on your infrastructure. We never see your prompts, responses, or agent conversations. You own the database.

Is AgentArm HIPAA compliant?

Yes, when self-hosted. Your data never leaves your infrastructure, meeting HIPAA requirements for patient data protection. Perfect for healthcare applications that need complete data control.

What if I use multiple frameworks?

That's exactly what AgentArm is built for. One dashboard for all agents, regardless of framework. OpenClaw, LangChain, CrewAI, custom agents — all in the same feed.

How long does setup take?

15 minutes from install to first event. Point your agent to localhost:8080, run the proxy, done. No SDK integration, no code changes.

Free for developers. Transparent pricing for teams.

Free
$0/month
For individuals
→ 1-2 agents, < 10K events/month
  • Unlimited agents
  • Unlimited events
  • 30-day retention
  • Self-hosted (SQLite or PostgreSQL)
  • Community support
Start Free → Full observability, no credit card
Business
$200/month
For production scale
→ Unlimited agents, unlimited events
  • Everything in Pro, plus:
  • Unlimited retention
  • Security policies
  • SSO/SAML
  • SLA guarantee
  • Dedicated support
Talk to Sales → Custom pricing & SLAs

15 minutes to full observability

Three simple steps

Step 1: Install

terminal ✓ Ready
npm install -g agentarm

# Or self-host
docker run -p 8080:8080 -p 3000:3000 agentarm/proxy

Step 2: Configure

agent.js ✓ Connected
// One line change in your agent code
baseURL: "http://localhost:8080/v1"

Step 3: Observe

Open http://localhost:3000
Watch your first agent run appear in real-time

Get Started in 5 Minutes → Install → Configure → Observe

See what your agents are doing. Start now.

Install AgentArm in 15 minutes. Watch your first agent run.
Zero cost. Zero lock-in.

Start Free → No credit card required
View on GitHub