Open source · Apache 2.0

Your AI agents fail.
Silently.

Real-time monitoring for AI agents. Dunetrace detects structural failures in real-time and fires a Slack alert within 15 seconds.

Get started free View on GitHub ↗
17
Detectors
<15s
Alert latency
0
Raw content sent
<500µs
Agent overhead
The Problem

Your logs say OK.
Your agent isn't.

API returns 200. No exceptions. But the agent called the same tool 12 times, burned through your token budget, and delivered a wrong answer, or no answer at all.

LangSmith and Langfuse answer "what happened?" after you already know something broke. Dunetrace answers a different question: is something breaking right now?

✗  Tool called 12× · API returns 200
✗  $12 in tokens burned · logs show no errors
✗  User got wrong answer · no alert fired
research-agent · run_f9c2 · ● live
step 01web_search"climate 2026"✓ 820tk
step 02web_search"climate 2026"⟳ 820tk
step 03web_search"climate 2026"⟳ 820tk
step 04web_search"climate 2026"⟳ 820tk
step 05web_search"climate 2026"⟳ 820tk
step 06web_search"climate 2026"⟳ 820tk
⚡ TOOL_LOOP detected HIGH
web_search called 6× with identical args
steps 2–6 · 4,100 tokens wasted
→ fix: add seen-queries deduplication
Slack #agent-alerts · delivered in 11s · First occurrence
How It Works

Zero-config. Zero latency impact.

01

Instrument once

Two lines of Python. The SDK patches OpenAI, Anthropic, httpx, and requests globally. No code changes per agent.

→
02

Detect automatically

17 structural detectors run on every completed run. Events are SHA-256 hashed. No raw content ever leaves your process.

→
03

Alert before users do

Slack or webhook. Plain-English explanation. Suggested fix. Rate context: first occurrence or systemic issue.

17 Structural Detectors

Catches what logs miss

Every detector runs automatically on every completed run. All thresholds configurable via detectors.yml — no code changes.

CRIT
Prompt Injection Signal
Input matches known injection / jailbreak patterns before hashing
HIGH
Tool Loop
Same tool called ≥3× in any 5-tool-call window
HIGH
Tool Thrashing
Agent alternates between exactly two tools repeatedly
HIGH
Retry Storm
Same tool fails 3+ times consecutively
HIGH
LLM Truncation Loop
finish_reason=length fires ≥2 times. Context overflowing.
HIGH
Empty LLM Response
Model returned zero-length output with finish_reason=stop
HIGH
Cascading Tool Failure
3+ consecutive failures across 2+ distinct tools
MED
Reasoning Stall
LLM:tool-call ratio ≥4×. Agent thinking without acting.
MED
Goal Abandonment
Tool use stops, then ≥4 consecutive LLM calls with no exit
MED
Context Bloat
Prompt tokens grow 3× from first to last LLM call
MED
RAG Empty Retrieval
Retrieval returned 0 results but agent answered anyway
MED
Slow Step
Tool call >15s or LLM call >30s detected
MED
Tool Avoidance
Final answer given without calling available tools
MED
Step Count Inflation
Run used >2× the P75 step count for this agent
MED
Cost Spike
Total token cost exceeded the per-agent P75 baseline by the configured inflation factor
MED
Session Latency
Total run wall-clock time exceeded the per-agent P75 baseline by the configured inflation factor
HIGH
First Step Failure
Error or empty output at step ≤2. Failing at the gate.

All detectors run automatically. Shadow mode lets you validate custom detectors against real traffic before they page anyone.

Dashboard

Every signal. Explained.

Live auto-refreshes every 15s. Plain-English explanations. Suggested fixes for every failure.

localhost:3000 — Overview
Dunetrace overview dashboard with risk trend and failure patterns
localhost:3000 — Analytics
Dunetrace analytics showing failure trends and patterns over time
localhost:3000 — Agent Detail
Dunetrace agent detail with run timeline and signal analysis
Quick Start

Running in under 5 minutes

  • 1
    Start the backend
    Clone the repo and spin up with Docker Compose. PostgreSQL included. No external dependencies.
  • 2
    Install the SDK
    Pure Python. Supports any framework. First-class LangChain and LangGraph support.
  • 3
    Add two lines
    Decorate your agent function. The SDK patches OpenAI, Anthropic, httpx and requests globally. Automatic.
  • 4
    Open the dashboard
    Navigate to localhost:3000. Detectors start firing immediately. Configure Slack for real-time alerts.
terminal
# 1. Start the backend
git clone https://github.com/dunetrace/dunetrace
cd dunetrace && cp .env.example .env
docker compose build && docker compose up -d

# 2. Install the SDK
pip install dunetrace
# or for LangChain:
pip install 'dunetrace[langchain]'
agent.py
from dunetrace import Dunetrace

dt = Dunetrace()
dt.init(agent_id="my-agent")

@dt.agent()
def run_agent(query: str) -> str:
    ...  # LLM + HTTP calls tracked automatically
Alerts

You shouldn't have to dig through logs to find out something broke.

Traditional monitoring never tells you. You find out when a user complains, then spend hours hunting through logs. Dunetrace fires a Slack alert within 15 seconds of a completed run, with a plain-English explanation and a suggested fix already attached.

✓  Slack Block Kit alerts
✓  Generic webhook (PagerDuty, Linear, custom)
✓  Weekly digest with 7-day aggregates
✓  At-least-once delivery with retry
Dunetrace
#agent-alerts
⚑ TOOL_LOOP detected
Agent: research-agent
Severity: HIGH
What: web_search called 8× in last 5-tool window
Why: Agent is stuck in a retrieval loop, likely failing to synthesize results
Fix: Add a tool-call ceiling or result-dedup step
Rate context: Recurring · seen in 3 of last 10 runs (30%)
Dunetrace Slack alert example
🔒
Privacy by Design

No raw content. Ever.

Every prompt, tool argument, and model output is SHA-256 hashed before transmission. Dunetrace detects structural patterns, not content. Your data never leaves your infrastructure.

prompt: "What are the current interest rates?" →  a3f8c1d9e2b47f5a0c6d8e1b3f2a9c47d8e1b3c5a2f9d6e0b4c7a1f3e8d2b5c9
Integrations

Works with your stack

Drop-in support for every major agent framework, in Python and TypeScript.

LangChain
LangGraph
CrewAI
AutoGen
Haystack
Langdock
Langfuse
FastAPI
Flask
TypeScript / Node
OpenTelemetry
Deep Analysis

Already using Langfuse? Connect your existing traces.

Dunetrace works alongside Langfuse, not instead of it. Langfuse tells you what happened in a run. Dunetrace tells you what went structurally wrong and fires an alert before you ever open a trace. Use both: get the alert from Dunetrace, then drill into the full trace in Langfuse for root cause analysis.

Read the Langfuse integration docs →
🔗
Dunetrace + Langfuse
Now live

Use Claude to debug your agents

The Dunetrace MCP server is live. Connect it to Claude, Cursor, or any MCP-compatible AI and ask questions like "Is my agent healthy?", "What failures happened in the last 24 hours?", or "Show me the details of run XYZ" — directly in your AI chat.

Set up MCP →
list_agents summarize_agent get_agent_health get_agent_patterns get_agent_runs get_agent_signals get_signal_detail search_signals get_run_detail get_instrumentation_guide
# 1. install
pip install dunetrace-mcp

# 2. ~/.claude.json
{
  "mcpServers": {
    "dunetrace": {
      "command": "dunetrace-mcp",
      "env": {
        "DUNETRACE_API_URL": "http://localhost:8002",
        "DUNETRACE_API_KEY": "dt_dev_test"
      }
    }
  }
}
Design Partner Program

We're looking for a design partner

We're working with a small group of teams to shape what Dunetrace becomes. If you're running AI agents in production and want early access, direct influence on the roadmap, and a team that will actually fix your problems — let's talk.

Contact us →
Open Source · Apache 2.0

Stop finding out from users.

Instrument your agent in 5 minutes. Get your first alert before your next user does.

Book a demo View on GitHub ↗ Join Discord Contact us