Open-source AI agent observability platformCA : 7Dpd...pump

SeeWhatYour
AIAgentActuallyDid.

Record every LLM call, tool execution, and file edit your AI agent makes. Replay sessions visually, pinpoint failures, and share annotated traces with your team — in seconds, not hours.

Start Free
View on GitHub
Session #41·auth refactor
Error
14.7s9 spans
1x
LLM
Tool
File
Error
0s3s6s9s12s14.7s
Select a span to inspect
Features

Everything you need to
debug autonomous agents

From session recording to visual replay to root cause analysis — Duskfall gives engineering teams full observability into autonomous agent behavior.

Session Recording SDK

A single import captures every LLM call, tool execution, file edit, and error as structured spans. Zero-config, framework-agnostic.

Visual Session Timeline

Every agent action rendered as color-coded nodes on an interactive timeline. Click to expand full prompt/response pairs, file diffs, and stack traces.

Session Replay & Scrubbing

Watch agent sessions replay in real-time or at 1x–10x speed. Scrub back and forth like a video player. Pause and inspect full state at any point.

Search & Filtering

Search across all sessions by error message, tool name, file path, model, or date range. Filter within a session to surface only what matters.

Sharing & Annotations

Generate shareable URLs for any session. Add annotations at specific points to explain what went wrong and build institutional knowledge.

How It Works

Three steps to clarity

1

Clone the repo

Clone Duskfall from GitHub and set up your agent observability in minutes.

git clone duskfalldotapp/duskfall
2

Run your agent

The SDK silently captures every event — LLM calls, tool usage, file edits, and errors — in real time.

record({ projectKey })
3

Debug visually

Open the dashboard, replay the session, and pinpoint exactly where the agent went off track.

duskfall.dev/dashboard
Integration

Three lines of code.
Full visibility.

Wrap your agent with the Duskfall recorder and every event is captured automatically. No manual instrumentation needed.

  • Framework agnostic — LangChain, CrewAI, custom pipelines
  • OpenTelemetry-compatible span format
  • Real-time streaming via WebSocket with REST fallback
  • SDKs for Node.js, Python, and Go
agent.ts
import { record } from '@duskfall/sdk';
import { runAgent } from './my-agent';

// Wrap your agent — that's it
const session = record({
  projectKey: 'my-project',
  metadata: {
    model: 'claude-opus-4-6',
    framework: 'custom'
  }
});

// Run your agent as normal
await runAgent(session);

// End when done
await session.end();
// → View at duskfall.dev/dashboard

Stop guessing.
Start replaying.

Connect your wallet to create a project, or explore the live demo with real agent sessions, annotated timelines, and failure analysis.

Start FreeStar on GitHub