Open-Source LLM Observability for Developers
Trace prompts, responses, latency, token usage, and tool calls locally. Replay sessions, inspect failures, and debug AI applications with full visibility.
$pip install tracellm-cliagent:start ├── retrieve.docs 118ms ├── tool:web_search 401ms ├── llm.generate 842ms └── success 2.14s
Before You Start
Requirements
Make sure your environment meets these prerequisites before installing TraceLLM.
Python 3.11+
TraceLLM requires Python 3.11 or newer.
MongoDB Atlas or Self-Hosted MongoDB
A running MongoDB instance for trace storage.
PyPI Installation
pip install tracellm-cliEnvironment Configuration
Set MONGO_URL and DB_NAME in your environment.
Note: Version 1 requires MongoDB for trace storage. SQLite support is planned for a future release.
Product demo
See TraceLLM in Action
Watch TraceLLM trace, replay, and debug a real LLM workflow from the terminal.
Trace
Capture prompts, latency, token usage, retries, and tool calls.
Replay
Replay complete execution flows step-by-step.
Monitor
Watch traces arrive live through the terminal and dashboard.
$pip install tracellm-cli$tracellm start$tracellm trace "Explain transformers"Installation
Get Started in Four Steps
Install TraceLLM, configure your MongoDB connection, and start tracing in minutes.
Install
Install the TraceLLM CLI package from PyPI.
pip install tracellm-cliConfigure
Create a .env file with your MongoDB connection string and database name.
Start
Launch the TraceLLM backend service and trace collector.
tracellm startDashboard
Open the dashboard in your browser to view live traces.
Why care?
Manual debugging hides the timeline. TraceLLM shows the run as it happened: prompt, tool, retry, latency, token count, output.
What breaks?
Slow calls, hidden retry loops, bad tool outputs, prompt drift, runaway token usage, and agent branches that never converge.
What changes?
You stop reconstructing behavior from print statements and start inspecting structured, replayable execution data.
Features
Everything You Need to Debug AI Applications
TraceLLM gives you full visibility into every LLM call, tool invocation, and agent decision.
Prompt Tracing
Capture full prompt inputs, model responses, parameters, and metadata for every LLM call.
Replay Sessions
Replay execution chains step-by-step to inspect inputs, outputs, retries, and intermediate state at any point.
Live Monitoring
Stream execution events in real time as your agent runs. See latency spikes, tool failures, and status changes instantly.
Tool Call Visibility
Every tool invocation — inputs, outputs, duration, and nested calls — is captured as a first-class event in the trace.
Token Usage Tracking
Track token consumption per model call, per session, and across your entire application.
Session Export
Export traces as structured JSON files for debugging, sharing, CI pipelines, or offline analysis.
Local First Architecture
All trace data stays on your machine. No external servers, no data leaving your environment.
Open Source
Fully open-source under the MIT license. Inspect, modify, and self-host without restrictions.
Architecture
How TraceLLM Works
From your application code to the dashboard — trace data flows through a simple, local pipeline.
FAQ
Frequently Asked Questions
Common questions about TraceLLM, its requirements, and architecture.
Do I need MongoDB?
Yes. TraceLLM Version 1 stores traces in MongoDB.
Is TraceLLM open source?
Yes. TraceLLM is fully open source under the MIT license.
Does data leave my machine?
No. TraceLLM is local-first. All trace data stays on your machine.
Do I need an API key?
No. TraceLLM does not require any API key or external service registration.
CLI
Built for the Terminal
The CLI is the primary interface. Start tracing where the bug appears, copy the command, and keep moving.
$pip install tracellm-cli$tracellm start$tracellm trace$tracellm replay$tracellm monitoragent:start ├── retrieve ├── tool:web_search ├── llm.generate └── success
Replay turns “why did it do that?” into a timeline
Instead of reading logs backwards, replay shows the execution tree in order. You can inspect the prompt, tool input, tool output, retry reason, latency, and final response at the exact point where behavior changed.
A visual layer on top of TraceLLM
The dashboard is secondary by design. Use it when you want to scan runs, compare trace timelines, inspect token usage, or share a visual debugging session. The terminal remains the default workflow.
Recent traces
Dashboard placeholder for future screenshot.
Latency view
Dashboard placeholder for future screenshot.
Token usage
Dashboard placeholder for future screenshot.
Replay detail
Dashboard placeholder for future screenshot.
Mascot
Meet Tracey
A tiny dinosaur that follows every prompt, tool call, retry, and model response so you never lose track of what your AI system is doing.
Never Debug Blind Again
TraceLLM records the full execution path of LLM applications and agents.
All locally.
All open source.
All from the terminal.
Trace
Capture every request and response.
Replay
Step through execution history.
Monitor
Watch traces arrive in real time.
Docs
Everything needed to start tracing
Documentation is organized around the developer path: install, run, replay, understand the architecture, then adapt examples.
Quick Start
Install TraceLLM and capture your first trace.
CLI Commands
Reference for start, trace, replay, monitor, and export.
Replay Engine
Understand replay checkpoints and execution trees.
Architecture
Collector, trace store, dashboard, and export model.
Examples
Common patterns for LLM apps and agent workflows.