Open-Source Observability Platform for AI Applications & Agents

Track prompts, latency, token usage, execution traces, tool calls, retries, hallucinations, and agent workflows in real time.

See TraceLLM monitoring AI agents in real time.

🎥 DEMO VIDEO HERE

https://your-demo-video-link.com

Modern AI applications are difficult to debug.

Traditional logs tell you:

Request received
Response generated
Request completed

They do NOT tell you:

• Which prompt was sent
• Which tools were called
• Why the model failed
• Where latency occurred
• Which step hallucinated
• Which retry fixed the issue
• How agent decisions evolved

TraceLLM solves this problem.

flowchart LR

A[AI Application]
--> B[TraceLLM SDK]

B --> C[Trace Collector]

C --> D[(SQLite)]

C --> E[Live WebSocket Stream]

E --> F[Dashboard]

D --> G[Replay Engine]

G --> F

sequenceDiagram

participant User
participant App
participant TraceLLM
participant LLM
participant SQLite

User->>App: Ask Question

App->>TraceLLM: Start Trace

TraceLLM->>LLM: Send Prompt

LLM-->>TraceLLM: Response

TraceLLM->>SQLite: Save Trace

TraceLLM-->>App: Return Result

App-->>User: Final Answer

Install directly from PyPI.

pip install tracellm-cli

Verify installation:

tracellm --help

Start TraceLLM:

tracellm start

Expected output:

✓ SQLite initialized
✓ API ready
✓ WebSocket ready
✓ Dashboard ready
✓ Opening browser...

TraceLLM now uses SQLite by default.

No MongoDB required.

No database setup required.

No .env configuration required.

Running:

tracellm start

automatically:

✓ Creates local SQLite database

~/.tracellm/traces.db

✓ Starts API server

✓ Starts dashboard

✓ Opens browser automatically

Install directly from PyPI.

pip install tracellm-cli

Verify installation:

tracellm --help

Start TraceLLM:

tracellm start

from openai import OpenAI
from tracellm.integrations.openai import trace_openai

client = OpenAI()

trace_openai(client)

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[
        {
            "role": "user",
            "content": "Explain quantum computing"
        }
    ]
)

from tracellm.integrations.langchain import trace_langchain

trace_langchain()

# existing LangChain code continues normally

One of TraceLLM's most powerful features.

Replay previous executions step-by-step.

Prompt
 ↓
Tool Call
 ↓
Tool Result
 ↓
LLM Response
 ↓
Final Output

Perfect for:

• Agent debugging
• Failure analysis
• Performance optimization
• Prompt engineering

graph TD

A[Agent]
--> B[TraceLLM]

B --> C[Live Dashboard]

C --> D[Prompt Stream]

C --> E[Tool Calls]

C --> F[Latency Metrics]

C --> G[Errors]

START TRACE
    ↓
Capture Prompt
    ↓
Capture Model Response
    ↓
Capture Tool Calls
    ↓
Calculate Latency
    ↓
Store Execution
    ↓
Stream Live Updates
    ↓
END TRACE

Total Traces       : 12,431
Successful Traces  : 12,007
Failed Traces      : 424

Success Rate       : 96.58%

Average Latency    : 1.4 sec

Tool Calls         : 31,204

Tokens Processed   : 8.7M

Meet Tracey.

The official TraceLLM dinosaur mascot.

Tracey helps users:

• Install TraceLLM
• Create first traces
• Troubleshoot issues
• Understand observability concepts
• Navigate the documentation

You'll find Tracey throughout the documentation helping guide you.

• CLI themes
• Multi-project support
• Better dashboards
• Search traces
• Export traces

• Trace filtering
• Trace comparison
• Dashboard customization

• OpenTelemetry support
• Distributed tracing
• Production deployment tooling

Contributions are welcome.

git clone https://github.com/YOUR_USERNAME/tracellm

cd tracellm

pip install -r requirements.txt

python main.py

Open a Pull Request.

• GitHub Discussions
• Issues
• Feature Requests
• Bug Reports

MIT License.

⭐ If TraceLLM helps you build better AI systems, consider starring the repository.

TraceLLM aims to become the open-source observability layer for AI applications.

Build agents.

Trace everything.

Debug faster.

Ship confidently.

🦖 Powered by Tracey.