Tracepad documentation
Tracepad is LLM observability and evals in a single binary: traces, prompt versions, datasets and eval runs, scores and annotation queues for LLM and agent applications, self-hosted on one embedded database. Point any OpenTelemetry-instrumented app at it, or one that sends with the Langfuse SDK, and read what it did in the browser, from the CLI, or through your coding agent over MCP.
New here? Quickstart goes from nothing to a trace on screen, by hand or through your coding agent. Already sending with the Langfuse SDK? Coming from the Langfuse SDK is the switch. Weighing your options? How Tracepad compares sets it beside the tools that cover the same ground.
Get started
- Quickstart: the first trace on screen.
- Agent setup: the page to hand a coding agent, which sets Tracepad up for a project and reads the first trace back.
- Agent upgrade: the page to hand it for an update, which backs the data up, restarts the server and checks it, with a way back.
- Installing the binary: download, verify, run as a service, upgrade, back up.
- Docker: the image, its volume, TLS in front, upgrades.
- Configuration: every environment variable, and what reaches the network.
- How Tracepad compares: what other tools need to run, and where they do more.
Send traces
- Sending traces: the endpoints, the auth schemes, the attribute mapping.
- Python, Node and Go: the packages, each a thin layer over OpenTelemetry.
- Media: images and files in traces.
- Coming from the Langfuse SDK: what carries over, and what goes through Tracepad's API instead.
Use it
Read it back
- Read API, CLI and MCP: three clients of one API, returning the same bytes.
- Coding agents: which door to use, and in what order.
- Taking the data out: the archive, and what reads it.
Operate it
- Accounts, Administration and Retention.
The source is on GitHub; the changelog lists what changed in each release.