Overview and setup

The app you'll build

Every time you use a coding agent like Claude Code, Codex, or Copilot, it stores metadata about every session on your laptop. The logs live in places like ~/.claude/projects/ as JSONL files, one JSON object per line. They contain usage data, token counts, model names, tool calls - valuable data trapped in an awkward nested format.

In this workshop, taught by Alena Astrakhantseva from dltHub, we turn those logs into structured tables and dashboards. We do it with dlt and the dltHub AI workbench, which lets a coding agent build pipelines from natural-language prompts.

By the end you'll have:

1. A dlt pipeline loading local Claude Code logs into DuckDB.
2. A marimo dashboard over that data with activity, models, tokens, and projects.
3. A REST API pipeline pulling agent traces from a hosted API.
4. A scheduled deployment on the dltHub Platform with a shareable dashboard.

The architecture looks like this:

Prerequisites

You'll need these accounts and tools:

• Python 3.11 or later
• uv package manager
• A coding agent: Claude Code, Codex, or Copilot
• A dltHub Platform account (free): app.dlthub.com
• Some local agent logs so ~/.claude/projects/ has JSONL files to load, and if you don't have any yet, use your agent for a bit and come back.

Scaffold the workspace

The dltHub AI workbench has its own scaffolding command, so run it in an empty folder:

uvx dlthub-init@latest

This creates a workspace with pyproject.toml, a .dlt/ config directory, .claude/ skills, and .mcp.json for the MCP server. It also creates __deployment__.py for cloud deployment and a virtual environment. a virtual environment.

When it asks to create a virtual environment and install dependencies, say yes. It runs uv sync for you.

Open the workspace in your agent

Open the scaffolded folder in your coding agent. The agent reads the router skill and dispatches to the right toolkit when you ask it to build a pipeline.

Confirm the workbench is running:

uv run dlthub ai status

DuckDB is our destination for local development. It's an in-process analytical database - no server to run, dlt writes to a .duckdb file on disk. No extra setup is needed because DuckDB comes as a dependency of dlt.

Filesystem pipeline →