Setup

Pick an agent, point it at a model, done. This writes config.yaml for you — no YAML wrangling.

1

Agent

Which runtime should generate your traces?

2

Model

The model the agent will run with.

Model ID Reasoning effort

3

API

Where requests are sent and how they're authenticated.

API provider Base URL API key

Advanced

concurrency · timeouts · output · publishing

Parallel sessions Session timeout (seconds) Traces folder Dataset name Developer instructions Publish to Hugging Face Private dataset

Prompts

Each prompt becomes one generated session. Add them here or upload an existing prompts.jsonl.

Generate

Run every prompt through your configured agent and collect the traces.

Extract

Stage existing local Claude, Codex, Cursor, Pi, or Hermes sessions into the Studio output folder.

Provider Output folder Model filter optional Skip anonymization Session sources

Interactive

The real agent CLI, in your browser. Teich spawns the same Docker container the batch runner uses and drops you into a native pi / codex / claude / hermes session — steer it however you like, then save the trace straight into your dataset.

New session

Agent

Model GitHub repo optional

Dataset Preview

Preview the output folder as a Hugging Face-style dataset before upload.

▦

Select a row to inspect it.

Agent

Model

API

Advanced

No prompts yet

New session