Who it is forA researcher reviewing a failed or uncertain rollout who needs to see what the agent observed and did, step by step.

The problemReading raw episode logs makes it hard to judge whether an action changed the world as the agent claimed.

The outcomeThe researcher gets replay.mp4 showing the reconstructed timeline of the episode.

What you need first

Python 3.10 or newer and uv installed
An OpenAI-compatible Planner LLM or VLM endpoint for any run that reasons over a task
One simulator backend installed through scripts/setup_envs.sh, such as libero

Steps

Run uv sync --extra dev in the Open ETA repository folder.
Locate the episode log folder from a previous run, such as an episode directory under outputs/.
Run uv run openeta-replay EPISODE_DIR --output replay.mp4 with EPISODE_DIR replaced by that folder.

Prompt for your agent

In the OpenETA repo, run `uv run openeta-replay <episode dir> --output replay.mp4` and report the output path and whether the file was created. Verify the timeline matches the episode's recorded turns.

Caveat. The README says replay rebuilds the timeline from episode logs, so it only shows what those logs recorded and cannot add missing evidence.

Use this in your AI agent

Ask your agent

Using the Deduped MCP, get the setup steps for OpenMOSS/OpenETA and "Turn an episode log into a replay video"

Add Deduped to Claude Code once

claude mcp add --transport http deduped https://dedupedai.com/mcp

Cursor, ChatGPT, others

Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer

What you need first

Steps

Prompt for your agent

Use this in your AI agent

Run the default town and inspect it in the dashboard

Compile your notes into a searchable Markdown wiki

Save a Bilibili or YouTube video as Markdown notes