First public release of TeachReplay — teach once, replay anywhere.
Teach-by-Demonstration: demonstrate a workflow once on a computer; TeachReplay records a structured, semantic trajectory (element roles and names — never raw coordinates) and compiles it into a reusable, parameterized skill: your demonstrated values become inputs with your values as defaults.
GUI + CLI in one workflow: browser actions and shell commands interleave; shell steps run for real, are recorded with their real exit codes, and replay with exit-code verification.
Runs on real computers:
- any SSH-reachable Linux machine with a virtual desktop (generic remote backend)
- the bot's cloud computer (Box) backend
- a zero-credential in-app demo computer for trying it instantly
Deterministic replay + explicit verification: every replay re-snapshots, matches targets semantically, verifies each step's effect, and ends in an explicit success or failure — "the engine stopped" is never success. Severe UI drift fails loudly instead of clicking wrong things.
Optional DeepSeek support: experimental provider driver with text models and deepseek-v4-flash-vision-exp image input.
Small real-world evaluation (8 tasks / 14 replays on a real remote Linux box, results in scripts/teach-benchmark-results.json): 14/14 verdicts correct, 12/12 normal replays, 6/6 changed-parameter replays, 2/2 explicit failure detections, 0 ground-truth mismatches. An honest sanity check, not a SOTA claim.
Quick start: pnpm install && pnpm dev:server && pnpm dev, then click the graduation-cap button. See the README for the demo video and full setup.
Built on OpenMausBot (fork base 084f594, Apache-2.0).