Resurf

AI Agents & Assistants

A realistic, reproducible test framework for AI browser agents using synthetic, stateful websites with failure‑mode injection

Python Latest v0.1.1 · 27d ago Security brief →

Features

Provides a production‑shaped synthetic e-commerce site (FastAPI + React + SQLite) in Docker
Injects deterministic failure modes such as latency, payment errors, and server failures per task
Offers auditable success evaluation via database state rather than LLM judgment
Supports multiple adapters (DOM/AX‑tree native, Stagehand Node bridge, vision‑only baseline)
Records full trajectory data including DOM snapshots, screenshots, actions, token counts, and latencies

View all 2 releases →

v0.1.0 New feature 27d

Notable features

resurf SDK with Environment, Task, Runner, Trajectory and adapter ABCs
shop_v1 synthetic e-commerce site (FastAPI + React + SQLite) with modifier middleware
Adapters: browser-use, stagehand, vision_baseline

Full changelog

Initial public release.

resurf SDK (import resurf): Environment, Task, Runner, Trajectory, adapter ABC.
shop_v1 synthetic e-commerce site (FastAPI + React + SQLite) with
modifier middleware (latency, payment_outcome, server_error_rate,
session_ttl, frozen_time).
Adapters: browser-use, stagehand, vision_baseline.
~10 failure-mode templates and ~10 bundled tasks across find / cart /
checkout / account / adversarial / mobile.
CLI: resurf task new | from-template | validate | try | list.
resurf-models: shared SQLModel schema + deterministic seed_database.