Harsh builds the unglamorous middle layer between AI agents and the real world.

Open Source Project

Pause a browser automation, hand the page to a human, resume when they're done.

Streams the live browser to an operator when an agent hits login, 2FA, or payment, then resumes on a URL match, an element, or a plain-English description of “done” that a vision model watches for. Discord, Slack, and email notifiers ship in the box.

Python · Playwright · Browser Automation · AI Agents · Human-in-the-Loop · browser-use

Open Source Contribution

Full-stack contributor to cua's open-source computer-use agent platform — SDK, cloud, and CI/CD.

Five months across the full stack: Android-on-Docker and QEMU container support, the Yutori n1 agent loop, the reusable release workflow behind every NPM and PyPI publish, plus AWS/Terraform, Modal/vLLM, and K3s persistence on the cloud side.

Python · TypeScript · Docker · QEMU · Terraform · AWS

Open Source Contribution

Open-source virtual desktop for AI computer use with VNC streaming.

Contributed the VNC streaming layer that puts a sandboxed desktop in the browser in real time — password protection, view-only mode, and single-application streaming. It is what E2B’s Surf computer-use demo runs on.

Python · TypeScript · VNC · Virtual Desktop · AI Agents · Streaming

Open Source Project

Reverse-engineer any website's internal API using natural language.

Captures a site’s AJAX traffic, works out its pagination and dynamic parameters, and generates extraction code that replays the internal API with full parameter control. Claude does the discovery; the output is a reusable Source, not a scraper.

Python · AI · Web Scraping · API Discovery · Anthropic Claude

Open Source Contribution

Unified SDK for provisioning, managing, and scripting remote sandboxes.

One TypeScript SDK over several cloud sandbox providers. Contributed custom template orchestration, cross-provider template builders, environment variable support, resilient command execution, and binary file operations.

TypeScript · Node.js · Docker · CLI · AI Agents · DevTools

Open Source Project

Safe execution environment for AI-generated code with patching capabilities.

Runs AI-generated code in a controlled interpreter: environment switching, automatic package installation, and runtime patches that block dangerous calls while still capturing stdout, stderr, images, and plots.

Python · Code Execution · Security · AI Safety · Sandboxing

Open Source Project

Find UI elements in web applications using natural language.

Natural-language element lookup for Playwright, Selenium, and Appium, in both DOM and visual modes. Written in Go so resolving a locator never becomes the slow step in a test run.

Go · UI Testing · Playwright · Selenium · Appium · AI

Open Source Project

Ultra-fast reranking engine with CLI, API, and Python SDK.

A CPU-first reimagining of FlashRank: CLI, FastAPI server, and Python SDK over lightweight cross-encoders, with schema-driven JSON/YAML processing for low-latency retrieval pipelines.

Python · Rust · CLI · FastAPI · Retrieval · LLM