DoClaw, OpenJarvis, GroundSource & GStack Explained

You want to experiment with AI agents, run models locally, or automate development workflows, but deployment complexity, privacy concerns, and slow tooling keep getting in the way. Recent launches from Baidu, Stanford, Google, and independent developers target these exact pain points by offering browser-run agents, device-first assistants, massive disaster datasets, and developer-focused AI toolchains.

Key Takeaway: These four systems address deployment friction, local-first AI, large-scale event extraction, and efficient developer automation—each lowering a different barrier to productive AI use.

Baidu DoClaw: Instant Browser Agents

DoClaw – a new Baidu service that runs AI agents directly in a web browser without manual deployment, servers, or API keys.

DoClaw provides a fully managed OpenClaw environment hosted on Baidu AI Cloud so users can open an interface and start using agents immediately. It includes built-in integrations with Baidu Search, Baidu Baike, and Baidu Scholar, and supports switching between multiple foundation models such as DeepSeek, KimiK 2.5, GLM5, and MinimaxM 2.5.

OpenClaw originally drew rapid interest (reported >100,000 GitHub stars and ~2 million visitors in one week) but adoption slowed because deployment was complex. DoClaw removes that friction by offering a plug-and-play cloud-hosted agent platform and a promotional subscription price.

Baidu previously released a rapid deployment solution for OpenClaw on Baidu AI Cloud and integrated OpenClaw into the Baidu consumer app, aiming to reach both developers and roughly 700 million monthly active users of the app.

Key Takeaway: DoClaw simplifies agent experimentation by providing an immediately available, cloud-hosted OpenClaw interface with integrated Baidu tools and multi-model support.

Stanford OpenJarvis: Local-First Personal Agents

OpenJarvis – an open-source framework from Stanford's Scaling Intelligence Lab for building personal AI agents that run entirely on your own device.

The research shows modern local language models on consumer hardware can handle about 88.7% of common chat and reasoning tasks at interactive speeds, and local AI efficiency improved ~5.3x between 2023 and 2025.

OpenJarvis is built in layers:

• Intelligence – manages available local models and matches them to hardware.
• Engine – runs models through runtimes like OLAMA, VLLM, SGLANG, or LLAMA.cpp.
• Agents – composes specialized agents for coordination and task handling.
• Tools and memory – indexes local documents, connects to local tools, and enables inter-agent protocols.
• Learning – refines behavior via supervised fine-tuning, reinforcement learning, and prompt optimization.

Developer features include hardware scanning, model configuration suggestions, a Jarvis Doctor diagnostic tool, hardware monitoring, and JarvisBench for measuring response speed and energy cost per query. The system provides a browser UI, desktop apps (macOS, Windows, Linux), a Python SDK, and a CLI.

Key Takeaway: OpenJarvis makes practical the local-first AI approach by optimizing model selection, runtime execution, efficiency monitoring, and modular agent design for personal devices.

Google GroundSource and the FlashFlood Dataset

GroundSource – a Google system that uses AI (Gemini) to read decades of global news and convert articles into structured data about flash floods.

Flash floods cause a majority of flood-related deaths: about 85% of flood deaths worldwide and over 5,000 fatalities annually according to the World Meteorological Organization. Satellite data and existing disaster databases were limited, with only ~10,000 major events logged—insufficient for large-scale prediction modeling.

Gemini reads multilingual news articles to identify real flood events, extracts event details (location, severity, urban flash flood classification), and geocodes locations using Google Maps APIs. The result: a dataset of 2.6 million urban flash flood events across more than 150 countries. Using that dataset, Google trained a prediction model that can estimate flash flood risk up to 24 hours ahead.

Research cited indicates even a 12-hour warning can reduce flood damage by around 60%.

Key Takeaway: GroundSource leverages news-to-data pipelines to create a 2.6M-event flash flood dataset and enables short-term flood risk predictions that can materially improve preparedness.

GStack: Structured AI for Software Development

GStack – an open-source toolkit by Gary Tan that structures AI-assisted coding into specialized workflows and preserves browser state for faster automation.

GStack splits development into eight AI workflows covering planning, code review, release prep, and automated tests. Instead of launching a fresh browser per task, GStack runs a long-lived headless Chromium, keeping cookies, sessions, tabs, and stored data active. Initial browser startup normally takes ~3–5 seconds, but with the persistent engine actions run in ~100–200 milliseconds.

Commands like slash-browse allow AI to log in, navigate flows, and capture screenshots; slash-qa analyzes code changes and tests affected routes automatically.

Key Takeaway: GStack improves automation speed and reliability for developer workflows by combining specialized AI workflows with a persistent headless browser and Bun-based tooling.

Conclusion

Recent developments span four complementary directions:

• DoClaw – cloud-hosted, browser-run OpenClaw agents to remove deployment friction.
• OpenJarvis – local-first agent infrastructure enabling private, efficient on-device AI.
• GroundSource – large-scale news-to-data extraction producing a 2.6M-event flash flood dataset and short-term predictions.
• GStack – structured AI workflows and a persistent browser to speed and stabilize developer automation.

Together, these projects illustrate how agent platforms, device-based intelligence, large-scale event datasets, and developer-focused automation are converging to make AI more accessible, private, and practical.

Key Takeaway: Each system targets a specific barrier—deployment, privacy/performance, data scarcity for disasters, and development efficiency—making AI more usable across users and teams.