It Runs Inside the Page: Unlike Playwright or Selenium, which drive the browser from outside, PageAgent lives in the page as plain JavaScript. It reads the live DOM as text (no screenshots, no multimodal models) and uses "DOM dehydration" to compress thousands of nodes so even small text models can act precisely.

Bring Any LLM: It works with any OpenAI-compatible endpoint like GPT, Claude, Qwen, DeepSeek, and Gemini, or fully offline via Ollama. There's even a free demo LLM baked into the CDN script for instant evaluation.

Real Use Cases, Day One: Ship a SaaS copilot that actually clicks buttons instead of describing them, collapse 20-click ERP/CRM workflows into one sentence, or make any web app accessible through voice and natural language. A built-in MCP server even lets Claude Code and Cursor control your browser.

60+ Databases, One Question: Sources like UniProt, PDB, Ensembl, ClinVar, and ChEMBL, each with its own schema and query language, get queried and synthesized automatically. It's pre-configured for genomics, single-cell, proteomics, and cheminformatics, and connects to NVIDIA BioNeMo models like Evo 2, Boltz-2, and OpenFold3.

Fully Reproducible Output: It renders 3D protein structures, genome browser tracks, and chemical structures natively, and every figure ships with the exact code, environment, and message history that produced it, so results stay traceable months later.

Real Results Already: In beta, a UCSF Brain Tumor Center team cut glioma germline analysis to roughly a tenth of its former time, and the Allen Institute built a multi-agent pipeline producing 100+ page literature reviews. Anthropic is also funding research projects with up to $30,000 in compute credits (apply by July 15).

Full Design-to-Code Cycle: The 7 skills cover everything from generating new design screens in Stitch to converting Stitch HTML into production-ready React and React Native components, complete with automated validation and design token consistency.

DESIGN.md Is the Secret Weapon: One skill analyzes your Stitch project and generates a DESIGN.md, a portable markdown file encoding your entire design system (colors, typography, spacing). Any agent reads it before generating UI, so screen ten matches screen one instead of drifting off-brand.

Multi-Page Sites From a Prompt: Combined with the Stitch MCP server, agents can map screens to routes and build complete multi-page websites. No more copy-pasting HTML from a design tool into your editor.

It Remembers Your Scene: Characters stay consistent, physics hold up, and the scene remembers what came before across multiple editing turns. You can change the camera angle, environment, or style without re-describing anything or losing the thread of your original video.

Truly Multimodal Input: It natively processes text, images, audio, and video together. Feed it a product photo, a style reference, and a text prompt in a single request and it blends all three into one coherent output.

Cheap and Available Now: It's rolling out in the Gemini app, Google Flow, and YouTube Shorts, plus the Gemini API and AI Studio for developers at $0.10 per second of output, a quarter of standard Veo 3.1's price. Every clip carries an invisible SynthID watermark.

A Two-Agent Quality Loop: It's not one-shot generation. A drafter agent writes your tailored CV (in LaTeX) and cover letter, then a separate reviewer agent critiques the drafts, forces revisions, and only then presents the final output.

It Knows You Better Than Your CV Does: The /expand command scans your GitHub repos, portfolio site, Kaggle, and Google Scholar to surface skills your documents never made explicit, each added to your profile with a source tag. The /upskill command then builds a prioritized skill-gap heatmap plus a learning plan with resources and time estimates.

Swap In Your Local Market: The search skills ship configured for Danish job portals (Jobindex, Jobnet), but the whole pattern is designed to be swapped for the job boards in your country.

Frontier Scale, Efficient Inference: The Mixture-of-Experts architecture holds 1.6T total parameters but dynamically activates only 33B to 56B per token (~48B on average), while the 1M context window lets it reason over an entire large codebase at once.

Benchmark Wins: Meituan reports 59.5 on SWE-bench Pro, edging past GPT-5.5's 58.6, plus 70.8 on Terminal-Bench 2.1 and 77.3 on SWE-bench Multilingual. It was also revealed as "Owl Alpha," the stealth model that quietly topped OpenRouter's developer charts for two months.

Zero Nvidia, Full Stack: Meituan claims it's the industry's first trillion-parameter model to complete full training AND inference on a 50,000-card domestic compute cluster, a direct answer to US export controls. Weights are live on Hugging Face and GitHub, built to slot into coding harnesses like Claude Code as the agent's "brain."