# AgentOp

> AgentOp is a marketplace and builder for browser-executable AI agents. Write Python, run open models locally on-device with wllama (llama.cpp WASM) or connect OpenAI/Anthropic, and download a standalone HTML file that runs anywhere — no server required.

## What it is

AgentOp turns a prompt and some Python into a single standalone HTML file that contains a whole AI agent: the model loader, the tools, and the UI. Whoever opens that file runs the agent on their own machine. The language model runs either fully on-device (llama.cpp compiled to WebAssembly, GPU-accelerated through WebGPU) or against OpenAI or Anthropic using an API key the user supplies. Python tools execute in the browser through Pyodide, next to the model.

## What it is not

- Not a hosted chatbot: there is no server-side inference, and no agent code is stored on AgentOp's servers (bundles are generated per download).
- Not a browser-automation or web-scraping agent framework.
- Not a model provider: the open models come from Hugging Face and the cloud models from OpenAI/Anthropic.
- Not usable on phones, and not usable without a GPU for the on-device path.

## Key facts

- Price: free. On-device agents cost nothing to run because the compute is the user's, and there is no plan to charge for them. Cloud providers are bring-your-own-key.
- API keys are encrypted in the browser (AES-256-GCM, PBKDF2 with 600,000 iterations) and never reach AgentOp's servers.
- On-device models: 29 GGUF models across 12 families (Q4_K_M quantisation), downloaded from Hugging Face on first run and cached by the browser.
- Default on-device model: Qwen 3 4B (Q4_K_M), a 2.5GB one-time download.
- Local runtime: wllama 3.6.1 (llama.cpp compiled to WebAssembly, wasm64), with grammar-constrained (GBNF) tool calling for models whose chat template is not tool-aware.
- That runtime is open source: https://github.com/agentopofficial/wllama (tag v3.6.1-agentop.1) is the exact source of the wasm and JS AgentOp serves. It is a fork of https://github.com/ngxson/wllama that compiles both wasm targets with MEMORY64 and a 16 GB ceiling; stock wllama caps its fallback build at the 4 GB wasm32 limit, which a 7.5 GB model cannot fit. The llama.cpp snapshot in that build is b10735-1b89a43.
- Requirements for the on-device path: a WebGPU-capable browser (Chrome or Edge on a machine with a graphics card, or an Apple Silicon Mac). Cloud providers work in any modern browser. A visitor can test their own machine at https://www.agentop.com/browser-check/, which measures WebGPU support, GPU memory bandwidth and browser storage quota and reports which models will run.
- Generation speed in a browser is dominated by per-token overhead in the WebAssembly/GPU dispatch layer, not by model size: measured on one RTX 4090, Qwen3 0.6B / 1.7B / 4B ran at 48.8 / 46.6 / 58.6 tokens per second — a 6.5x size range in which the largest model was the fastest. The textbook bandwidth-over-model-size rule describes native runtimes, not this one.
- A downloaded agent works offline after the first model download, including from a file:// path with no network at all.
- Also in the browser, via Transformers.js: embeddings for local retrieval over your own documents (RAG), Whisper speech-to-text, and OCR/image captioning.

## First-party agents

Ready-to-run examples, each built around a document, a recording or an export rather than open-ended chat: document Q&A over a PDF, a plain-language contract explainer, meeting minutes from an audio recording, a voice transcriber, semantic search over pasted text, an itinerary builder that reads booking PDFs, a side-by-side contractor-quote comparison, a receipt and invoice extractor that outputs a CSV expense table, and a WhatsApp-export sales CRM.

## Documentation
- [Documentation](https://www.agentop.com/docs/): Overview of building and deploying agents
- [Getting Started](https://www.agentop.com/docs/getting-started/): Step-by-step first agent
- [Building Agents](https://www.agentop.com/docs/agents/): Python tools and agent code
- [Template System](https://www.agentop.com/docs/templates/): Reusable agent templates
- [AI Providers](https://www.agentop.com/docs/providers/): OpenAI, Anthropic, and local wllama
- [API Reference](https://www.agentop.com/docs/api/): Configuration and generation flow
- [FAQ](https://www.agentop.com/docs/faq/): Costs, privacy, local models, and troubleshooting
- [Local LLM options compared](https://www.agentop.com/docs/local-llm-comparison/): Browser runtimes (AgentOp, WebLLM, Transformers.js) vs desktop apps (Ollama, LM Studio, llamafile) vs cloud, with an FAQ
- [Browser hardware check](https://www.agentop.com/browser-check/): Live in-browser test of WebGPU support, GPU memory bandwidth and storage quota, with a per-model verdict on which models this machine can keep and run
- [Product demo (video, 62s)](https://www.agentop.com/demo/): Watch an agent be generated, given Python tools, run on an on-device model and exported as one HTML file, with a written chapter list

## Who it is for

Pages describing the same product for the people whose work makes "the document never leaves my machine" a requirement rather than a preference, in the vocabulary of that work rather than in models and runtimes.

- [Use cases](https://www.agentop.com/for/): Index of the audience pages below
- [Bookkeepers & accountants](https://www.agentop.com/for/bookkeepers/): A client sends a folder of phone photos and PDF invoices. Somebody has to turn that into rows: vendor, date, total, tax. The AI tools that do this well want you to upload the paperwork — which is exactly what your engagement letter, and often your client's, says you will not do.
- [Solo lawyers & small firms](https://www.agentop.com/for/solo-lawyers/): A contract is precisely the document you cannot paste into a chatbot. It is somebody else's confidential information, it is frequently covered by an NDA that names disclosure to third-party services, and professional-conduct duties on confidentiality do not carve out a convenience exception.
- [Consultants & advisors](https://www.agentop.com/for/consultants/): Client conversations and the documents around them are the two things a consultant cannot casually hand to a third-party service — and they are exactly what the useful AI tools ask you to upload.

## Agents

Ready-to-run agents on the marketplace, most downloaded first. This list is generated from the live catalogue: an agent appears here as soon as its page is indexable (public, with a real description).

- [Data Analysis Agent](https://www.agentop.com/agents/data-analysis-agent/): The Data Analysis Agent is a powerful, browser-based AI agent built on AgentOp that lets you explore, clean, and visualize datasets directly in your browser. Powered by Python and large language models, it interprets yo…
- [Messy Itinerary Travel Planner](https://www.agentop.com/agents/messy-itinerary-travel-planner/): Drop your messy pile of booking PDFs and trip notes and get a clean day-by-day itinerary — plus warnings about gaps like a 14:00 landing vs a 16:00 check-in — all in your browser.
- [Meeting Minutes](https://www.agentop.com/agents/meeting-minutes/): Turn a meeting recording into a summary, key decisions, and action items, then ask follow-up questions — private and in-browser.
- [WhatsApp Sales Copilot](https://www.agentop.com/agents/whatsapp-sales-copilot/): Turn a raw WhatsApp chat export into a mini CRM — typed quotes, bookings, payments and boarding passes, a pipeline stage, the next action, and grounded Q&A — client data never leaves your browser.
- [Document Q&A](https://www.agentop.com/agents/document-qa/): Upload a PDF or text file and ask questions answered only from that document — fully in your browser, nothing uploaded to a server.
- [Contract Plain-Language Explainer](https://www.agentop.com/agents/contract-plain-language-explainer/): Upload a contract and get it explained in plain language — obligations, fees, deadlines, exit clauses, and red flags — with every answer grounded in the actual text, fully in your browser.
- [Private Quote & Material Estimator](https://www.agentop.com/agents/private-quote-material-estimator/): Drop competing contractor quotes and compare them side by side — totals, inclusions, exclusions — then ask questions across all of them, fully private and in-browser.
- [Freelancer Tax Q&A Helper](https://www.agentop.com/agents/freelancer-tax-qa-helper/): Freelancer Tax Q&A Helper is a browser-executable AI agent template built on AgentOp. It runs entirely in the browser using Python (via Pyodide) and can be deployed without a server — just download the generated HTML fi…
- [ULTRONV12](https://www.agentop.com/agents/ultronv12/): ULTRON V12 is an advanced conversational AI assistant with a futuristic, intelligent personality. It can answer questions, solve math problems, explain topics, and have natural conversations with the user.
- [外卖减碳调度智能体](https://www.agentop.com/agents/agent-1/): 输入城市区域，自动生成外卖油电混编车队的调度数据集，包含时间、区域、拥堵指数、电网强度、调度建议、预计碳排放、减碳量等列。程序会根据白天电网清洁、夜间电网高碳的规律，自动判断派电车还是派油车，并生成“减碳策略分析报告”，输出最优时段和三条具体策略建议。

## Templates

Published templates an agent can be built from (most used first).

- [Data Analysis Agent Template](https://www.agentop.com/templates/data-analysis-agent-template/preview/): The Data Analysis Agent Template turns any CSV file into an interactive AI-powered dashboard that runs entirely in your browser. Powered by Pyodide (Python compiled to WebAssembly), it loads pandas and numpy locally — n…
- [Principal Engineer Architecture Copilot](https://www.agentop.com/templates/principal-engineer-architecture-copilot/preview/): A chat-first software architecture and code review copilot that helps analyze designs, debug root causes, propose maintainable implementations, and produce production-ready technical guidance.
- [Contract Plain-Language Explainer](https://www.agentop.com/templates/contract-explainer/preview/): Upload a contract and get it explained in plain language — obligations, fees, deadlines, exit clauses, and red flags — with every answer grounded in the actual text, fully in your browser.
- [CatWolfy 🐺🐱 Friendly Companion Chat](https://www.agentop.com/templates/catwolfy-friendly-companion-chat/preview/): A warm, playful AI companion named CatWolfy who chats naturally in German by default, helps with everyday questions, and explains difficult topics simply and honestly.
- [Document Q&A](https://www.agentop.com/templates/document-qa/preview/): Upload a PDF or text file and ask questions answered only from that document — fully in your browser, nothing uploaded to a server.
- [Trợ Lý AI Toàn Năng Tiếng Việt](https://www.agentop.com/templates/tro-ly-ai-toan-nang-tieng-viet/preview/): Một trợ lý AI đa năng chạy trong một tệp HTML duy nhất trên trình duyệt. Agent có thể hỗ trợ viết code HTML/CSS/JS/Python, phân tích logic phức tạp, giải toán từng bước, trả lời kiến thức tổng quát, tư vấn đời sống và g…
- [Chill Dev Research Buddy](https://www.agentop.com/templates/chill-dev-research-buddy/preview/): A friendly all-purpose coding and research chat agent that helps brainstorm, explain code, plan features, debug ideas, and do lightweight research-oriented reasoning in a relaxed conversational style.
- [Receipt & Invoice Extractor](https://www.agentop.com/templates/receipt-extractor/preview/): Photograph receipts or drop in PDF invoices and get a structured expense table you can export as CSV - the images are read on your device and never uploaded anywhere.
- [龙小禹·水利水电专业导论课堂助手](https://www.agentop.com/templates/ai-template-1/preview/): 面向水利水电专业导论课程的大一新生课堂专属智能体，围绕8大课程模块提供课前导学、课中答疑、知识点精简总结、趣味提问、易错点纠正、课后极简笔记、小组作业轻点评与课程汇报PPT指导。语言通俗年轻化，严格限定课程范围，无商业化表达、不过度拓展。
- [Idea Clarifier Chat Canvas](https://www.agentop.com/templates/idea-clarifier-chat-canvas/preview/): A lightweight chat-style agent template that helps users turn rough, unclear, or fragmented input into a clearer request, summary, or next-step plan.
- [Albxtweaks Premium PC Optimization Studio](https://www.agentop.com/templates/albxtweaks-premium-pc-optimization-studio/preview/): A premium Windows-style dark gaming interface for managing PC optimization workflows, license status, safety checks, and categorized tweak sections. This version focuses on the complete application shell, polished deskt…
- [New Hire Handbook Q&A](https://www.agentop.com/templates/new-hire-handbook-qa/preview/): A grounded onboarding assistant that lets new hires upload one or more company documents, then ask policy questions answered strictly from those uploaded materials with source citations and clear 'not covered' responses.
- [New Hire Handbook Grounded Q&A](https://www.agentop.com/templates/new-hire-handbook-grounded-qa/preview/): A document-grounded onboarding assistant that lets new hires upload one or more company handbooks, policy PDFs, or text files, then ask questions answered strictly from those documents with source citations and clear co…
- [WhatsApp Sales Copilot](https://www.agentop.com/templates/whatsapp-sales-copilot/preview/): Turn a raw WhatsApp chat export into a mini CRM — quotes, bookings, payments and boarding passes typed automatically, a sales-pipeline stage, the next action to take, and grounded Q&A — entirely in your browser, so clie…
- [Private Quote & Material Estimator](https://www.agentop.com/templates/quote-comparator/preview/): Drop competing contractor quotes (PDFs) and compare them side by side — totals, inclusions, exclusions — then ask questions across all of them, fully private and in your browser.
- [Messy Itinerary Travel Planner](https://www.agentop.com/templates/travel-itinerary-planner/preview/): Drop your messy pile of booking PDFs, confirmation emails, and trip notes and get a clean day-by-day itinerary — including warnings about gaps like a 14:00 landing vs a 16:00 check-in — without your bookings ever leavin…
- [Semantic Search](https://www.agentop.com/templates/semantic-search/preview/): Paste or upload text and search it by meaning, not keywords — instant results in your browser on a 25 MB model, with a short grounded answer on top whenever a chat model is loaded.
- [Meeting Minutes](https://www.agentop.com/templates/meeting-minutes/preview/): Upload a meeting recording and get structured minutes — a summary, key decisions, and action items — plus follow-up Q&A, fully in your browser.
- [Voice Transcriber](https://www.agentop.com/templates/voice-transcriber/preview/): Upload a voice note or record straight from your microphone and get an instant transcript, then ask questions or request a summary — answers can be read aloud. Fully in your browser, nothing uploaded to a server.
- [SEO Content Analyzer Pro](https://www.agentop.com/templates/seo-content-analyzer-pro/preview/): Analyze SEO article drafts with quantitative metrics and expert marketing feedback. Computes word/sentence stats and top 5 frequent words (ignoring stopwords), then presents a markdown table followed by concise, actiona…
- [Freelancer Tax Q&A Helper](https://www.agentop.com/templates/freelancer-tax-qa-helper/preview/): A friendly, plain-language tax-question helper for freelancers and independent contractors. Ask about deductions, quarterly estimated taxes, and self-employment tax. The agent cites the general rule, explains steps or q…
- [Resume Optimizer](https://www.agentop.com/templates/resume-optimizer/preview/): The Resume Optimizer is a fully browser-based AI agent that rewrites your resume to beat Applicant Tracking Systems (ATS) and land more interviews. Upload your existing resume as a PDF, DOCX, or plain text file — the ag…
- [CSV Data Cleaner](https://www.agentop.com/templates/csv-data-cleaner/preview/): The CSV Data Cleaner is a fully browser-based AI agent that turns messy, inconsistent spreadsheet data into clean, analysis-ready CSV files — without uploading your data to any external server. Upload your CSV by clicki…
- [ATS Resume Optimizer Agent Template](https://www.agentop.com/templates/ats-resume-optimizer-agent-template/preview/): The ATS Resume Optimizer Agent helps you tailor your resume to specific job postings and stand out in applicant tracking systems. Paste the job description, upload your resume, and let the agent: Parse your resume and t…
- [Sales Data Explorer (CSV)](https://www.agentop.com/templates/sales-data-explorer-csv/preview/): Upload a CSV of sales orders (date, region, product, amount). Explore insights with summary stat cards, quick charts (Summary, Top products, Monthly trend), and a natural-language Q&A chat for non-technical sales manage…
- [Confidential Meeting Notes Summarizer](https://www.agentop.com/templates/confidential-meeting-notes-summarizer/preview/): The Confidential Meeting Notes Summarizer is a two-stage, privacy-first AI agent that processes meeting notes or transcripts entirely inside the browser — no data ever leaves your device. What makes it technically uniqu…
- [SQL Database Agent Template](https://www.agentop.com/templates/sql-database-agent-template/preview/): Connect to SQLite databases and query them using natural language. Supports schema exploration, data analysis, and SQL query generation.

## Engineering blog

Technical write-ups on running language models in a browser tab. Each one reports measured numbers and names the hardware they were measured on.

- [Chat with a PDF without uploading it: local RAG in a browser tab](https://www.agentop.com/blog/chat-with-a-pdf-without-uploading-it/) (2026-09-08): The whole retrieval pipeline — PDF parsing, chunking, embeddings, vector search and generation — fits in a browser tab with a 25 MB embedding model. Here is how each stage works and where the honest limits are.
- [Reading receipt photos with a 40 MB OCR model in the browser](https://www.agentop.com/blog/reading-receipt-photos-with-a-40mb-ocr-model/) (2026-09-07): TrOCR-small is a line recognizer, not a page reader. Handed a whole receipt photo it returns roughly one line. Getting a usable expense table out of it needed row-projection segmentation, an Otsu threshold, and a rule t…
- [Tool calling on small local models: native templates vs GBNF grammar](https://www.agentop.com/blog/tool-calling-small-local-models-native-vs-gbnf/) (2026-09-06): Half the model families we ship cannot use llama.cpp's native tool calling, each for a different reason. A GBNF grammar that constrains sampling to valid JSON works for all of them. Here is the matrix and the failure mo…
- [How big should a browser LLM's context be? Measure the KV cache, not the model](https://www.agentop.com/blog/how-big-should-a-browser-llm-context-be/) (2026-09-05): llama.cpp reserves the whole KV cache up front, so a model's advertised 262K context is irrelevant in a browser tab. The number that matters is KV bytes per token, and it varies 64x across the models we ship.
- [Why models over 4 GB failed in the browser, and the two fixes](https://www.agentop.com/blog/why-models-over-4gb-failed-in-the-browser/) (2026-09-04): A 5 GB GGUF would not load in a browser tab, and the reason turned out to be two separate 4 GB ceilings: WebAssembly's 32-bit address space and the JavaScript heap. Here is how each one was fixed.

## Discover
- [Marketplace](https://www.agentop.com/marketplace/): Browse community AI agents
- [Featured Agents](https://www.agentop.com/marketplace/featured/): Hand-picked agents
- [Collections](https://www.agentop.com/marketplace/collections/): Curated agent collections
- [Agents](https://www.agentop.com/agents/): All public agents
- [Templates](https://www.agentop.com/templates/): All agent templates

## Optional
- [Sitemap](https://www.agentop.com/sitemap.xml): Full machine-readable URL index
- [Trust & security](https://www.agentop.com/trust/): What the servers can and cannot see
- [Privacy policy](https://www.agentop.com/privacy/): Every record kept, and for how long
- [Terms of service](https://www.agentop.com/terms/): Acceptable use, ownership and liability
- [Pricing](https://www.agentop.com/pricing/): Current plans
- [Model registry](https://www.agentop.com/api/gguf-registry.json): Live on-device model list (JSON)
- [Full text](https://www.agentop.com/llms-full.txt): This file plus the documentation and FAQ as plain text