WebBrain

Product Introduction

1. Definition: WebBrain is a free, open-source AI browser agent distributed as a browser extension for Chromium-based browsers (Chrome 116+, Edge, Brave, Opera, Vivaldi) and Firefox 109+. Technically categorized as an AI-powered browser automation and assistive intelligence layer, it implements a side-panel conversational interface with dual interaction modes: a read-only Ask Mode and a full agentic Act Mode. The extension manages page context via content scripts, captures token-conscious screenshots through iterative JPEG compression, and routes all natural language processing to a user-selected LLM backend — either a local inference server (llama.cpp, Ollama, LM Studio, vLLM) or a cloud API (OpenAI, Anthropic Claude, OpenRouter, Grok, Gemini, DeepSeek, Mistral). It is MIT-licensed, telemetry-free, and operates without mandatory user accounts.

2. Core Value Proposition: WebBrain exists to deliver a privacy-preserving, self-hostable alternative to proprietary browser AI plugins such as Claude for Chrome. It decouples the AI agent experience from any single vendor by supporting any OpenAI-compatible endpoint, enabling users to run the entire agent workload on-device with local LLMs — eliminating query costs, data egress, and cloud dependency. The product targets users who require a browser agent for page comprehension, structured data extraction, and multi-step workflow automation but refuse to compromise on data sovereignty, open-source auditability, or provider flexibility. Additional differentiators include a dedicated vision model slot, optional encrypted cloud sync, an MCP server bridge for connecting external coding agents, and a fair-usage cloud subscription priced at $5/month.

Main Features

1. Dual-Mode Agent Architecture (Ask Mode / Act Mode): The extension ships with a read-only Ask Mode as the default, which permits page-level Q&A, summarization, and information extraction without any possibility of mutating the page. Act Mode unlocks the full browser agent toolkit — clicking, typing, scrolling, form filling, navigation, and script execution — governed by a task-level permission gate that requires explicit user approval before consequential actions. This design mirrors a progressive trust model: users start in a safe, non-intrusive state and escalate to full automation only when the task demands it. The Act Mode instruction pipeline parses natural language goals into discrete browser operations, executes them via content scripts, and observes results through screenshot capture and DOM inspection.

2. Multi-Provider LLM Routing with Local-First Support: WebBrain connects to four primary provider types: llama.cpp for arbitrary local GGUF models via its OpenAI-compatible server, native OpenAI APIs, Anthropic Claude native APIs, and OpenRouter for access to 100+ models. Any OpenAI-compatible endpoint is also supported, including Together AI, Groq, Mistral, Gemini's OpenAI-compat layer, DeepSeek, and vLLM-served models. The default provider is llama.cpp, meaning out-of-the-box usage can be fully offline with zero API keys. A "Compact mode" compresses context for local models with as little as 8k-token windows, while the recommended minimum is 16k tokens; automatic conversation compaction prevents token overflow in longer sessions.

3. Dedicated Vision Model Separation: Rather than forcing a single multimodal model to handle both planning and screenshot interpretation, WebBrain supports a split architecture: a fast, text-only model handles reasoning, tool calls, and instruction planning, while a separate vision-capable model processes page screenshots. This reduces inference cost and latency because the expensive multimodal model only handles image-to-text extraction, while the cheaper text model manages the bulk of the reasoning workload. Both models are user-configured, enabling combinations such as a local Llama-based planner paired with a cloud vision model, or fully local configurations using Qwen 3.6 35B with vLLM serving.

4. Token-Conscious Screenshot and Context Pipeline: Before any screenshot leaves the user's machine, WebBrain resizes the image and iteratively applies JPEG compression to minimize the token footprint of image inputs. Complementing this, smart context management trims conversation history and caps tool outputs so long sessions do not produce unbounded cloud API bills. The system automatically compacts the conversation as it approaches the model's context limit, preserving task state while keeping token consumption predictable.

5. Automated Form Filling with Profile Auto-fill and Memory: Act Mode can complete web forms by recalling previously supplied entity data — demonstrated in the provided content with tax forms (legal business name, EIN, payroll wages) and signup forms. The optional Profile Auto-fill feature stores a plaintext bio (name, work email, company, throwaway password) locally in the browser, allowing the agent to breeze through low-stakes registration forms without redundant prompting. This feature is off by default, never leaves the device, and the agent stops before submission to let the user review and send.

6. Cloud Privacy and Consent Management (Cookie & Paywall Aware): The agent automatically dismisses consent-management banners (OneTrust, Cookiebot, Didomi, Quantcast) before reasoning about page content, reducing visual noise and improving screenshot clarity. It also detects paywall walls and explicitly informs the user of content access restrictions rather than hallucinating article content or attempting unauthorized bypass. CAPTCHA handling is available as an opt-in plug-in: users can supply a CapSolver API key, enabling the agent to solve reCAPTCHA v2/v3, hCaptcha, and Cloudflare Turnstile challenges mid-task — no CAPTCHA service is contacted unless enabled.

7. MCP Server Bridge for Coding Agents: A newly introduced MCP (Model Context Protocol) server, launched via npx -y @webbrain/mcp-server, connects external coding agents — Claude Code, Codex, Cursor, OpenCode, and any stdio MCP client — to the user's signed-in Chromium browser. The bridge exposes six task-level tools (webbrain_run, webbrain_extract, webbrain_status, webbrain_respond, webbrain_abort, webbrain_connection) and passes delegated goals to the WebBrain extension, which executes them in the live browser session under the existing permission gate. This eliminates the need for cookie export, headless login orchestration, or credential replay. Firefox is not supported by the MCP bridge; Chrome, Edge, Brave, Opera, and Vivaldi are.

8. Multilingual UI and Localized Marketing Site: The extension interface ships in English, Español, Français, Türkçe, and 中文, with automatic detection of the browser's language on first use and a globe icon for manual switching. The marketing site is localized to match, making the tool accessible to non-English-speaking users in privacy-critical markets.

9. Optional Encrypted Cloud Sync and Managed Subscription: For users who want cross-device consistency without self-hosting infrastructure, WebBrain Cloud — priced at $5/month per device profile — provides encrypted syncing of memories, Profile auto-fill text, and supported provider configuration (including API keys). The browser encrypts the sync vault with a user-chosen password before upload; WebBrain Cloud stores only ciphertext and cannot read the contents, nor can it recover a forgotten sync password. Chat history, browser history, page screenshots, and legacy OAuth tokens are explicitly excluded from sync. The subscription is tied to a browser + OS GUID device identifier rather than an email account, and billing is managed from the extension's Account section.

Problems Solved

1. Pain Point — The Cost and Lock-In of Proprietary Browser AI Plugins: Claude for Chrome requires a Claude Pro subscription at $20/month and is restricted exclusively to Anthropic's models. Other browser-native AI plugins follow a similar closed model. WebBrain eliminates this vendor lock-in by supporting all OpenAI-compatible APIs and local inference engines, making the per-query cost effectively zero when running on a local llama.cpp server. This directly addresses the keyword cluster around "free AI browser extension" and "self-hosted browser agent."

2. Pain Point — Data Privacy and Exfiltration Risk in Cloud-Only AI Browsing Tools: Browser AI plugins that route all page content, screenshots, and form data through a third-party cloud create an adversarial surface: sensitive corporate data, personal medical information, financial records, and login credentials pass through un-audited intermediaries. WebBrain solves this with local-first inference (llama.cpp, Ollama), no telemetry, no mandatory accounts, and MIT-licensed source that security teams can audit. The read-only Ask Mode default further reduces risk surface because the agent cannot mutate pages unless explicitly placed into Act Mode.

3. Pain Point — Prompt Injection and Automated Action Safety: An AI browser agent operating on arbitrary web pages is inherently exposed to hidden prompt injections embedded in page content. WebBrain addresses this by defaulting to read-only Ask Mode, requiring explicit confirmation before consequential actions, and applying stricter handling rules to sensitive fields. This reduces the blast radius of a malicious prompt designed to trick the agent into performing unintended actions.

4. Target Audience — Detailed User Personas:

   • Privacy-Sensitive Knowledge Workers: Legal professionals, healthcare administrators, and financial analysts who handle confidential documents in browsers and need summarization or extraction without exfiltrating data to cloud LLM providers.
   • React/Web Developers and DevOps Engineers: Users who maintain local LLM infrastructure (vLLM, llama.cpp, Ollama) and need a production browser agent that can inspect pages, suggest CSS/layout changes with DevTools context (as shown in the "Acme Studio" example), and debug live applications — without sharing proprietary codebases with third-party APIs.
   • Digital Marketing Managers and SEO Specialists: Runs product catalog extraction, competitor price monitoring, and signup form automation at scale, using token-conscious screenshot compression to keep cloud API costs predictable.
   • Tax Preparers and Small Business Admins: Form-heavy workflows such as quarterly tax returns where the agent recalls known entities (EIN, legal name, payroll data) from memory and pre-fills fields, requiring human review before submission.
   • AI/ML Researchers and Local Model Enthusiasts: Users evaluating on-device reasoning models (e.g., Qwen 3.6 35B vs. Gemma 4) who need a vision-requiring agentic test bed that runs entirely offline.
   • Open-Source Contributors and Security Auditors: Teams that require MIT-licensed, auditable browser automation software with no hidden telemetry.

5. Use Cases — Specific Scenarios Where WebBrain Is Essential:

   • Article and PDF comprehension: Ask natural-language questions about long-form content, with answers grounded in the actual page text rather than generic model training data.
   • Form automation: Quarterly tax return filing (QT-4 forms), CRM data entry, and multi-field signup workflows with values recalled from memory and a mandatory pre-submission review step.
   • E-commerce data extraction: Pulling all product names and prices from a 248-item catalog into structured data for price comparison or market research.
   • Creative/development collaboration: "How can I space these blocks farther apart?" — the agent inspects the live page's DevTools grid, proposes concrete CSS changes (e.g., gap: 10px → 28px, margin-top: 32px), and presents the diff to the user.
   • Video download automation: Locating the video stream on a media page, fetching the MP4, and saving it to the Downloads folder with format and file size confirmation.
   • Coding agent delegation via MCP: A headless CLI coding assistant (Cursor, Codex, Claude Code) delegates browser tasks — like checking a production page after a deploy — to WebBrain, which executes in the developer's signed-in browser session while the terminal agent remains in control.
   • Offline research environments: Air-gapped or VPN-isolated workstations where cloud LLM APIs are unavailable or prohibited; WebBrain with llama.cpp performs page summaries and extraction entirely on-device.

Unique Advantages

1. Differentiation vs. Claude for Chrome and Other Proprietary Browser Plugins: WebBrain is MIT-licensed and free forever, whereas Claude in Chrome is proprietary and requires a $20/month Pro subscription. WebBrain supports local model inference (llama.cpp, Ollama), multi-provider routing through OpenAI-compatible APIs, and Firefox deployment — all absent in Claude's plugin. The fully offline configuration option means organizations can adopt WebBrain in environments where Anthropic's cloud-only architecture is outright unusable. Native comparison benchmarks in the content summary show WebBrain running Gemma 4 31B on-device finishing the same task faster than Claude in Chrome on its cloud backend.

2. Differentiation vs. AI Agent Frameworks (OpenClaw, Browser-Use, Playwright-based tools): These frameworks are developer SDKs requiring Python/Docker setup, CDP or Playwright infrastructure, and typically headless browser orchestration. WebBrain is an install-and-use browser extension with an embedded side-panel chat UI, designed for non-programmers. It runs as a content script inside the user's actual signed-in browser session, meaning it inherits real session cookies, corporate SSO, and user permissions without credential replay or headless login hacks. The two categories are complementary rather than competitive: developers can use the WebBrain MCP server to let their Python SDK pipelines delegate browser interaction to the extension.

3. Key Innovation — Modular Model Routing and Vision/Text Split: Unlike single-model agent implementations, WebBrain lets users assign a fast text-only model for planning and tool invocation while independently selecting a vision-capable model for screenshot interpretation. This separation is a genuine cost optimization: text models are dramatically cheaper per token than equivalent-size multimodal models, and screenshots represent the largest token expenditure in browser agent workloads. Combined with iterative JPEG compression of screenshots, the pipeline keeps image token consumption minimal. The result is a system that makes local model usage practical (a 16k-token local model with Compact mode can run an agent session) and makes cloud usage affordable.

4. Key Innovation — Task-Level Permission Gate with MCP Integration: The MCP bridge preserves a human-in-the-loop control point even when the task originates from an external coding agent. The browser extension maintains the same ask-before-act boundary, so delegated goals are subject to the same safety checks as directly issued commands. This is architecturally superior to naive automation approaches that grant an agent full CDP control with no intermediate oversight.

Frequently Asked Questions (FAQ)

1. Is WebBrain really a free alternative to Claude's browser plugin? Yes. WebBrain provides the same core browser-agent capabilities — reading pages, extracting data, clicking, form filling, multi-step workflow automation — but is MIT-licensed and free forever. Claude's proprietary plugin requires a $20/month Claude Pro subscription and restricts users to Anthropic's models. WebBrain additionally supports local LLM inference via llama.cpp, Ollama, LM Studio, and vLLM, works on Firefox in addition to Chromium browsers, and never requires an account or telemetry.

2. Can WebBrain run completely offline with local AI models? Yes. The default provider is llama.cpp, which runs any GGUF model locally on the user's machine. To go fully offline, download a GGUF model, start llama-server, and configure WebBrain to point to it — no API keys, no internet connection, and no data leaving the device are required. Ollama is also supported through its OpenAI-compatible endpoint at http://localhost:11434/v1. For local models, WebBrain recommends at least a 16k-token context window (8k works with Compact mode), and the agent auto-compacts the conversation as it approaches the limit. A recommended model as of April 21, 2026, is Qwen 3.6 35B, using an INT4 AutoRound-quantized build on consumer GPUs and optionally served via vLLM with speculative decoding for maximum speed.

3. How does WebBrain protect against prompt injection and unsafe automated actions? WebBrain starts in read-only Ask Mode, so the agent cannot click, type, or modify anything without the user explicitly switching to Act Mode. In Act Mode, the extension asks for confirmation before consequential actions such as form submission, payment, or navigation away from the current page. The system additionally defends against hidden prompt injections embedded in untrusted web content — a known attack vector where a malicious page attempts to overwrite the agent's system instructions. Sensitive fields (passwords, credentials, financial data) are handled with stricter rules, and CAPTCHA solving is disabled by default, requiring the user to bring their own CapSolver API key to enable it.

4. What is the WebBrain MCP server and how does it differ from using a headless browser framework? The WebBrain MCP server (npx -y @webbrain/mcp-server) is a Model Context Protocol bridge that lets external coding agents such as Claude Code, Codex, Cursor, and OpenCode delegate browser tasks to WebBrain running in the user's signed-in Chromium browser. It exposes six tools — webbrain_run, webbrain_extract, webbrain_status, webbrain_respond, webbrain_abort, and webbrain_connection — and preserves the extension's task-level permission gate, so every delegated action still requires user consent in the browser. Unlike headless frameworks (Playwright, CDP-based tools), this approach uses the existing authenticated browser session, so there is no cookie export, no credential replay, and no separate login flow. The same session that the human is looking at is the session the coding agent operates through. Firefox is not supported by the MCP bridge.

5. How much does WebBrain Cloud cost and what does the encrypted Cloud Sync actually sync? WebBrain Cloud costs $5/month per device profile with unlimited normal personal usage under a fair-usage policy. The subscription is tied to a device identifier (browser + OS GUID), not an email account. Cloud Sync, included for active subscribers, synchronizes WebBrain memories, Profile auto-fill text, and supported provider configuration including API keys. Critically, the user's browser encrypts the sync vault with a password before uploading — WebBrain Cloud only stores ciphertext and cannot read memories, profile data, provider settings, or API keys. The sync vault excludes chat history, browser history, page screenshots, and legacy OAuth tokens. If the sync password is forgotten, the old vault cannot be decrypted; users must reset the cloud vault and upload a fresh encrypted copy from a device that still has the data locally.