什么是 io.github.xidik12/oculo?
由 AI 驱动的原生浏览器,内置 12 个 MCP 工具;每页上下文开销约 30 tokens,轻量高效。
README
Oculo is a full-Chromium desktop browser that speaks the Model Context Protocol. Point Claude Code, Cursor, Windsurf, or any MCP client at it and your agent can see and drive real web pages — read the DOM, click, fill forms, extract data, run multi-step pipelines — through 12 compact tools that answer in under 300 tokens per flow.
Cursor : VSCode :: Oculo : Chrome
See it in action
Your agent describes intent; Oculo resolves it into a few tiny tool calls and hands back terse, redacted results — not megabytes of screenshots.
You: "Log into GitHub and star the oculo repo"
Claude Code calls:
1. act({action: "navigate", url: "https://github.com/login"})
2. act({action: "login", site: "github.com"}) # vault lookup, password never leaves the OS keychain
3. act({action: "navigate", url: "https://github.com/xidik12/oculo"})
4. act({action: "click", text: "Star"})
→ 4 tool calls, <100 tokens of response
You: "Fill out the contact form on example.com"
Claude Code calls:
1. act({action: "navigate", url: "https://example.com/contact"})
2. page() # see the form structure
3. fill({fields: {"Name": "...", "Email": "..."}, submit: true})
→ 3 tool calls
Oculo runs a real Chromium engine on your machine and adds an agent-first control layer on top. That combination is the point:
| Capability | |
|---|---|
| Native browser | Full Chromium engine — not a wrapper, extension, or headless scraper |
| 12 MCP tools | page, act, fill, read, run, media, shell, tabs, research, preview, translate, lens |
| < 300 tokens / flow | Compact text responses by default — cheaper than screenshot-based approaches |
| Self-healing automation | Selector caching + DOM diffing — 44%+ faster on repeated workflows |
| Multi-provider AI | Built-in chat with Claude, OpenAI, Gemini, Grok, OpenClaw, Ollama |
| 4-level security | auto / notify / confirm / blocked permission gate on every action |
| OS keychain vault | Credentials encrypted via electron.safeStorage (macOS Keychain / Windows DPAPI) |
| PII redaction | Credit cards, SSNs, JWTs, API keys, Bearer tokens stripped from every MCP response |
| Anti-injection | Content boundary markers + regex-based prompt-injection detection |
| 19 stealth patches | Navigator, WebGL, canvas, WebRTC, audio, font, battery, and screen fingerprint defenses |
| Headless mode | Run without a window — Docker support included |
| Cross-platform | macOS, Windows, Linux |
| Python SDK | pip install oculo — sync and async clients |
Each tool does one thing and returns the smallest useful answer. Token cost is the response size the agent pays for.
| Tool | What it does | Token cost |
|---|---|---|
page | Describe the current page — headings, forms, buttons, links. Compact, a11y (ref-tagged), and markdown modes | ~30–80 |
act | Navigate, click, hover, scroll, type, press keys, login via vault, manage tabs, cookies, proxy, recording | ~1 line |
fill | Fill form fields by label/placeholder matching, optional submit. Text, select, checkbox, contenteditable | ~1 line |
read | Extract structured data — search results, tables, lists, articles | compact |
run | Multi-step pipeline with conditionals (page/act/fill/read/wait/if). Cached for replay | header + last |
media | Generate images (Nano Banana 2 / DALL·E 3) or video (Veo 3.1); image-to-image editing | file path |
shell | Execute shell commands non-interactively (ls, npm, git, python, …) | stdout + stderr |
tabs | List all open browser tabs with URLs and titles | compact |
research | Deep web research — opens multiple tabs, reads pages, synthesizes findings | synthesized |
preview | Pre-fetch a URL without navigating away from the current page | page description |
translate | Translate page content or specific text to any language | translated text |
lens | Visual analysis of the current page via screenshot + AI vision | description |
Bonus: webmcp_list and webmcp_call discover and invoke page-declared tools via the WebMCP protocol.
Why HTTP instead of stdio? Electron's <webview> is only reachable from the renderer process. The main process — where stdio lives — can't touch page content. The HTTP bridge crosses that boundary via main-to-renderer IPC.
Port discovery. On startup Oculo writes port:authtoken to ~/.oculo-port. The bin/oculo-mcp.mjs bridge reads it automatically, so tool definitions stay discoverable even when the app is closed — only execution requires Oculo to be running.
1. Install
Grab the latest build from Releases, or run from source:
git clone https://github.com/xidik12/oculo.git
cd oculo
npm install
npm run dev
2. Register with your MCP client
Claude Code
claude mcp add oculo -- node ~/oculo/bin/oculo-mcp.mjs
Cursor / Windsurf — add to your MCP config (.cursor/mcp.json or equivalent):
{
"mcpServers": {
"oculo": {
"command": "node",
"args": ["/path/to/oculo/bin/oculo-mcp.mjs"]
}
}
}
3. (Optional) Headless & Docker
Run without a visible window for CI/CD, scraping, or server-side automation:
node bin/oculo-headless.mjs # convenience launcher
npx electron . --headless # or raw flags
npx electron . --headless --headless-auto-approve # auto-approve CONFIRM actions
OCULO_HEADLESS=1 npm run dev # or via env var
docker compose up # containerized headless (Xvfb)
4. (Optional) Python SDK
pip install oculo
from oculo import OculoClient
client = OculoClient() # auto-discovers port from ~/.oculo-port
print(client.page()) # describe the page
client.act("navigate", url="https://example.com")
client.fill({"Email": "hi@oculo.com", "Message": "Hello!"}, submit=True)
results = client.read("search results", format="json")
An async AsyncOculoClient with the same surface is also available.
Every action passes through a permission gate, and every response is redacted before it reaches the model.
Permission levels
| Level | Actions | Behavior |
|---|---|---|
| Auto | navigate, page, read, scroll, screenshot, back, forward, reload, hover, listTabs, switchTab, preview, translate, lens | Executes silently |
| Notify | click, type, fill, select, press, submit, newTab, closeTab | Executes + OS notification |
| Confirm | payment, delete_account, change_password, send_email, download, oauth, shell, evaluate, setProxy, startRecording | Native dialog approval required |
| Blocked | read_vault, export_cookies, export_tokens, disable_security | Always rejected |
Credential vault
- Encrypted with
electron.safeStorage→ OS Keychain (macOS) / DPAPI (Windows). - Passwords are never returned via IPC or MCP — only domain + username are exposed.
act({action: "login", site: "github.com"})retrieves and fills credentials without the model ever seeing them.
Redaction & anti-injection
- Every MCP response passes through a redactor that strips credit-card numbers, SSNs, JWTs, API keys, private keys, and Bearer tokens.
- Page content handed to the model is wrapped in boundary markers, and regex-based detection blocks prompt-injection attempts embedded in that content.
Stealth (19 patches)
Navigator (webdriver, languages, plugins, mimeTypes, connection, hardwareConcurrency, deviceMemory), window (chrome API, dimensions), WebGL (vendor/renderer spoofing), canvas (per-call fingerprint randomization), WebRTC (IP-leak prevention), AudioContext, font enumeration blocking, Battery API, and screen-resolution randomization.
Self-healing automation
After a successful act/fill, selectors are cached with stability scores — id and data-testid = 10, aria-label = 9, role+name = 8, text = 7, css = 5. On the next run, DOM diffing picks the strategy:
- > 80% similarity — replay from cache, no LLM call.
- 50–80% — fall back to alternative selectors.
- < 50% — re-engage the AI for a fresh resolution.
That's where the 44%+ speed-up on repeated workflows comes from.
Built-in AI chat
Oculo also ships a side-panel chat that talks to multiple providers directly:
| Provider | Auth | Models |
|---|---|---|
| Claude | API key or CLI subscription | Opus, Sonnet, Haiku |
| OpenAI | API key or Codex CLI | GPT-4o, GPT-4o mini, o1, o3 |
| Gemini | API key | 2.0 Flash, 1.5 Pro, 1.5 Flash |
| Grok | API key | Grok 2, Grok 2 Mini |
| Ollama | Local (no key) | Any pulled model |
| OpenClaw | API key | OpenClaw models |
Building from source
npm run build # production build
npm run dist:mac # macOS DMG + ZIP
npm run dist:win # Windows NSIS + portable
npm run dist:linux # Linux AppImage + deb
npm run typecheck # TypeScript
npm run lint # ESLint
npm run test # Vitest
Prerequisites: Node.js 20+, npm (not pnpm/yarn — native modules require npm), macOS / Windows / Linux.
<details> <summary><strong>Project structure</strong></summary>src/
main/ Electron main process
ai/agent.ts Multi-provider AI controller
captcha/ CAPTCHA detection + solvers
data/ Bookmarks, downloads, history, session recording
engine/ Page describer, extractor, form-detector, pipeline, resolver,
selector-cache, dom-differ, tab-manager
mcp/server.ts HTTP MCP server (ports 19516–19520, auth token)
mcp/tools/ act, fill, page, read, run tool handlers
network/proxy.ts HTTP/SOCKS proxy manager
security/ Vault, permissions, redactor, audit, anti-injection
preload/index.ts contextBridge API
renderer/
App.tsx Root browser UI component
components/ TabBar, AddressBar, ChatPanel, WebViewContainer,
bookmarks, downloads, find, history, layout, common
shared/ Types, constants, IPC channels, AI provider definitions
bin/
oculo-mcp.mjs stdio-to-HTTP MCP bridge (for Claude Code / Cursor)
oculo-headless.mjs Headless mode launcher
sdk/python/ Python SDK (pip install oculo)
Dockerfile Container deployment
docker-compose.yml Docker Compose for headless mode
Contributing
See CONTRIBUTING.md for development setup, architecture details, and how to add a new MCP tool.
Donate
If Oculo saves you time, you can support development — BTC: 12yRGpUfFznzZoz4yVfZKRxLSkAwbanw2B
License
MIT © 2026 Salakhitdinov Khidayotullo
<p align="center"> Built by <a href="https://github.com/xidik12">Salakhitdinov Khidayotullo</a> · <a href="https://getoculo.com">getoculo.com</a> </p>
常见问题
io.github.xidik12/oculo 是什么?
由 AI 驱动的原生浏览器,内置 12 个 MCP 工具;每页上下文开销约 30 tokens,轻量高效。
相关 Skills
前端设计
by anthropics
面向组件、页面、海报和 Web 应用开发,按鲜明视觉方向生成可直接落地的前端代码与高质感 UI,适合做 landing page、Dashboard 或美化现有界面,避开千篇一律的 AI 审美。
✎ 想把页面做得既能上线又有设计感,就用前端设计:组件到整站都能产出,难得的是能避开千篇一律的 AI 味。
网页应用测试
by anthropics
用 Playwright 为本地 Web 应用编写自动化测试,支持启动开发服务器、校验前端交互、排查 UI 异常、抓取截图与浏览器日志,适合调试动态页面和回归验证。
✎ 借助 Playwright 一站式验证本地 Web 应用前端功能,调 UI 时还能同步查看日志和截图,定位问题更快。
网页构建器
by anthropics
面向复杂 claude.ai HTML artifact 开发,快速初始化 React + Tailwind CSS + shadcn/ui 项目并打包为单文件 HTML,适合需要状态管理、路由或多组件交互的页面。
✎ 在 claude.ai 里做复杂网页 Artifact 很省心,多组件、状态和路由都能顺手搭起来,React、Tailwind 与 shadcn/ui 组合效率高、成品也更精致。
相关 MCP Server
GitHub
编辑精选by GitHub
GitHub 是 MCP 官方参考服务器,让 Claude 直接读写你的代码仓库和 Issues。
✎ 这个参考服务器解决了开发者想让 AI 安全访问 GitHub 数据的问题,适合需要自动化代码审查或 Issue 管理的团队。但注意它只是参考实现,生产环境得自己加固安全。
Context7 文档查询
编辑精选by Context7
Context7 是实时拉取最新文档和代码示例的智能助手,让你告别过时资料。
✎ 它能解决开发者查找文档时信息滞后的问题,特别适合快速上手新库或跟进更新。不过,依赖外部源可能导致偶尔的数据延迟,建议结合官方文档使用。
by tldraw
tldraw 是让 AI 助手直接在无限画布上绘图和协作的 MCP 服务器。
✎ 这解决了 AI 只能输出文本、无法视觉化协作的痛点——想象让 Claude 帮你画流程图或白板讨论。最适合需要快速原型设计或头脑风暴的开发者。不过,目前它只是个基础连接器,你得自己搭建画布应用才能发挥全部潜力。