# Browserbase MCP — Cloud Browser Automation Tools > Browserbase MCP server exposes automation tools (navigate, act, observe, extract) backed by Browserbase + Stagehand, letting agents operate real web pages. ## Install Merge the JSON below into your `.mcp.json`: ## Quick Use 1. Add to your MCP config (self-hosted via NPM): ```json { "mcpServers": { "browserbase": { "command": "npx", "args": ["@browserbasehq/mcp"], "env": { "BROWSERBASE_API_KEY": "", "BROWSERBASE_PROJECT_ID": "", "GEMINI_API_KEY": "" } } } } ``` 2. Restart your MCP client. 3. Ask the agent to `navigate` then `extract` a page summary. ## Intro Browserbase MCP server exposes automation tools (navigate, act, observe, extract) backed by Browserbase + Stagehand, letting agents operate real web pages. - **Best for:** teams that want web automation as tools (testing, scraping, form flows) without running local browsers - **Works with:** MCP clients, Browserbase project + API key, Stagehand model provider key (default Gemini in README) - **Setup time:** 10 minutes ## Practical Notes - Exposes 6 tools: start/end/navigate/act/observe/extract (README tool table) - Supports both hosted SHTTP and self-hosted STDIO; flags like viewport and port are configurable (README config section) ## Make Browser Automation Predictable Browser tools become flaky when prompts are vague. A few tactics help: - Always `observe` before you `act` so the agent uses stable element affordances. - Keep your extraction requests structured (“extract a table with columns X/Y/Z”) instead of “summarize”. - For CI-style runs, pin viewport size and avoid infinite-scroll pages. If you need a 100% local deployment, the README documents two options: ```bash git clone https://github.com/browserbase/mcp-server-browserbase.git cd mcp-server-browserbase npm install && npm run build ``` Then run the server via `node` (or Docker) and point your MCP client at it. ### FAQ **Q: What credentials do I need?** A: At minimum: `BROWSERBASE_API_KEY` + `BROWSERBASE_PROJECT_ID`; a model API key may be required depending on your setup. **Q: Hosted vs self-hosted?** A: Hosted SHTTP is the easiest; self-hosted STDIO is available via `npx @browserbasehq/mcp` or Docker. **Q: How do I reduce flakiness?** A: Use observe→act, keep instructions structured, and pin viewport/inputs. ## Source & Thanks > Source: https://github.com/browserbase/mcp-server-browserbase > License: Apache-2.0 > GitHub stars: 3,331 · forks: 359 --- ## 快速使用 1. 加入 MCP 配置(NPM 自托管方式): ```json { "mcpServers": { "browserbase": { "command": "npx", "args": ["@browserbasehq/mcp"], "env": { "BROWSERBASE_API_KEY": "", "BROWSERBASE_PROJECT_ID": "", "GEMINI_API_KEY": "" } } } } ``` 2. 重启 MCP 客户端。 3. 让 Agent 先 `navigate` 再 `extract` 做一次页面摘要。 ## 简介 Browserbase MCP server 提供浏览器自动化 tools(navigate/act/observe/extract 等),基于 Browserbase + Stagehand 让 Agent 操作真实网页并抽取信息。 - **适合谁:** 希望把网页自动化当工具层(测试、抓取、表单流程)但不想本地跑浏览器的团队 - **可搭配:** MCP 客户端、Browserbase 项目与 API key、Stagehand 模型提供方 key(README 默认 Gemini) - **准备时间:** 10 分钟 ## 实战建议 - 提供 6 个工具:start/end/navigate/act/observe/extract(见 README) - 同时支持 hosted SHTTP 与自托管 STDIO;端口/视窗等参数可通过 flags 配置(见 README) ## 让浏览器自动化更可控 浏览器工具容易“飘”,通常是因为指令太泛。几条经验能显著提升稳定性: - `act` 之前先 `observe`,让 Agent 基于可见的可操作元素做选择。 - 抽取请求尽量结构化(例如“抽一张表,列为 X/Y/Z”),不要只说“总结一下”。 - 类 CI 任务建议固定 viewport,避开无限滚动页面。 如果你要 100% 本地部署,README 也提供了方案: ```bash git clone https://github.com/browserbase/mcp-server-browserbase.git cd mcp-server-browserbase npm install && npm run build ``` 之后用 `node`(或 Docker)起服务,再让 MCP 客户端指向该 server。 ### FAQ **需要哪些凭据?** 答:至少需要 `BROWSERBASE_API_KEY` 与 `BROWSERBASE_PROJECT_ID`;按 README 的模型配置可能还需要模型方 API key。 **Hosted 和自托管怎么选?** 答:README 推荐 hosted SHTTP 最省事;也支持通过 `npx @browserbasehq/mcp` 或 Docker 自托管。 **怎么降低不稳定?** 答:先 observe 再 act、指令结构化、固定 viewport 与输入。 ## 来源与感谢 > Source: https://github.com/browserbase/mcp-server-browserbase > License: Apache-2.0 > GitHub stars: 3,331 · forks: 359 --- Source: https://tokrepo.com/en/workflows/browserbase-mcp-cloud-browser-automation-tools Author: MCP Hub