Firecrawl 表示其重建的 MCP 将上下文使用量减少 50%

Why it matters

Firecrawl is competing to become the web-access layer for AI agents. Lower context use and simpler authentication make repeated browsing cheaper to run and easier to distribute.

Firecrawl, founded by Caleb Peffer (@CalebPeffer), Eric Ciarla (@ericciarla), and Nicolas Silberstein Camara (@nickscamara_), released a rebuilt Model Context Protocol server on July 30th that it says consumes 50% less context on search, scrape, and interact calls.

https://x.com/firecrawl/status/2082864634764075135?s=46

The release announced on X also adds browser-based OAuth onboarding for people and a keyless route for agents. Firecrawl is packaging those changes as a faster way to give any MCP-compatible AI client access to current web pages without forcing users to install a local server or manually copy credentials into a configuration file.

Firecrawl's founders arrived at the problem while building Mendable, a chat product that ingested documentation. They repeatedly had to solve the same underlying problem: collecting web pages, handling JavaScript and rate limits, cleaning the results, and converting them into data that a language model could use. They eventually separated that infrastructure from the chatbot and built Firecrawl around the API layer.

Peffer previously described that shift in Firecrawl's August 2025 Series A announcement. Firecrawl's about page says the founders also built SideGuide and now operate from San Francisco. The three founders participated in Y Combinator's Summer 2022 batch.

降低 MCP 上下文成本

Thursday 更新中的核心主张涉及每次 Firecrawl 运行 /search、/scrape 或 /interact 时消耗的模型上下文窗口量。MCP 服务器会用工具定义和每次调用返回的信息占用上下文，从而减少模型指令、对话历史、检索到的文档和推理可用的空间。

Firecrawl 表示，重建后的 MCP 在这三种调用中的上下文消耗均减少了一半。这是公司基准测试结果，公告并未证明在所有模型和客户端配置中都存在通用的 token 减少。实际目标很明确：让代理在上下文限制迫使它们压缩、丢弃或重新加载信息之前，执行更多的网页操作。

The Firecrawl MCP documentation exposes tools for web search, individual-page scraping, site mapping, crawling, document parsing, autonomous research, and browser interaction. The interact tool can click, navigate, fill forms, and operate dynamic pages after a scrape, extending Firecrawl beyond the read-only retrieval handled by a conventional search API.

身份验证现在有三条路径。交互式 MCP 客户端可以使用 OAuth，将用户引导到浏览器进行登录，而不是要求他们检索并粘贴 API key。服务器、脚本和其他无人值守的工作负载可以继续通过授权头发送 Firecrawl key。一个带速率限制的无密钥端点允许代理在没有账户或凭证的情况下开始使用。

Firecrawl's current MCP material says keyless access covers Search, Scrape, and Parse, while OAuth or an API key opens the broader tool surface. That boundary gives Firecrawl a low-friction entry point without offering unlimited anonymous access to heavier operations such as crawling and autonomous research.

Firecrawl 进一步简化入门流程

无密钥组件建立在 Ciarla 于 6 月 16 日 发布的版本之上。Firecrawl 最初将无密钥访问定位为一种方式，供代理、原型、研讨会和小型项目在开发者创建账户之前使用网页数据。新的 MCP 发布将这一分发策略与 OAuth 和更低的上下文使用结合起来。

发布时间点反映了网络数据提供商的竞争格局。Firecrawl 在其替代方案列表中列出了 Tavily, Exa, Apify, Bright Data, Browser Use, Browserbase, Jina AI, and others。搜索质量和抓取可靠性仍然是核心的购买考量，但代理开发者还必须考虑 token 消耗、身份验证、工具模式，以及某个集成是否能跨多个 AI 客户端工作。

Firecrawl 正在试图将这些运行细节作为自己的优势。开源的 Firecrawl repository 提供了底层的网页抽取软件，而 Firecrawl 则销售托管访问和更高流量的计划。Firecrawl 表示它已处理超过 50 亿次请求，并注册了超过 125 万名开发者；这些数据为公司自报。

根据其网站，Firecrawl 已筹集 1620 万美元。其中包括在 2025 年 8 月 19 日宣布的 1450 万美元的 Series A，由 Nexus Venture Partners 领投，得到 Y Combinator、Zapier、Shopify CEO Tobias Lutke、Postman CEO Abhinav Asthana 以及 Mux 创始人 Matt McClure 的支持。

更低的上下文消耗为那些构建需要频繁网页调用的代理的开发者提供了直接的成本和可用性主张。OAuth 和无密钥访问解决了同一押注的分发端：Firecrawl 希望其 MCP 端点成为代理在需要来自实时网页信息时采取的默认路径。