← MCP

Firecrawl MCP ServerFirecrawl MCP Server

Firecrawl

MCPMCP 抓取Scraping 搜索Search MITMIT 远程托管Remote-hosted keylesskeyless 浏览器交互Browser interaction partial lifecycle: active 核验Checked 2026-10-23

Firecrawl 官方 MCP server（MIT）：web 上下文全家桶——搜索、抓正文、解析 PDF、站点级 crawl/map、真实浏览器 interact、页面 monitor。三条硬口径：keyless 免费档只有 3 工具； 发版线在 npm 不在 GitHub releases；API key 绝不进 URL 也不进 agent 聊天（官方 README 明令）。仓已从 mendableai/ 迁到 firecrawl/ org（301 坐实）。

Firecrawl's official MCP server (MIT): the full web-context kit — search, scrape, PDF parsing, site-level crawl/map, real-browser interact, page monitoring. Three hard facts: the keyless free tier exposes only 3 tools; the release line is npm, not GitHub releases; the API key must never go into the server URL or the agent chat (explicit in the README). The repo moved from mendableai/ to the firecrawl/ org (301 confirmed).

适合与不适合When to use it, when not to

适合：要「搜索→抓取→结构化→浏览器交互→监控」全链一面包齐；keyless 起步再升 key。不适合：要逐步驱动的长浏览器会话（README 明写不是它的模型）；只要单页读取的极简场景。

Fits: agents wanting the whole chain — search → scrape → structured JSON → browser interact → monitor — in one package; keyless start then upgrade to a key. Not a fit: long browser sessions driven step-by-step (the README says that is not its model); minimal single-page reading (the fetch reference is lighter).

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-23

出品属地Origin: 国外出品foreign-made　Firecrawl（mendable inc.，美国）。Firecrawl (mendable inc., US).
国内使用Direct access: 直连不可达direct blocked　firecrawl.dev 与 mcp.firecrawl.dev 直连实测 000（网络受限，非死链）；GitHub 仓与 raw README 经代理 200（2026-10-23）。firecrawl.dev and mcp.firecrawl.dev measured 000 direct (network restricted, not dead); the GitHub repo and raw README returned 200 via proxy (2026-10-23).
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-23 检索）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (searched 2026-10-23); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

11 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 它自己不调模型——被 MCP 客户端调用的接口实现。它把 Firecrawl（「AI agent 的 context API」）的搜索/抓取/解析/浏览器交互包装成工具给 agent 用。
It calls no model itself — an interface implementation called by MCP clients. It wraps Firecrawl (the "context API for AI agents") search/fetch/parse/browser-interaction capabilities as tools for agents.
运行位置Where it runs: 官方托管远程 server（https://mcp.firecrawl.dev/v2/mcp）+ 本地 npx（firecrawl-mcp）双线；官方 README 明写支持 cloud 与 self-hosted。GitHub 7,673★，最近推送 2026-10-23（API 核验）；发版线在 npm（firecrawl-mcp 3.34.1），GitHub releases 停在 v3.2.1（2025-09-26）——落后于 npm。
Official hosted remote server (https://mcp.firecrawl.dev/v2/mcp) + local npx (firecrawl-mcp), two lines; the official README states support for both cloud and self-hosted. GitHub 7,673★, last push 2026-10-23 (API verified); the release line lives on npm (firecrawl-mcp 3.34.1), GitHub releases stopped at v3.2.1 (2025-09-26) — behind npm.
本地文件Local files: 不读写本机文件——输出网页正文（markdown/结构化 JSON）与解析产物（PDF 等文档）。
Reads and writes no local files — it outputs web page text (markdown/structured JSON) and parsed artifacts (PDF and other documents).
关机后的任务Tasks after shutdown: 托管线在 Firecrawl 侧；monitor 工具族可在 Firecrawl 侧按日程盯页面变更并出 diff/告警—— 这是「服务端任务」不是本机后台。
The hosted line lives on Firecrawl's side; the monitor tool family can watch page changes on a schedule on Firecrawl's side and emit diffs/alerts — that is a "server-side task", not a local background process.
工具与扩展Tools & extensions: 完整 profile 默认注册 27 个工具（README 口径，含 feedback 工具； FIRECRAWL_NO_SEARCH_FEEDBACK/NO_ENDPOINT_FEEDBACK 环境变量可减）： scrape / search / parse / crawl / map / interact（真实浏览器点击·填表·导航）/ monitor 族 / credit_usage / agent 任务 / developer·gov·research 垂直搜索索引族。托管 keyless 端点只暴露 3 个（scrape/search/parse）；search-only 端点固定 9 个。
The full profile registers 27 tools by default (README wording, including the feedback tool; the FIRECRAWL_NO_SEARCH_FEEDBACK/NO_ENDPOINT_FEEDBACK env vars can trim it): scrape / search / parse / crawl / map / interact (real-browser clicks·form filling·navigation) / the monitor family / credit_usage / agent tasks / the developer·gov·research vertical search index families. The hosted keyless endpoint exposes only 3 (scrape/search/parse); the search-only endpoint is fixed at 9.
上下文与记忆Context & memory: 无 server 侧跨会话记忆；interact 会话可经 scrapeId 跨调用续接（官方 README：不能从客户端单调用内逐步驱动，每调用跑完一个 prompt/code 回合即交还控制权）。
No server-side cross-session memory; interact sessions can continue across calls via scrapeId (official README: you cannot drive it step by step from within a single client call — each call runs one prompt/code round and hands control back).
权限与限制Permissions & limits: keyless 免费档可用（scrape/search/parse 三工具，限流）；其余工具要 API key。官方 README 三条硬口径：key 绝不进 server URL、绝不进 agent 聊天、只配在客户端或 secret manager。OAuth 端点（/v2/mcp-oauth）走交互式账户连接。
The keyless free tier works (scrape/search/parse three tools, rate-limited); the other tools need an API key. Three hard rules from the official README: the key never goes into the server URL, never into the agent chat, only into the client or a secret manager. The OAuth endpoint (/v2/mcp-oauth) uses interactive account connection.
适合什么任务Best for: 适合：要「搜索→抓正文→结构化 JSON→浏览器交互→站点级 crawl/map→页面监控」全链一面包齐的 agent；要 keyless 起步再升 key 的渐进接入。不适合：要长期持有浏览器会话逐步驱动（README 明写这不是它的交互模型）；只要单页读取的极简场景（fetch reference 更轻）。
Fits: agents that want the full chain covered in one package — "search→fetch text→structured JSON→browser interaction→site-level crawl/map→page monitoring"; gradual onboarding that starts keyless and upgrades to a key. Not a fit: holding a browser session long-term and driving it step by step (the README states plainly this is not its interaction model); minimal scenarios that only need single-page reading (the fetch reference is lighter).
传输方式Transport: streamable-http（托管 https://mcp.firecrawl.dev/v2/mcp；另有 /v2/mcp-oauth 与 /v2/mcp-search 两端点）+ stdio（npx -y firecrawl-mcp）双线；README 另提 SSE 支持。核验：仓 README（2026-10-23 直抓）。
streamable-http (hosted https://mcp.firecrawl.dev/v2/mcp; plus /v2/mcp-oauth and /v2/mcp-search endpoints) + stdio (npx -y firecrawl-mcp), two lines; the README also mentions SSE support. Verified: repo README (fetched directly 2026-10-23).
认证机制Authentication: keyless 免费档（限流）→ API key（Authorization Bearer，配客户端/secret manager） → OAuth 交互式连接（/v2/mcp-oauth）。本地 npx 线用 FIRECRAWL_API_KEY 环境变量。
Keyless free tier (rate-limited) → API key (Authorization Bearer, configured in the client/secret manager) → OAuth interactive connection (/v2/mcp-oauth). The local npx line uses the FIRECRAWL_API_KEY env var.
权限范围Permission scope: 读面为主（搜索/抓取/解析/监控）；interact 能在真实浏览器里点击填表—— 「读」与「在网页上动手」的分界要按目标站点的授权边界自己把关。无本机写面。
Mostly read (search/fetch/parse/monitor); interact can click and fill forms in a real browser — the boundary between "reading" and "touching things on a web page" must be gated yourself against the target site's authorization boundaries. No local write surface.

头号误解Top misconception

把 GitHub releases 当发版线——releases 停在 v3.2.1（2025-09），npm 已到 3.34.x，看 npmTreating GitHub releases as the release line — releases stop at v3.2.1 (2025-09) while npm is at 3.34.x; watch npm
把 keyless 档当全功能——托管 keyless 只有 3 工具，crawl/map/agent 等要 keyTreating the keyless tier as full-featured — hosted keyless exposes 3 tools only; crawl/map/agent etc. need a key
把 API key 塞进 server URL 或 agent 聊天——官方 README 明令禁止，只进客户端/secret managerPutting the API key in the server URL or the agent chat — the README forbids both; client config or secret manager only
把 interact 当可逐步驱动的浏览器会话——每调用跑完一个回合交还控制权，跨调用靠 scrapeId 续Treating interact as a step-by-step driven browser session — each call runs one prompt/code turn and returns control; persistence across calls goes through scrapeId
旧仓名 mendableai/firecrawl-mcp-server——已迁移到 firecrawl/ org（301 坐实），引用要换Referencing the old repo name mendableai/firecrawl-mcp-server — it moved to the firecrawl/ org (301 confirmed); update references

未知项清单Known unknowns

27 工具的完整参数面与 feedback 工具行为（README 有清单与裁剪口径，未逐条清点）
keyless 档与 key 档的速率/配额数值（README 只说 rate-limited；按 D9 与快照纪律不记数字）
self-hosted 部署线的功能边界（README 声明支持，未逐条核验）

The full parameter surface of the 27 tools and the feedback tool's behavior (the README has an inventory and trimming wording, not enumerated item by item)
Rate/quota numbers for the keyless vs keyed tiers (the README only says rate-limited; per D9 and snapshot discipline no numbers are recorded)
The functional boundaries of the self-hosted deployment line (README claims support, not verified item by item)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-23，内容更新日content updated on 2026-10-23。

类型Type	名称Name	链接Link
repo	firecrawl/firecrawl-mcp-server · 仓库（MIT；7,673★，最近推送 2026-10-23，GitHub API 核验；旧名 mendableai/ 301 迁移坐实）firecrawl/firecrawl-mcp-server · repository (MIT; 7,673★, last push 2026-10-23, GitHub API verified; old name mendableai/ 301 migration confirmed)	https://github.com/firecrawl/firecrawl-mcp-server
changelog	npm · firecrawl-mcp（发版线：3.34.1；GitHub releases 停在 v3.2.1）npm · firecrawl-mcp (release line: 3.34.1; GitHub releases stopped at v3.2.1)	https://www.npmjs.com/package/firecrawl-mcp
docs	Firecrawl · 官方文档 MCP 页（hosted 端点与客户端配置线）Firecrawl · official docs MCP page (hosted endpoints and client configuration line)	https://docs.firecrawl.dev/mcp-server

实测记录Measured runs

本档未在真实客户端逐项跑通；工具清单与权限来自官方 README。

This site has not completed measured runs. The test protocol is at protocols/mcp/_protocol.md.