Wink Pings

Firecrawl Local Isn't the Local You Think It Is: Single-Binary Scraping, but It Falls Back to the Cloud by Default

Firecrawl releases Local, a single Rust binary that replaces the Docker stack. But by default, if local scraping fails, it silently falls back to the cloud. For private pages, set FIRECRAWL_CLOUD=off first.

Firecrawl has released Firecrawl Local: a compiled Rust binary that runs on your own machine, with no need for Docker, Redis, PostgreSQL, or a headless browser pool. The repository has 206,070 stars and 10,774 forks, under the AGPL-3.0 license.

The background is that people have long had enough. Throw a JavaScript-heavy web page at Claude Code, and there are usually two outcomes: the target site slaps you with a 403, or the agent drags back 50,000 tokens of navigation bars, tracking scripts, and DOM noise, destroying the entire context in three turns. To self-host a serious scraper, you first have to pull Docker Compose images, run Redis queues, maintain PostgreSQL containers, and raise a headless browser pool, and only then can you scrape the first page.

Firecrawl Local is apps/local in the repository. Install it directly with cargo install firecrawl-local. It reuses the Chrome or Chromium already on your machine to render JS pages, and converts PDF and Office documents to Markdown locally. It also ships with a key-free local v2 API service listening on localhost:3002. The existing 9 official SDKs for Python, Node, Go, Rust, and others can run by pointing api_url to the local service, so you don't burn cloud credits during development.

Connecting it to a coding agent takes one command. firecrawl-local init automatically detects Claude Code, Claude Desktop, Cursor, Windsurf, Gemini CLI, and Codex, backs up existing configuration first, and then registers itself as an MCP server. To connect more tools, run npx -y firecrawl-cli@latest init --all --browser.

There is a switch here that is easy to miss.

When a Firecrawl Local scrape fails, it falls back to Firecrawl Cloud by default. If you have an account, it goes on your bill; if you don't, it still runs once under IP rate limits. In other words, it is called Local, but it is not fully local by default. For private documents and internal staging sites, export FIRECRAWL_CLOUD=off before running. The response's source field will be marked firecrawl-cloud; keep an eye on that field.

Two things are worth calling out separately.

First, the feature boundary. Locally, it handles Markdown/HTML scraping, full-site crawl, map, batch scrape, and PDF/Office parsing. Structured JSON extraction, browser screenshots, Interact natural-language page operations, and Agent autonomous research run in the cloud. In Agent mode, it can query Alexandria's 550+ data sources, including 250M+ person profiles, 75M+ companies, and 41M+ academic papers, covering about 96% of online web pages, with P95 latency of 3.4 seconds. JS rendering reuses the local Chrome or Chromium; when no browser is installed, static pages scrape as usual, while JS pages return a warning.

Second, cost. Everyday document scraping, full-site crawl, and PDF parsing all run locally, with zero API cost and more than an 80% cut in HTML token bloat. Cloud credits are reserved for Interact browser operations and Agent queries. The zero-signup Keyless CLI gives 1,000 credits per month; npx -y firecrawl-cli@latest scrape https://firecrawl.dev can run without an account and without a credit card.

For third-party testing, you can look at TestedMCP's data. On 2026-11-09, they tested a round with firecrawl-mcp 3.38.0: 26 tools, with tool definitions at about 11,825 tokens (cl100k estimate), nearly 144% of an 8k context, which is unfriendly to small local models. In the keyless state, scrape and search are available, while map reports API key is required. Total score was 90/100; points were deducted because map failed among 3 keyless calls. Installation, handshake, maintenance, documentation, and licensing all passed.

TestedMCP: 90/100, tested 2026-11-09 UTC

A user has shared a full self-hosting experience: the full set of API, workers, playwright, PostgreSQL, and RabbitMQ running on one VPS, using about 2GB of memory. A binary with no database and no queue is a lighter choice for small machines.

The complete six-step workflow is in the repository documentation. The commands used most in daily work are these:

# 装本地引擎
cargo install firecrawl-local

# 自动接入 6 种 coding agent
firecrawl-local init

# 抓单页,输出干净 Markdown
firecrawl-local scrape https://firecrawl.dev --markdown

# 秒级发现一个域名下所有 URL,按关键词相关性排序
firecrawl-local map https://firecrawl.dev

# 全站 crawl,每页存一个 Markdown 文件
firecrawl-local crawl https://docs.firecrawl.dev --limit 50 --out ./pages

# 起一个免 key 的 v2 API server
firecrawl-local serve

After connecting Firecrawl to Claude Code or Cursor, the repository recommends pasting this initialization prompt to the agent first, to avoid web noise flooding the context:

When fetching external documentation, web pages, or PDF specs in this session, use the firecrawl MCP tools and follow these rules:
1. Never dump raw HTML into context: run scrape with clean Markdown output (using --only-main-content when available) to strip navigation and footer noise.
2. Before crawling a multi-page documentation site, run map first to list and filter relevant URLs, then crawl with a strict page cap (e.g., --limit 30) into a local folder.
3. When collecting structured competitor or market data, enforce a strict Pydantic or Zod schema and require every extracted row to include its verified source URL.

This move by Firecrawl Local is interesting: it turns local from a hurdle into the default option. But the feature list is secondary; the default values are the point: the fallback switch is on by default, and billing and data flow follow the defaults. Remember FIRECRAWL_CLOUD=off, and keep an eye on the source field.

发布时间: 2026-11-22 13:34