Web Fetch Mcp

web-fetch-mcp

MCP server that fetches any URL and returns clean Markdown — not raw HTML.

[](https://github.com/jdevalk/web-fetch-mcp/actions/workflows/ci.yml) [](LICENSE) [](https://bun.sh) [](https://www.typescriptlang.org)

Claude Code's built-in WebFetch returns raw HTML. web-fetch-mcp tries two strategies in order:

1. Content negotiation — sends Accept: text/markdown and returns the body if the server honours it (fast, zero browser overhead).
2. Headless browser — launches Obscura, navigates to the URL, and extracts Markdown via the LP.getMarkdown CDP method. Falls back to DOM.getOuterHTML + turndown if the CDP result is too sparse.

Results are cached in-memory for 15 minutes.

---

Quick start

1. Clone and install

git clone https://github.com/jdevalk/web-fetch-mcp.git
cd web-fetch-mcp
bun install

2. Download the Obscura binary for your platform

bun run setup

3. Register with Claude Code

claude mcp add web-fetch-mcp -- bun run /path/to/web-fetch-mcp/index.ts

To make Claude always prefer this server over the built-in WebFetch, add "denyTools": ["WebFetch"] to ~/.claude/settings.json.

---

Tool reference

fetch

Fetch a URL and return its content as clean Markdown.

| Parameter | Type | Default | Description | |---|---|---|---| | url | string | — | The URL to fetch | | force_browser | boolean | false | Skip content negotiation and go straight to the headless browser |

Response shape:

{
  markdown: string;
  url: string;       // final URL after redirects
  title: string;     // page  or first H1
  contentType: "negotiated" | "browser" | "fallback";
}

---

How it works

Strategy 1 — content negotiation

Sends Accept: text/markdown, text/plain;q=0.9, */*;q=0.5. If the server responds with content-type: text/markdown or text/plain, the body is returned directly without launching a browser. Most servers don't honour this, but those that do (e.g. joost.blog) get a near-instant response.

Strategy 2 — Obscura CDP

Obscura is launched once as a persistent subprocess on port 9222. Each fetch opens a fresh browser target so pages are fully isolated and don't share cookies or state. After navigation, LP.getMarkdown walks the DOM into structured Markdown. If the result is too sparse (< 50 chars), the server falls back to DOM.getOuterHTML` + turndown.

Cache — 15-minute TTL

A 15-minute in-memory TTL cache (matching Claude Code's built-in WebFetch) avoids duplicate fetches within a session.

---

Platform support

| Platform | Status | |---|---| | macOS arm64 | ✅ | | macOS x8664 | ✅ | | Linux x8664 | ✅ | | Linux aarch64 | ✅ |

---

Contributing

See [CONTRIBUTING.md](CONTRIBUTING.md).

Security

To report a vulnerability, see [SECURITY.md](SECURITY.md).

License

MIT — see [LICENSE](LICENSE).

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jdevalk
• Source: jdevalk/web-fetch-mcp
• License: MIT

Install and usage instructions live in the source repository linked above.