dsh-webfetch
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
dsh-webfetch
A zero-dependency webfetch MCP server: fetches any http(s) URL and returns clean content — text, simplified Markdown, raw HTML, or formatted JSON. Built for DeepSeek Harness (dsh), works with any MCP client that supports stdio servers.
Zero dependencies: single
server.js, hand-rolled stdio JSON-RPC 2.0 — nonpm install, no SDK version conflicts.Model-friendly output: main-content extraction (
<main>/<article>), entity decoding, absolute links, image alt text, chunked reading.Polite & safe:
robots.txtcompliance, Cloudflare-challenge retry, SSRF protection, bounded 5 MiB reads.
Design references
Compared three existing implementations (2026-08) and adopted the useful parts with zero-dependency equivalents:
Source | Adopted design |
Official mcp-server-fetch |
|
OpenCode | Format-aware |
Main-content-first (lightweight Readability heuristic), SSRF protection, relative-link absolutization, image alt |
Deliberately not adopted: @mozilla/readability, turndown, undici, unpdf — dependency-free equivalents cover the common cases.
Related MCP server: superFetch MCP Server
Quick start
Requires Node.js ≥ 20 (uses global fetch).
git clone https://github.com/withlovehub/dsh-webfetch.git
cd dsh-webfetch
# Optional but convenient: puts the `dsh-webfetch` command on your PATH
npm link # or: npm install -g .
node test.js # 32 in-process test cases — all green means your Node is readyThe server speaks newline-delimited JSON-RPC 2.0 on stdio. Once published to npm, the one-liner is npx -y dsh-webfetch.
Smoke-test the server by hand (expect an initialize response and a tools/list response on stdout):
printf '%s\n' \
'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"probe","version":"0"}}}' \
'{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}' \
| node server.jsPowerShell equivalent:
@(
'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"probe","version":"0"}}}',
'{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}'
) | node server.jsClient setup
The server is a plain stdio MCP server — any MCP client works. Pick your client below; every example uses node + the path to server.js. Two shortcuts make life easier:
Path style: use forward slashes even on Windows (
C:/dev/dsh-webfetch/server.js) — Node accepts them and you avoid JSON backslash escaping.After
npm link: setcommandtodsh-webfetchand dropargsentirely.
Most clients require an absolute path; ${workspaceFolder}-style variables only work where documented (VS Code).
DeepSeek Harness
File: ~/.dsh/cordis.patch.yml
- insert:
- id: mcp-webfetch
name: '@deepseek-ai/dsh-mcp-client'
config:
serverName: web
transport: stdio
command: node
args:
- '/path/to/dsh-webfetch/server.js'Restart dsh web; the model gains the mcp__web__fetch tool.
Claude Desktop
File: claude_desktop_config.json (Windows: %APPDATA%\Claude\, macOS: ~/Library/Application Support/Claude/, Linux: ~/.config/Claude/). Restart Claude afterwards.
{
"mcpServers": {
"webfetch": {
"command": "node",
"args": ["C:/path/to/dsh-webfetch/server.js"]
}
}
}Claude Code
File: .mcp.json in your project root, or add it from the CLI:
claude mcp add webfetch -- node /path/to/dsh-webfetch/server.js{
"mcpServers": {
"webfetch": {
"type": "stdio",
"command": "node",
"args": ["/path/to/dsh-webfetch/server.js"]
}
}
}Verify with claude mcp list.
Cursor
File: .cursor/mcp.json in your project (or Cursor Settings → MCP → Add new MCP server).
{
"mcpServers": {
"webfetch": {
"command": "node",
"args": ["C:/path/to/dsh-webfetch/server.js"]
}
}
}Verify in Cursor Settings → MCP (green status dot).
VS Code Copilot
File: .vscode/mcp.json — supports the ${workspaceFolder} variable, so a cloned repo can self-configure:
{
"servers": {
"webfetch": {
"type": "stdio",
"command": "node",
"args": ["${workspaceFolder}/server.js"]
}
}
}Verify: Command Palette → MCP: List Servers.
OpenCode
File: opencode.json (project) or global config.
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"servers": {
"webfetch": {
"type": "local",
"command": ["node", "/path/to/dsh-webfetch/server.js"]
}
}
}
}Verify with opencode2 mcp list (v1: opencode mcp list).
Anything else
Any MCP client that speaks stdio works with the same command + args shape — the protocol is universal.
Tool parameters
Parameter | Type | Default | Description |
| string (required) | — | Full URL, http/https only ( |
| text / markdown / html / json | text | text = cleaned content; markdown = links/headings/images kept; html = raw HTML; json = pretty-printed JSON |
| 1000–200000 | 60000 | Maximum characters returned this call |
| 0–10000000 | 0 | Start content from this character index — pair with |
| 3000–120000 | 20000 | Request timeout |
Behavior
Main-content first: if the text of a
<main>/<article>block is ≥ 30% of the page, only that block is returned — nav/footer noise is filtered (lightweight Readability).HTML cleaning: script/style/svg dropped, block tags become newlines, entities decoded,
<title>extracted.Markdown mode: relative hrefs/srcs absolutized against the final URL (
javascript:/mailto:degrade to plain text), images become, h1–h6 become ATX headings; links wrapping images keep the alt text.JSON APIs: auto
JSON.parse+ indent; invalid JSON returnsisErrorwith a preview instead of silently passing raw text.Binary: returns metadata only (status/content-type/bytes) — bytes are never fed into the model context.
robots.txt: respected by default (
dsh-webfetchgroup, then*group, longest-prefix match); blocked URLs returnisError; results cached 30 min; setDSH_WEBFETCH_IGNORE_ROBOTS=1to disable. EmptyDisallow:means allow-all per RFC 9309.Cloudflare challenge: on
403+cf-mitigated: challenge, retries once with a transparent UA.SSRF protection: rejects
metadata.google.internal,169.254.*,0.0.0.0.Bounded reads: max 5 MiB per response (declared limit → error; undeclared → streaming cutoff).
Errors: failures return MCP
isError: true; diagnostics go to stderr only (stdout is the protocol channel).
Known limitations
No JavaScript execution — SPAs return the raw shell, not the rendered app.
No PDF extraction / true Readability / proxy support (zero-dependency line; add when actually needed).
Malformed HTML with quotes inside hrefs degrades to plain text rather than crashing.
License
MIT
This server cannot be deployed
Maintenance
Related MCP Connectors
Web scraping MCP server: any URL to Markdown or JSON, with JavaScript rendering and batch jobs.
MCP server (stdio): fetch web pages as clean readable markdown via the AgentForge API
Free remote MCP server for fetching public web pages through a rotating proxy pool.
Hosted MCP server: render any URL to print-stable PDF, PNG, WebP or Markdown
Related MCP Servers
- AlicenseAqualityDmaintenanceA powerful MCP server for fetching and transforming web content into various formats (HTML, JSON, Markdown, Plain Text) with ease.47,548 npm42MIT
- AlicenseAqualityCmaintenanceAn MCP server that fetches web pages and extracts clean, AI-friendly Markdown content using Mozilla Readability. It provides secure web access for LLMs with built-in SSRF protection and automated content cleaning for improved context retrieval and summarization.1515 npmMIT
- AlicenseNot gradedqualityCmaintenanceA lightweight MCP server for parsing HTML, fetching URLs, rendering terminal-style screenshots, and executing JavaScript on static HTML without external dependencies.3MIT
- AlicenseNot gradedqualityCmaintenanceA lightweight MCP server that fetches URLs and returns clean, readable markdown, enabling AI assistants to read any webpage with article extraction, caching, and security features.3MIT