web

web

WebFetchRaw: a web page's own content in a file, where WebFetch gives a small model's answer about it.

Claude Code's WebFetch converts the page to markdown, hands it to a small model with your prompt and returns that model's reply. Asking it for the page verbatim does not get the page. On 2026-10-13, asked for example.com's text "verbatim, character for character", it answered:

I can't reproduce the full page verbatim because quotes are limited to 125 characters, but here is the closest I can provide:
"Example Domain. This domain is for use in documentation examples without needing permission."
The page also says the domain isn't a service and shouldn't be used for testing or monitoring.

The page says "This is not a service; avoid relying on it for testing and monitoring purposes." WebFetchRaw writes the page itself to a file and returns one line:

ok: wrote 34038 bytes to /tmp/…/webfetch-raw/docs.python.org/20261012-221501-3_library_json.html-1a2b3c4d.md (markdown, text/html)

The page enters the context only when the agent Reads that file, whole or in parts.

   the model calls mcp__web__WebFetchRaw { url, path?, output_format?, … }
      └─ hooks/register.ts answers the call itself (a Claude Mod: no MCP server)
           └─ $.process.run: uv runs scripts/webfetch_raw.py, the input as JSON on stdin
                ├─ an HTML page → trafilatura → markdown · text · cleaned html (main content, metadata on top)
                ├─ anything else (JSON, text, markdown, a PDF), or output_format "raw" → the body byte for byte
                └─ stdout: the one ok: line → the tool's result · stderr: error: … → an error result

• In the tool list from the start (isDeferred: false), beside WebFetch, so the model reaches for it when it needs exact words, code, numbers or links. Asked for "the exact wording of the paragraph on example.com", a claude -p session picked it over WebFetch without being told.
• Not in this module: the fetch and the extraction. A hooks module has no npm and no Node for an HTML extractor, and $.http.fetch returns text only (a PDF would not survive), so a Python script does both, one process per call; nothing runs between calls.
• Where the file goes: path (absolute; folders are made), else <tmp>/webfetch-raw/<host>/<time>-<slug>-<hash>.<ext>.

It replaces an earlier MCP server, webfetch-verbatim (tool WebFetchVerbatim). Changed on the way: no MCP server; raw is the body as sent (it was prefixed with a header line, and decoded as text); content that is not HTML is saved as sent, with its own extension; the byte count is bytes.

Needs

uv, and CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 in the env of your Claude Code settings (ak setup switches it on with this plugin; installed with /plugin install web@agentic-kit, set it yourself). The first call builds the plugin's environment (trafilatura, lxml): a few seconds, once. A call goes through Claude Code's permission check like any tool's; mcp__web__WebFetchRaw in permissions.allow lets it run without asking.

Check it

uv run pytest -q                                  # in plugins/web: the script, against a local HTTP server
claude plugin test plugins/web                    # the mod: registration, the call, its errors
claude plugin validate plugins/web