ProofFetch: web fetch receipts for AI agents

Three tools, one fetch engine

What the "proof" is (and isn't)

• content_sha256 is the hex SHA-256 of the raw response body bytes. For static pages, you can recompute it yourself: curl -s https://example.com/ | sha256sum gave the same hash as the example below.
• Dynamic pages (ads, CSRF tokens, timestamps) return different bytes on every request, so their hash changes from one fetch to the next. Plan for that when you use UrlWatch.
• The receipt is not cryptographically signed or notarized by a third party. It's a structured, hash-anchored record of one fetch that your agent can log, diff and cite.

Real request and response

These were captured from the live server on 2026-10-08. The output is unedited apart from line breaks.

curl -sS -X POST https://proofetch.hardenbranch.workers.dev/v1/proof \
  -H 'Content-Type: application/json' \
  -d '{"url":"https://example.com"}'

{
  "ok": true,
  "product": "ProofFetch Lite",
  "fetched_at": "2026-10-08T17:20:44.976Z",
  "content_sha256": "25ddf2c883e0d1958ea971d279a7e4f0fd446724ee3db7db19dadabd4a62e484",
  "title": "Example Domain",
  "text_markdown": "This domain is for use in documentation examples without needing permission. This is not a service; avoid relying on it for testing and monitoring purposes.",
  "status_code": 200,
  "final_url": "https://example.com/"
}

ReadClean on the same URL (POST /v1/readclean):

{
  "ok": true,
  "product": "ReadClean",
  "title": "Example Domain",
  "text_markdown": "This domain is for use in documentation examples without needing permission. This is not a service; avoid relying on it for testing and monitoring purposes."
}

UrlWatch with the previous hash (POST /v1/urlwatch, body {"url":"https://example.com","previous_sha256":"25dd…e484"}):

{
  "ok": true,
  "product": "UrlWatch",
  "fetched_at": "2026-10-08T17:20:45.377Z",
  "content_sha256": "25ddf2c883e0d1958ea971d279a7e4f0fd446724ee3db7db19dadabd4a62e484",
  "changed": false,
  "previous_sha256": "25ddf2c883e0d1958ea971d279a7e4f0fd446724ee3db7db19dadabd4a62e484"
}

A blocked request (cloud metadata IP) returns HTTP 400 and isn't billable:

{
  "ok": false,
  "product": "ProofFetch Lite",
  "error": {
    "code": "ssrf_blocked",
    "message": "blocked_ip"
  },
  "chargeable": false
}

MCP setup (Claude, Cursor, any MCP client)

The server speaks MCP over Streamable HTTP (JSON-RPC 2.0, protocol 2025-03-26) at https://proofetch.hardenbranch.workers.dev/mcp. This page covers its three read-only tools: proof_fetch, read_clean and url_watch. The same endpoint also exposes two newer tools, site_crawl and sec_section. They aren't on the marketplaces yet; see /llms.txt.

Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "proofetch": { "url": "https://proofetch.hardenbranch.workers.dev/mcp" }
  }
}

Claude Code:

claude mcp add --transport http proofetch https://proofetch.hardenbranch.workers.dev/mcp

Claude Desktop (claude_desktop_config.json, through the mcp-remote bridge):

{
  "mcpServers": {
    "proofetch": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://proofetch.hardenbranch.workers.dev/mcp"]
    }
  }
}

Metered through a marketplace (stable listing, one wallet or key for many tools):

# Agent Wonderland MCP (see docs.agentwonderland.com/consumers/setup)
{ "mcpServers": { "agentwonderland": { "command": "npx", "args": ["-y", "@agentwonderland/mcp"] } } }
# then ask your agent to run:
run_agent({ slug: "proof-fetch-lite", input: { "url": "https://example.com" } })

# Cracked (REST; also exposed as run_tool in https://cracked.ai/mcp)
curl -s https://cracked.ai/v1/run -H "Authorization: Bearer $CRACKED_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"provider":"cp-prooffetch-lite-2","endpoint":"/post/v1/proof","input":{"url":"https://example.com"}}'

Raw JSON-RPC check:

curl -sS -X POST https://proofetch.hardenbranch.workers.dev/mcp -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Pricing

You pay per successful call, at the price the marketplace lists. Failed calls (invalid URL, blocked address, DNS failure, timeout, body over 2 MB) return HTTP 4xx/5xx with "chargeable": false and aren't charged.

Prices as listed on 2026-10-08. Cracked's listing page says it adds a $0.001 platform fee per successful run. Agent Wonderland charges $0.10 per request for each tool, so Cracked or run.pay is the cheaper way to call ReadClean and UrlWatch. Full run.pay service IDs: ProofFetch 9a326c8a-bbac-4b11-9750-8a4fe52e22de, ReadClean 38c95c4f-ead6-4c76-a0f3-e77c70c5b262, UrlWatch bf603002-2584-47f8-97fb-ce1927cba95b.

Limits and guarantees

• Public http/https URLs only. No other schemes, and no user:pass@ credentials in the URL.
• SSRF guards: blocks localhost, .local/.internal names, private, loopback, link-local, CGNAT and test ranges, IPv6 ULA/link-local, and cloud metadata (169.254.169.254). Hostnames are resolved over DNS-over-HTTPS and rejected if any answer is private. Every redirect (max 5) is checked again.
• 15 second timeout, 2 MB maximum body, Markdown capped at 200,000 characters.
• One plain HTTP GET. No JavaScript rendering, no screenshots, no headless browser, no proxies, no logins or cookies. Pages that render client-side may come back mostly empty.
• Non-HTML text (JSON, XML, plain text) comes back in a code block. Binary content gets a short note with the byte count and content type, and the SHA-256 still covers the raw bytes.

FAQ

Is a 404 page billed? If the fetch completes, the call succeeds and status_code reports what the target returned (for example, 404). That counts as a successful call. Only errors on our side (blocked, invalid, DNS, timeout, too large) are non-billable.

Does it check robots.txt? No. It doesn't crawl. It makes one on-demand GET for the single URL your agent passes, with the User-Agent ProofFetch-Lite/1.0.0. You're responsible for fetching only pages you're allowed to.

Can it reach my intranet? No. Private and internal addresses are blocked by design.

Machine-readable docs? /llms.txt, /openapi.yaml, /health.

Contact

Questions, bulk use or a feature you need: ktstepp85@gmail.com (Kyle Stepp).