ProofFetch

• CollectionSearch & Web Intelligence
• Classificationagent-native
• MCPAvailable
• Agent SkillUnavailable
• InterfacesURL onboarding · MCP · REST · A2A · openapi · http · API

• Website ↗
• Docs ↗
• Full dossier on GitHub ↗

What it does

ProofFetch fetches one public web page or text-layer PDF and returns it as numbered, citable paragraphs (p1, p2, …) so an agent can quote a specific paragraph by ID. Each response keeps the fetch time and two SHA-256 fingerprints: rawSha256 over the exact bytes the server sent and contentSha256 over the cleaned text.

Every fetch also runs a deterministic hidden-instruction (prompt-injection) scan and returns each match with its excerpt. A flag is a warning to review, not a safety guarantee; the text stays as the page had it.

Anyone can check a response with the open-source verifier, which recomputes both fingerprints from the page bytes and confirms every paragraph comes from that text in order. The service reads public resources only: no cookies, credentials, browser actions, JavaScript rendering, or private networks.

Agent-facing interfaces

• Interaction pattern: remote MCP + REST, with an llms.txt agent guide.
• Catalog protocols: URL onboarding, MCP, REST, A2A, openapi, http, API
• MCP: ✅ Available (hosted remote server)
• Agent Skill: ⚠️ Not yet published
• Onboarding type: url

How agents onboard

Interaction pattern: remote MCP + REST, with an llms.txt agent guide.

# Agent guide:
Read https://prooffetch.steellantern.com/llms.txt and follow the instructions.

# MCP (Claude Code):
claude mcp add --transport http proof-fetch https://prooffetch.steellantern.com/mcp

# REST: get a free key (1,000 extractions, shown once), then fetch a page
POST https://prooffetch.steellantern.com/v1/keys/free
POST https://prooffetch.steellantern.com/v1/evidence   (Authorization: Bearer <key>; body: url, idempotencyKey)

See the full onboarding section in the dossier.

Protocol surface