ProofGraph MCP: Evidence-First Entity Graph Server

Section 1

The Core Problem It Solves: RAG Context Bloat & Hallucination

Traditional Vector RAG pipelines stuff unfiltered web pages and long markdown dumps into the model's prompt. This causes context saturation, sky-high inference bills, and hallucinated citations.

Traditional Raw RAG Pipeline

Unstructured Token Dumping

User Question ⟶ Vector Search / Web Fetch ⟶ Dump 5,000–20,000 Noisy Tokens ⟶ LLM Generates Answer

Why this fails in production:

High Token Costs & Sluggish Latency: Feeding entire articles, navbars, and boilerplate headers drives up API costs exponentially and increases Time-To-First-Token (TTFT).
Context Pollution ("Lost in the Middle"): Attention degrades as context windows fill with irrelevant paragraphs, causing agents to overlook the exact needle of truth.
Hallucinated Citations & Stitched Lies: Large language models hallucinate URL paths or weave conflicting statements into confident yet fictitious answers.

The ProofGraph Solution

Minimum Sufficient Context + Cryptographic Proof

Question ⟶ Entity Resolution ⟶ Claim Verification ⟶ Evidence Ranking ⟶ Token Optimizer ⟶ Compact Proof Packet (≤1,500 Tokens) ⟶ AI Agent

The Architectural Breakthrough:

Minimum Sufficient Context: Injects only the precise, corroborated atomic evidence nodes required to answer the prompt with zero filler tokens.
Cryptographic Provenance Chains: Every claim links to SHA-256 body hashes, canonical URLs, and immutable retrieval timestamps.
Objective Contradiction Handling: Surfaces conflicting facts side-by-side with confidence scores rather than guessing.

Section 2 & 3

Core Capabilities & Architectural Pillars

Built specifically for autonomous AI agents requiring deterministic truth, hardened network sandboxing, and strict token efficiency.

Capability	How It Works	Real-World Benefit
Deterministic Entity Resolution	Uses normalized string metrics (Jaro-Winkler) combined with domain, GitHub org, and package registry verification.	High Precision Resolves aliases (AI Build Infra, aibuildinfra) without false-positive entity merges.
Cryptographic Provenance	Every proof links to a source URL, verified retrieval timestamp, and SHA-256 body hash.	Audit Trail Answers "Why does the AI believe this?" with an unbroken, tamper-proof audit trail.
Contradiction Detection	If two sources conflict, ProofGraph marks claims as conflicting and shows both sides with respective confidence scores.	Zero Bias Prevents silent, arbitrary assumptions, biases, or hallucinated reconciliations.
Token Budget & Compression	Semantic deduplication clusters repeated excerpts into 1 primary quote + corroborating source counts.	~75%–90% Savings Achieves 75%–90% context token reduction while retaining complete, verifiable evidence.
Enterprise SSRF Protection	Hardened network layer blocking private subnets (RFC 1918), loopback (127.0.0.1), and cloud metadata endpoints (`169.254.169.254`).	Hardened Security Safe for fully autonomous AI agent execution without internal network intrusion vulnerabilities.

Section 4

The 12 Available MCP Tools in ProofGraph

AI agents (such as Google Antigravity, Claude Desktop, and Cursor) can call these 12 specialized tools natively through the Model Context Protocol.

Primary Retrieval Tool

get_evidence_packet

Returns the smallest evidence packet sufficient to answer a prompt within a strict token budget (default 1,500 tokens).

params: question, max_tokens, entity_id, claim

Verification

verify_claim

Verifies whether an assertion is supported, contradicted, conflicting, or unverified against indexed knowledge graphs.

params: claim, entity_id, sources

Provenance Audit

trace_claim

Generates an unbroken provenance chain:
Claim ⟶ Excerpt ⟶ Source ⟶ URL ⟶ Timestamp ⟶ SHA-256.

params: claim_id, format

Discovery

search_entities

Fuzzy and multi-signal search for canonical entities by legal name, developer alias, keyword, or domain.

params: query, limit, types

Entity Profile

get_entity

Retrieves canonical profile, normalized metadata, connected relationships, and valid Schema.org JSON-LD.

params: entity_id, include_schema

Graph Traversal

find_relationships

Explores 1-hop, 2-hop, or 3-hop relationship paths between entities with edge confidence weighting.

params: entity_a, entity_b, max_depth

Synthesis

explain_entity

Synthesizes a factual profile backed by verified sources, structured edges, and confidence indicators.

params: entity_id, perspective

Arbitration

compare_claims

Detects discrepancies between conflicting statements objectively, scoring citation authority across both sides.

params: claim_a, claim_b

Audit Footprint

analyze_web_presence

Audits an entity’s footprint across 8 platform categories (official domains, GitHub, package registries, documentation, etc.).

params: entity_id, categories

Diagnostics

audit_entity_consistency

Diagnostic tool checking cross-platform consistency across domains, GitHub orgs, npm scopes, and Schema.org graphs.

params: entity_id, platforms

Atomic Evidence

find_evidence

Retrieves compact, ranked supporting evidence for a specific assertion without document bloat.

params: assertion, max_results

Interoperability

export_graph

Exports verified knowledge graphs into standard enterprise formats: JSON, JSON-LD, CSV, or GraphML.

params: format, scope, entity_id

Section 5

Proven Benchmark Results: 90.7% Token Reduction

Evaluated against traditional raw document retrieval across standard enterprise and research benchmark queries.

Traditional Raw Context Dumping

~22,500

Tokens dumped per query into model prompt. High latency, high cost, context degradation.

ProofGraph Evidence Packets

~2,088

Tokens needed for comprehensive answer. Compact, deduplicated, and ranked evidence.

Total Context Reduction

90.7%

Token savings with 100% traceable, cryptographically grounded evidence precision.

Evidence Precision: 100% Traceable & Cryptographically Grounded with SHA-256 Hashes.

Section 6

Where ProofGraph Lives & Quickstart

Install directly via npm or connect to Google Antigravity, Claude Desktop, or Cursor using standard MCP JSON configuration.

MCP Identifier

io.github.AI-BuildInfra/proof-graph

Standard Protocol ID

Claude Desktop / Antigravity / Cursor mcp_config.json

{
  "mcpServers": {
    "proofgraph": {
      "command": "npx",
      "args": [
        "-y",
        "@aibuildinfra/proofgraph"
      ]
    }
  }
}

Install globally with npm: npm install -g @aibuildinfra/proofgraph

View Documentation

Looking for Human Design & Anti-AI Slop Tooling?

Explore HumanCraft UI MCP

Need Custom AI Agent Tooling or Enterprise Knowledge Graphs?

From specialized Model Context Protocol (MCP) servers to high-performance agentic RAG and ERPNext backbones, AI Build Infra architects production-grade infrastructure.

Book a Free Engineering Consultation Chat on WhatsApp

ProofGraph — Evidence-First Knowledge & Entity Graph MCP Server