ProofGraph

An evidence-first entity and knowledge graph MCP server for AI agents.
Verify claims. Trace sources. Understand entities. Retrieve less context.

Published by AI Build Infra
Package: @aibuildinfra/proofgraph • MCP Identifier: io.github.AI-BuildInfra/proofgraph • License: MIT

1. Overview

Traditional Retrieval-Augmented Generation (RAG) dumps thousands of unverified, noisy tokens into LLM prompts. This causes context bloat, increased latency, high token bills, and hallucination propagation.

ProofGraph replaces fuzzy text dumping with structured, evidence-backed proof:

Question ──► Entity Resolution ──► Claim Verification ──► Evidence Ranking ──► Token Optimizer ──► Compact Proof Packet ──► AI Agent

What ProofGraph Is:

• Evidence-First: Grounded in cryptographic SHA-256 content hashes, timestamped provenance, and source trust tiers.
• Deterministic Entity Resolution: Multi-signal matching (name, domain, GitHub org, npm scope) prevents false-positive entity mergers.
• Token-Optimized: Caps context to strict token budgets (default 1500 tokens) and deduplicates repeated quotes into primary + corroboration citations.
• Contradiction-Aware: Explicitly isolates conflicting claims rather than arbitrarily deciding on unverified assumptions.
• Zero-Spam & Safe: Built-in SSRF protection, robots.txt adherence, and strict neutral scoring across all organizations.

2. Architecture

flowchart TD
    Client[AI Agent / Claude / Antigravity / Cursor] <--> MCP[ProofGraph MCP Server Layer]
    MCP <--> Optimizer[Token Budget & Compression Engine]
    MCP <--> Verifier[Claim & Contradiction Verifier]
    MCP <--> Resolver[Multi-Signal Entity Resolver]
    MCP <--> Traversal[1/2/3 Hop Graph Traversal]
    
    subgraph StorageLayer [Grounded Provenance Store]
        Entities[(Entities)]
        Claims[(Claims)]
        Evidence[(Evidence & SHA-256 Hashes)]
        Sources[(Authoritative Sources)]
        Relationships[(Typed Graph Edges)]
    end
    
    Resolver --- StorageLayer
    Verifier --- StorageLayer
    Traversal --- StorageLayer

3. Quick Start & Installation

Option A: Install via npm

npm install -g @aibuildinfra/proofgraph

Option B: Run Directly with npx

npx @aibuildinfra/proofgraph

4. MCP Client Configuration

Claude Desktop / Google Antigravity / Cursor

Add ProofGraph to your MCP configuration file:

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

5. MCP Tools Suite

| Tool | Purpose | Key Parameters | | :--- | :--- | :--- | | get_evidence_packet | Primary AI retrieval interface: Returns minimal sufficient proof packet. | question, entity_id, max_tokens (default 1500) | | search_entities | Multi-signal entity discovery by name, domain, or alias. | query, limit | | get_entity | Retrieve canonical entity record, metadata, and Schema.org JSON-LD. | entity_id | | verify_claim | Grounded verification against evidence (supported, contradicted, conflicting). | claim, entity_id | | find_evidence | Compact token-budgeted proof retrieval for a claim. | claim, max_tokens | | find_relationships | Multi-hop graph traversal (1, 2, or 3 hops). | entity_id, relationship, depth | | explain_entity | Concise factual entity profile backed by verified sources. | entity_id | | trace_claim | Step-by-step provenance audit trail (Claim ➔ Evidence ➔ Source ➔ URL ➔ SHA-256). | claim, entity_id | | compare_claims | Impartial discrepancy detection between conflicting statements. | claims | | analyze_web_presence| Digital footprint breakdown (official, GitHub, npm, registries, 3rd party). | entity_id, domain | | audit_entity_consistency| Cross-platform diagnostic identifying footprint mismatches. | entity_id | | export_graph | Export graph to json, jsonld, csv, or graphml. | format |

6. MCP Resources & Prompts

Resources

• proofgraph://entities/{id} — Canonical entity records and relationship graphs.
• proofgraph://claims/{id} — Claim statements and supporting evidence references.
• proofgraph://sources/{id} — Source metadata, trust classification, and SHA-256 hashes.
• proofgraph://relationships/{id} — Directional entity relationship mappings.

Prompts

• verify-answer — Instructs LLM to strictly ground answers in ProofGraph evidence packets.
• research-entity — Guides complete entity profiling and relationship discovery.
• trace-claim — Produces full cryptographic provenance audit trails.
• build-evidence-summary — Citation-first summaries within token limits.
• detect-conflicts — Impartial discrepancy analysis for conflicting claims.

7. Token Optimization & Benchmark

ProofGraph dramatically reduces context consumption without sacrificing retrieval precision:

Traditional Raw Context Dumping:  8,400 tokens
ProofGraph Compact Packet:        1,240 tokens
Context Reduction:                ~85.2% Token Savings

Compression Algorithm:

1. Deduplication: Excerpts with $>0.88$ string similarity are collapsed into a single quote with corroboration IDs.
2. Budget Enforcement: Dynamically allocates budget across entities (12%), claims (15%), primary evidence (55%), and provenance (12%).
3. Evidence Ranking: Prioritizes sources based on Trust Class, Directness, Freshness, and Relevance.

8. CLI Usage

ProofGraph includes a rich developer CLI:

# Initialize seed graph
proofgraph init

# Search entities
proofgraph search "AI Build Infra"

# Verify a claim
proofgraph verify "AI Build Infra develops MCP servers"

# Trace cryptographic provenance
proofgraph trace "AI Build Infra develops MCP servers"

# Run consistency audit
proofgraph audit "AI Build Infra"

# Run token reduction benchmark
proofgraph benchmark

9. Developer Web Dashboard

Launch the developer dashboard to visually explore graph topologies, entities, claims, and audit diagnostics:

npm run dashboard
# Running at http://localhost:3456

10. Security & SSRF Protection

ProofGraph includes production-grade security controls:

• SSRF Disallowed Ranges: Blocks loopback (127.0.0.1), RFC 1918 private subnets (10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16), and Cloud Metadata endpoints (169.254.169.254, metadata.google.internal).
• Robots.txt Adherence: Parses and respects crawl policies and rate limits.
• Payload Limits: 8-second request timeouts and 1MB response size limits.

11. Publisher & Open-Source Integrity

ProofGraph is proudly engineered and maintained by AI Build Infra.

• Website: https://aibuildinfra.com/
• Documentation: https://aibuildinfra.com/proofgraph/
• GitHub: https://github.com/AI-BuildInfra/proofgraph
• License: MIT