MCP-native evidence graphs for EVM + Solana

The security graph behind AI-assisted smart contract audits.

ilold maps authority, state, value flow, call paths, traces, findings, and fix proof into a domain-specific knowledge graph for smart contract security: an audit evidence graph humans inspect visually and agents query through MCP.

Open EVM repo See the workflow

audit evidence graph · evm risk path closed

entrypointpublic call missing guardauthority gap value pathfunds reachable fix proofpath closed

code -> risk -> fix proof

Current work spans working analyzers and deployable review workflows

Source index Evidence graph MCP server Fix proof

Audit workflow

From codebase to evidence, without losing the path.

ilold connects the pieces that usually live in separate tools: source, traces, findings, reports, and the fix that claims to close the issue. Every step remains inspectable by humans and queryable by agents.

Index source

AST, imports, modifiers, storage writes, ABI surfaces, instructions, and reachable entrypoints.

Build graph

Authority edges, call paths, state dependencies, value movement, traces, and detector output.

Query through MCP

Agents ask precise questions like which public paths reach funds or which signer protects a transfer.

Export proof

Reports link findings back to source spans, runtime evidence, graph paths, and verified fixes.

MCP-native

Give AI agents security context they can actually use.

Most AI coding tools start by searching raw files. ilold starts from a deterministic security graph. Agents query the knowledge graph through narrow MCP resources and tools instead of rebuilding context from raw files.

ilold.mcp security context server

tool which_paths_reach_funds()

tool explain_authority("transfer")

tool trace_value_flow("USDC")

tool verify_fix("finding-42")

source spans · traces · graph paths · report proof

Agents query facts

Claude Code, Codex, Cursor, and Copilot-style clients can ask security questions through MCP instead of rebuilding context every run.

Auditors inspect paths

The same graph is visual: authority, state, value, calls, traces, and reports stay reviewable by humans.

Findings become evidence

A report item links back to source spans, traces, affected paths, and the proof that a fix closed the route.

Critical assets

Map every path to funds and privileged control.

Static analyzers are useful for known patterns. The expensive failures are usually violated assumptions: who can call what, which state changed, where value moved, and whether a fix actually broke the exploit path. ilold makes those relationships reviewable by humans and queryable by MCP-connected agents.

Funds

Withdrawals, token transfers, treasury movement, accounting deltas, and paths to balances.

Authority

Owners, roles, signers, upgrade controls, admin routes, and authorization assumptions.

State

Storage writes, account changes, configuration, invariants, and dependency edges.

External behavior

Calls, delegatecalls, CPI, oracles, bridges, token programs, and cross-system effects.

Fix proof

Changed paths, closed exploit routes, regression evidence, and report-linked verification.

Technical architecture

A security evidence layer, not another chatbot.

ilold is designed as infrastructure: adapters collect chain and source context, the graph builder normalizes security relationships, the evidence store preserves provenance, and MCP makes the graph available to agents, auditors, CI, and reports. Under the hood, contracts, functions, modifiers, storage slots, signers, token flows, traces, findings, and fixes become typed nodes and edges with source provenance.

Adapters EVM / Solana

source, ABI, traces, instructions, detector output

Graph builder Security relationships

authority, calls, state, value flow, findings

Evidence store Queryable provenance

nodes, edges, source spans, report history, fix proof

Interfaces Humans + agents

visual explorer, MCP server, exports, CI/API

EVM graph engine

Dependency graphs and path evidence for Solidity systems.

ilold-evm indexes Solidity projects into graph-backed review state: dependency graphs, call paths, traces, slices, sessions, findings, report export, and MCP tools for Claude Code, Codex, Cursor, and security automation.

Path analysisfollow functions, modifiers, calls, and affected code slices

Finding contextconnect detector output to reachable paths and reports

Agent accesslet Claude Code, Codex, Cursor, and other clients query graph facts

EVM dependency graph in ilold — EVM dependency graph and review context

Solana graph canvas in ilold — Solana scenarios, account diffs, CPI context

Current Solana workbench

Runtime evidence for accounts, instructions, and CPI behavior.

The Solana workbench adds execution context: LiteSVM scenarios, account diffs, scenario forking, CPI paths, Markdown export, and typed MCP tools for repeatable review workflows.

Scenario runscapture what changed across accounts and instructions

CPI contexttrace cross-program behavior and review impact

Audit exportsship concise Markdown reports from graph-backed sessions

Product proof

Graph views, terminal evidence, MCP tools, and reports from the current work.

source spans

Solidity graph canvas

Explore code relationships, paths, and slices as reviewable graph state.

trace proof

Trace output

Evidence trails that explain where claims come from.

report context

Review workbench

Sessions, context, reports, and MCP agent tools in one workspace.

Web2 precedent

Proven in Web2 security. Missing in smart contracts.

Knowledge graphs, code property graphs, and attack-path graphs are becoming the context layer for AI-assisted security. CodeQL made code queryable. Joern made vulnerability research graph native. BloodHound made attack paths operational. ilold brings that evidence-first model to EVM and Solana security through MCP.

CodeQL code as queryable security data Joern / CPG code property graphs for vulnerability discovery BloodHound attack-path graphs as practical security infrastructure Sourcegraph large-codebase context for developers and AI-assisted workflows CodeGraph MCP-native code context for modern AI coding agents LLMxCPG LLMs over code property graph slices for vulnerability detection CKG-LLM contract knowledge graphs plus LLM-assisted vulnerability review ilold MCP-native evidence graphs for EVM and Solana audit workflows

How it fits

Static analyzers find patterns. ilold preserves security evidence.

The goal is not to replace Slither, fuzzers, manual auditors, or AI agents. The goal is to connect their outputs into a graph that can be reviewed, queried, and reused across the audit lifecycle.

Static analyzers Find known code patterns and detector matches. ilold links detector output to reachable paths, source spans, reports, and fix proof.

Fuzzers and traces Explore execution behavior and runtime state changes. ilold stores those traces as graph evidence connected to authority and value flow.

Manual audits Find semantic and business-logic failures. ilold keeps the reasoning path reviewable instead of trapped in notes and screenshots.

AI security agents Reason over security context and propose hypotheses. ilold builds the smart contract evidence graph those agents need to reason reliably.

$ mcp.query("which paths can transfer value?")

$ mcp.query("what role or signer protects this call?")

$ mcp.query("what changed in this pull request?")

$ mcp.query("did fix #42 close the value path?")

returns graph evidence, not guesses

MCP context layer

Not an AI auditor. The evidence layer agents need.

AI security agents are only as good as their context. ilold gives agents deterministic graph facts, source spans, traces, reports, and prior findings so Claude Code, Codex, Cursor, and audit bots start from evidence.

Security posture

MCP, but security-first.

Security context is sensitive. ilold should expose evidence to agents through narrow, reviewable interfaces rather than handing an LLM arbitrary shell access or unscoped repository state.

Read-only firstGraph resources are safe to inspect before write-capable tools are added.

Explicit schemasEach MCP tool has named inputs, bounded outputs, and predictable evidence types.

Local indexingPrivate code can be indexed locally with secrets redaction and controlled exports.

Audit trailAgent queries can be logged with the exact evidence returned for review.

ilold

Pilot the evidence graph behind AI-assisted audits.

Use ilold as the graph-backed context layer for EVM and Solana smart contract review, audit preparation, AI agents, and fix proof.

Open GitHub Discuss MCP integration