Understand-Anything MCP Server

Understand-Anything MCP Server

This MCP Server bridges Egonex-AI/Understand-Anything with Claude Desktop and other MCP clients, giving LLMs native access to full-codebase structural graphs, architecture analysis, and CI/CD validation to prevent breaking changes.

Privacy Policy

[!IMPORTANT]Link to Privacy Policy

Data Processing Details:

• License keys and email addresses are securely stored for billing purposes.
• Purely local, no network calls: ua_status, ua_scan, ua_graph_summary, ua_explain, ua_onboarding_doc.
• Sends graph data to the backend (on both Free and Pro): ua_precheck, ua_find_callers, ua_impact_analysis, ua_rules, ua_ci_check, ua_validate_graph. When these tools are used, the full local graph object is sent per request to our backend for processing, license, and quota validation.
• No source code contents are transmitted, only graph metadata (file paths and import relationships). All backend graph processing is done purely in-memory per-request and is never persisted.

Features

Most code intelligence tools tell your agent what's connected. ua-mcp tells your agent what it's allowed to change — and flags risky changes before they happen.

• Advanced Context Analysis: Export targeted contexts or the full graph for advanced LLM reasoning.
• Open-source with Premium Tiers: Free essential features, with advanced capabilities for Pro users.

[!TIP]System Instruction Recommended: For the best experience, add the following to your AI assistant's system prompt or custom instructions:"Always call ua_precheck before modifying any file in this project."

How it works

1. Your AI agent decides it wants to modify a critical file (e.g., src/auth.ts).
2. The agent (following its system instructions) triggers ua_precheck before making the edit to run an Architectural Safety Linter.
3. The server analyzes the graph to determine the "blast radius" and checks it against your rules.
4. If the blast radius is too large or violates a rule, a Safety Checkpoint (Elicitation Prompt) interrupts the agent, asking for your explicit confirmation before proceeding.

Configuration Rules (.ua-rules.json)

You can define specific boundaries in a .ua-rules.json file in the root of your workspace to dictate what the LLM is allowed to touch.

{
  // Understand-Anything Architectural Rules
  // Define constraints that agents and developers must respect.
  "rules": [
    {
      "id": "no-ui-db-import",
      "description": "UI layer must never import database layer directly",
      "from_pattern": "src/ui/**",
      "to_pattern": "src/db/**",
      "severity": "error"
    },
    {
      "id": "auth-required-for-payments",
      "description": "Payment modules must always be reachable from auth",
      "requires_path_through": "src/auth/**",
      "for_pattern": "src/payments/**",
      "severity": "error"
    }
  ]
}

Quick Start

Install the server globally:

npm install -g ua-mcp

Client Configuration

Add the following to your MCP client configuration file:

Claude Desktop (claude_desktop_config.json)

Add the following to your Claude Desktop config file (usually ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "understand-anything": {
      "command": "npx",
      "args": ["-y", "ua-mcp"],
      "env": {
        "UA_PROJECT_PATH": "/path/to/your/project",
        "UA_LICENSE_KEY": "your_license_key_here"
      }
    }
  }
}

[!WARNING]Claude Desktop Bug (July 2026): Claude Desktop currently has a bug where MCP elicitation requests (form prompts) silently auto-cancel instead of showing the dialog UI. If your agent says the precheck was cancelled without you seeing a prompt, this is a Claude Desktop client bug, not a server failure.

Cursor

1. Go to Settings > Features > MCP.
2. Click + Add new MCP server.
3. Name: understand-anything
4. Type: command
5. Command: npx -y ua-mcp

Continue

Add to your config.json under mcpServers:

"understand-anything": {
  "command": "npx",
  "args": ["-y", "ua-mcp"]
}

Available Tools & Tiering

The Understand-Anything MCP Server operates on a tiered licensing model.

Core Tools (Free Tier)

Available out of the box with no license required.

• ua_precheck: Pre-flight architectural risk check (10 checks/day, default critical-path rules only)
• ua_status: Returns MCP health status.
• ua_scan: Forces a re-scan of the workspace.
• ua_graph_summary: Returns aggregated node/edge statistics.
• ua_architecture_report: Groups files by top-level modules.
• ua_dependency_report: Identifies files with the most incoming dependencies (fan-in).
• ua_explain: Retrieves 1-hop dependencies for a specific file.
• ua_onboarding_doc: Generates onboarding context.

Premium Tools (Pro Tier)

• ua_precheck: Unlimited pre-flight checks with configurable critical paths and .ua-rules.json enforcement
• ua_rules: Enforces custom .ua-rules.json boundaries.
• ua_rules_check: Validates the rules schema.
• ua_find_callers: Retrieves reverse dependencies up to 2 hops.
• ua_impact_analysis: Retrieves full transitive closure of reverse dependencies.
• ua_validate_graph: Checks the knowledge graph schema for corruption.
• ua_ci_check: Analyzes Git PR diffs for architectural impact.

[!IMPORTANT]The True Enforcement Backstop: While local agents rely on system instructions to run ua_precheck, the ua_ci_check tool is designed to be your unbypassable safety net.By running ua_ci_check in your GitHub Actions and requiring it as a status check in GitHub Branch Protection, you ensure that no rogue edits can ever be merged into production without explicit approval.

Pricing

Get your license key:

• Pro - Monthly ($10)
• Pro - Lifetime ($50)

Troubleshooting

• Server fails to start: Ensure you have Node.js v18 or later installed.
• License key error: Verify your key in the .env file or Claude config matches the one on your dashboard.
• Path not found: Ensure UA_PROJECT_PATH is absolute or resolves correctly relative to where the server runs.

License

MIT License