Silentfail

---

The problem

You connect MCP servers to Claude, Cursor, VS Code. They break — no logs, no warnings, no clue which one is slow, dead, or eating your entire context window.

Every connected server has a token cost. Claude reads all tool schemas before your conversation even starts. Five servers with 20+ tools can burn 5,000+ tokens just sitting there. And if one of those servers has a malicious tool description with a prompt injection? You'd never know.

What SilentFail does

Discover	Finds all MCP configs across Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, and project-level files
Connect	Spawns each server, measures startup time, lists all tools
Test	Calls each tool with smart inferred parameters to verify it actually works
Audit	Calculates exact token cost per server and per tool
Detect	Finds tool name conflicts across servers
Scan	50+ security patterns — prompt injection, data exfiltration, dangerous commands, encoded payloads
Report	Clear, actionable output in CLI, chat, or visual dashboard

Quick start

git clone https://github.com/decksaga/silentfail.git
cd silentfail
npm install
npm run build

SilentFail works in two modes: MCP Server (Claude scans itself) or CLI (you run it from the terminal).

---

Mode 1: MCP Server (recommended)

Add SilentFail to your Claude Desktop config and let Claude diagnose its own setup.

claude_desktop_config.json

{
  "mcpServers": {
    "silentfail": {
      "command": "node",
      "args": ["/absolute/path/to/silentfail/dist/server.js"]
    }
  }
}

Where is this file?

OS	Path
Windows	%APPDATA%\Claude\claude_desktop_config.json
macOS	~/Library/Application Support/Claude/claude_desktop_config.json
Linux	~/.config/claude/claude_desktop_config.json

Then restart Claude Desktop and ask:

"Scan my MCP servers"

Claude calls the scan tool, gets the full diagnostic report, and presents it in whatever language you used. On first use, it asks if you want results in chat or the visual web dashboard. It remembers your preference.

Tools exposed

Tool	Description
scan_mcp_servers	Full diagnostic: health, tokens, conflicts, security, recommendations. Set test_tools: true to verify each tool works.
open_silentfail_dashboard	Run scan + open interactive web dashboard on localhost:3777

---

Mode 2: CLI

Run it directly from the terminal for quick diagnostics or CI integration.

silentfail              # Quick health check
silentfail --test       # Full scan + call each tool to verify
silentfail --dashboard  # Open visual web dashboard
silentfail --json       # JSON output for scripts/CI
silentfail --help       # All options

To use silentfail as a global command:

npm link

Example output

  SilentFail — Scan Report
  ──────────────────────────────────────────────────

  OVERVIEW
  ──────────────────────────────────────────────────
  Configs found:    3 (Claude Desktop, Claude Code, Cursor)
  Servers:          4 healthy, 1 failed
  Total tools:      23
  Schema tokens:    ~4,812 (consumed before you type anything)
  Conflicts:        1
  Scan time:        6204ms

  market-pulse                                    OK
     Response: 412ms
     Tokens:   ~651 (8 tools, ~81 per tool)
     Tools:
       get_price (110 tok) passed
       get_stock_price (94 tok) passed
       get_market_summary (69 tok) passed

  broken-server                                   ERROR
     Error: Script not found: /old/path/server.js

  TOKEN BUDGET
  ──────────────────────────────────────────────────
    browser-tools             ████████████░░░░░░░░  2,340 tok (49%)
    file-system               ██████░░░░░░░░░░░░░░  1,180 tok (25%)
    market-pulse              ███░░░░░░░░░░░░░░░░░    651 tok (14%)

  SECURITY SCAN
  ──────────────────────────────────────────────────
  Risk level: CLEAN
  Scanned: 4 server(s), 23 tool(s)
    No security issues detected.

  RECOMMENDATIONS
  ──────────────────────────────────────────────────
    [broken-server] Server is broken: Script not found.
       Fix the configuration or remove this server.
    [browser-tools] Heavy schema cost: ~2,340 tokens for 12 tools.
       Consider if you use all 12. Each unused tool wastes context.
    [market-pulse] Healthy and efficient. 8 tools, ~651 tokens.

---

Smart tool testing

SilentFail doesn't just check if servers respond — it calls each tool.

It reads the schema, infers valid test inputs (AAPL for stocks, USD/EUR for forex, bitcoin for crypto), calls the tool, and categorizes the result:

Result	Meaning
Passed	Tool works, returned data
Broken	Runtime error — tool is dead
Rejected	Tool works but rejected the test input (validation is working correctly)
Skipped	Couldn't infer safe params — manual test recommended
Timeout	Took too long to respond

If it says broken, it's broken. No false positives.

Security scanning

Every scan automatically checks all server configs and tool schemas against 50+ patterns:

Category	Examples
Prompt injection	ignore previous instructions, fake system prompts, role hijacking, prompt leaking
Data exfiltration	Suspicious URLs, POST requests, webhook endpoints, base64-encode-then-send
Dangerous commands	eval(), rm -rf, reverse shells, child_process spawns
Sensitive env vars	AWS keys, API tokens, database credentials in server configs
Encoded payloads	Base64, hex, unicode, HTML entities hiding malicious content

  SECURITY SCAN
  ──────────────────────────────────────────────────
  Risk level: CRITICAL
  Scanned: 3 server(s), 15 tool(s)

    [CRITICAL] shady-server -> execute_query
      Tells the model to ignore previous instructions
      Evidence: "ignore all previous instructions and..."
      Remove this server immediately.

Token budget explained

Every MCP tool you connect has a JSON schema that Claude reads before your conversation starts. This is invisible — there's no UI for it, no warning.

SilentFail measures the exact token cost of each server and each tool, so you can see:

• Which servers cost the most context
• Which tools are worth keeping
• Where you're wasting tokens on tools you never use

A clean setup with low token overhead means more context for your actual work.

Visual dashboard

Run with --dashboard (CLI) or ask Claude to open_silentfail_dashboard (MCP mode) to get an interactive web UI:

• Server health cards with status indicators
• Token budget visualization with gradient bars
• Per-tool test results with pass/fail badges
• Security findings with severity levels
• Language toggle (English / Español)

Served on localhost:3777. Dark theme. No external dependencies.

Supported clients

SilentFail auto-discovers configs from all major MCP clients:

Client	Config path
Claude Desktop	%APPDATA%/Claude/claude_desktop_config.json
Claude Code	~/.claude/settings.json
Cursor	~/.cursor/mcp.json
VS Code	~/.vscode/mcp.json
Windsurf	~/.windsurf/mcp.json
Project-level	.mcp.json, .claude/settings.json

Architecture

  Discovery         Connect          Test Tools        Security          Report
 ┌─────────────┐   ┌──────────────┐  ┌───────────────┐ ┌──────────────┐  ┌──────────────┐
 │ Find configs │──>│ Spawn each   │─>│ Infer params  │>│ 50+ patterns │─>│ CLI output   │
 │ across all   │   │ MCP server   │  │ from schema   │ │ Injection,   │  │ MCP report   │
 │ clients      │   │ via stdio    │  │ Call & verify  │ │ exfil, etc.  │  │ Dashboard    │
 └─────────────┘   └──────────────┘  └───────────────┘ └──────────────┘  └──────────────┘

Key files:

src/
  discovery.ts   — Config auto-detection across all clients
  scanner.ts     — Server connection, tool listing, testing
  security.ts    — Pattern-based security analysis (50+ rules)
  server.ts      — MCP server mode (Claude Desktop integration)
  dashboard.ts   — Web dashboard with i18n (EN/ES)
  index.ts       — CLI entry point
  types.ts       — TypeScript interfaces

FAQ

Does SilentFail scan itself?

No. When running as an MCP server, SilentFail automatically excludes itself from the scan to avoid recursion.

Is it safe to run tool tests?

Yes. SilentFail only calls tools with read-only, minimal test inputs (stock tickers, currency codes). It never sends destructive operations. If it can't infer safe params, it skips the tool.

Can I use it in CI?

Yes. Use silentfail --test --json for machine-readable output. Parse the JSON for automated checks.

What Node version do I need?

Node 18 or higher. Uses ES modules and the MCP SDK.

License

MIT — do whatever you want with it.