Veridict MCP Server

Veridict

Know if you can trust your MCP server.

Agents call tools. Some fail silently. Veridict gives you a signal.

Part of the emerging Agent Trust Stack (Runtime Verification Layer).

Quickstart (30 seconds)

npm install veridict

import { withVeridict } from "veridict";               // 1. import
// ... register your tools as usual ...
withVeridict(server, { name: "my-server" });            // 2. wrap
await server.connect(transport);

Done. Every tool call is now logged. Two new tools are automatically added:

• veridict_stats — execution statistics
• veridict_can_i_trust — trust judgment (YES / CAUTION / NO)

What you get

{
  "verdict": "yes",
  "score": 0.97,
  "reason": "reliable",
  "recentSuccessRate": 0.98,
  "recentExecutions": 85,
  "failureBreakdown": { "timeout": 2, "error": 1 },
  "stats": {
    "totalExecutions": 1247,
    "successRate": 0.992,
    "avgLatencyMs": 142
  }
}

Two-layer design:

• Layer 1 (instant decision): verdict + score + reason — all an agent needs
• Layer 2 (details on demand): stats + failureBreakdown + recentSuccessRate — dig deeper when needed

v0.2.0 — What's new

• Failure classification — failures are categorized as timeout, error, validation, or unknown
• Time decay — recent executions (last 7 days) are weighted 70%, all-time 30%
• Trust score — score field (0-1) replaces confidence for clearer semantics
• Failure breakdown — see exactly what types of failures are occurring
• Agent-friendly reasons — short, actionable: "elevated timeout rate", "reliable", "high error rate"

Verbose mode

See executions in real-time:

withVeridict(server, { name: "my-server", verbose: true });

Output (stderr):

[veridict] monitoring "my-server"
[veridict] search_docs ok 120ms
[veridict] create_item ok 85ms
[veridict] fetch_data FAIL [timeout] 30201ms — timeout
[veridict] search_docs ok 94ms

CLI

npx veridict                       # Show all tracked servers
npx veridict stats my-server       # Detailed stats
npx veridict trust my-server       # Trust judgment

Trust judgment logic

Blended rate = recent 7-day success rate (70% weight) + all-time rate (30% weight).

Recent performance matters more than history.

Failure types

Options

withVeridict(server, {
  name: "my-server",            // Required: server identifier
  instanceId: "prod-1",         // Optional: distinguish instances
  dbPath: "./my-logs.db",       // Optional: custom DB path (default: ~/.veridict/executions.db)
  minExecutions: 20,            // Optional: min data for judgment (default: 10)
  verbose: true,                // Optional: log to stderr (default: false)
});

Layer 1 integration (MCP Trust Kit)

Combine static analysis with runtime monitoring for full-stack trust:

# Step 1: Static scan with MCP Trust Kit (v0.5.0+)
npx mcp-trust-kit scan --json-out layer1-baseline.json --cmd node my-server.js

import { withVeridict, parseLayer1Report } from "veridict";
import report from "./layer1-baseline.json";

withVeridict(server, {
  name: "my-server",
  baseline: parseLayer1Report(report),  // Layer 1 → Layer 2
});

What this does:

• Static risks (e.g., dangerous_fs_write) are factored into trust judgment
• Critical static risks downgrade a "yes" verdict to "caution" even at 99% success rate
• scan_timestamp is preserved for Layer 3 temporal decay logic
• The full Layer 1 report is stored in baseline.raw for cross-org consumers

Part of the Agent Trust Stack:

Layer 1: MCP Trust Kit (pre-deploy)  → "Is this server safe to run?"
Layer 2: Veridict (runtime)          → "Is this server actually reliable?"
Layer 3: SATP/XAIP (cross-org)      → "Should I trust this across boundaries?"

How it works

1. Wraps all registered MCP tool handlers
2. Logs every execution: input hash, output hash, success/fail, failure type, latency
3. Stores in local SQLite (~/.veridict/executions.db)
4. Provides trust judgment based on time-weighted execution history

Future

• Cross-agent trust verification
• OpenTelemetry export
• On-chain audit trails
• Enterprise compliance reporting

Early users welcome

If you're building MCP servers or agents, I'd love your feedback. Try Veridict and tell me what breaks.

License

MIT