MCP Server

@kakunin/mcp — Model Context Protocol server. Lets AI agents query their own scope, check risk score, and append to the audit log in real time.

Overview

@kakunin/mcp exposes three tools via the Model Context Protocol:

Tool	Purpose
`verify_agent_scope`	Check if the agent is authorised to perform an action before executing it
`check_risk_score`	Get the rolling 30-day risk profile and self-throttle guidance
`audit_log_append`	Voluntarily log a behavioral event to the immutable audit trail

These tools run in the agent's own context — the agent can query its identity and compliance status without human intervention.

Quickstart

KAKUNIN_API_KEY=kkn_live_... KAKUNIN_AGENT_ID=agt_... npx @kakunin/mcp

Or install globally:

npm install -g @kakunin/mcp
KAKUNIN_API_KEY=kkn_live_... KAKUNIN_AGENT_ID=agt_... kakunin-mcp

MCP client configuration

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "kakunin": {
      "command": "npx",
      "args": ["-y", "@kakunin/mcp"],
      "env": {
        "KAKUNIN_API_KEY": "kkn_live_...",
        "KAKUNIN_AGENT_ID": "agt_..."
      }
    }
  }
}

Cursor / other MCP clients

Same pattern — provide command, args, and env. The server communicates over stdio.

Custom agent (via MCP SDK)

import { Client } from '@modelcontextprotocol/sdk/client/index.js';
import { StdioClientTransport } from '@modelcontextprotocol/sdk/client/stdio.js';

const transport = new StdioClientTransport({
  command: 'npx',
  args: ['-y', '@kakunin/mcp'],
  env: {
    KAKUNIN_API_KEY: process.env.KKN_KEY!,
    KAKUNIN_AGENT_ID: process.env.KKN_AGENT_ID!,
  },
});

const client = new Client({ name: 'my-agent', version: '1.0.0' });
await client.connect(transport);

// Now call tools...
const result = await client.callTool({
  name: 'verify_agent_scope',
  arguments: { action: 'transaction_initiated', amount_usd: 25000, venue: 'euronext' },
});

Tools

`verify_agent_scope`

Check whether the agent is authorised to perform a specific action. Verifies the active X.509 certificate, permitted_actions scope, financial limits, and revocation status.

Call this BEFORE executing — not after.

// "Can I initiate a 25K EUR trade on Euronext?"
const result = await client.callTool({
  name: 'verify_agent_scope',
  arguments: {
    action: 'transaction_initiated',
    amount_usd: 25_000,
    venue: 'euronext',
  },
});

Input:

Parameter	Type	Required	Description
`action`	`string`	✓	Scope string (`write:invoices`), API path (`/api/v1/trade`), or plain description
`venue`	`string`	—	Trading venue or system — checked against `permitted_venues` in the certificate
`amount_usd`	`number`	—	Transaction amount — checked against `max_single_trade_usd`

Output:

{
  "allowed": true,
  "reason": "Action \"transaction_initiated\" is within the scope of this certificate.",
  "certificate_status": "active",
  "permitted_actions": ["read:invoices", "transaction_initiated"],
  "checked_at": "2026-05-20T10:00:00.000Z"
}

{
  "allowed": false,
  "reason": "Transaction amount $75,000 USD exceeds the per-trade limit of $50,000 USD encoded in this certificate.",
  "certificate_status": "active",
  "permitted_actions": ["transaction_initiated"],
  "checked_at": "2026-05-20T10:00:00.000Z"
}

`check_risk_score`

Get the agent's rolling 30-day risk profile. Returns an actionable recommendation the agent can act on without parsing the raw score.

const result = await client.callTool({
  name: 'check_risk_score',
  arguments: {},
});

No input required. The agent ID is set at server startup via KAKUNIN_AGENT_ID.

Output:

{
  "score": 0.12,
  "band": "low",
  "trend": "stable",
  "recommendation": "Risk score 0.12 is LOW. Normal operations permitted.",
  "recent_high_risk_events": [],
  "checked_at": "2026-05-20T10:00:00.000Z"
}

Recommendation patterns:

band	trend	Recommendation
`high`	any	Pause operations. Notify human operator. Revocation check queued.
`medium`	`increasing`	Reduce transaction frequency. Wait for trend to stabilise.
`medium`	other	Caution — avoid high-value or irreversible ops without human confirmation.
`low`	`increasing`	Monitor — consider logging additional context.
`low`	other	Normal operations permitted.

Drift trend is insufficient_data for the first 30 days after an agent is created — not enough baseline to compute drift.

Voluntarily log a behavioral event to the immutable audit trail. Returns the risk score for the event synchronously (p99 200ms). High-risk events (score >= 0.85) automatically trigger a certificate revocation check.

const result = await client.callTool({
  name: 'audit_log_append',
  arguments: {
    event_type: 'transaction_initiated',
    metadata: { amount: 25000, currency: 'EUR', venue: 'euronext', instrument: 'EUR/USD' },
    session_id: 'sess_abc123',
  },
});

Input:

Parameter	Type	Required	Description
`event_type`	`ActionType`	✓	Type of behavioral event (see table below)
`metadata`	`object`	✓	Key-value context for the event. Avoid PII — stored in immutable log.
`session_id`	`string`	—	Optional — group related events in the audit trail

Action types:

Type	Description
`api_call`	Standard API invocation
`authentication_attempt`	Login / token request
`authentication_failure`	Failed auth
`data_access`	Read operation
`data_mutation`	Write / delete operation
`transaction_initiated`	Financial transaction
`transaction_anomaly`	Transaction outside normal pattern
`unauthorized_access_attempt`	Scope violation attempt
`message_signed`	Agent signed a message with KMS
`message_verification_failed`	Signature verification failed

Output:

{
  "inserted": true,
  "tx_id": "evt_8f3c2a91d4",
  "risk_score": 0.12,
  "risk_band": "low",
  "revocation_check_queued": false,
  "logged_at": "2026-05-20T10:00:00.000Z"
}

Recommended usage pattern

A well-behaved agent using Kakunin MCP looks like this:

// 1. Before any significant action — check scope
const scope = await mcp.callTool('verify_agent_scope', {
  action: 'transaction_initiated',
  amount_usd: 25_000,
  venue: 'euronext',
});

if (!scope.allowed) {
  throw new Error(`Kakunin blocked action: ${scope.reason}`);
}

// 2. Check risk — consider self-throttling if medium/high
const risk = await mcp.callTool('check_risk_score', {});
if (risk.band === 'high') {
  await notifyHumanOperator(risk.recommendation);
  return; // refuse to proceed
}

// 3. Execute the action
const trade = await executeTrade({ amount: 25_000, venue: 'euronext' });

// 4. Log the event
await mcp.callTool('audit_log_append', {
  event_type: 'transaction_initiated',
  metadata: { amount: 25_000, currency: 'EUR', venue: 'euronext', trade_id: trade.id },
});

Environment variables

# Required
KAKUNIN_API_KEY=kkn_live_...   # or kkn_test_... for sandbox
KAKUNIN_AGENT_ID=agt_...       # agent's ID from kkn.agents.create()

# Optional
KAKUNIN_BASE_URL=https://api.kakunin.ai/v1  # override for testing

Sandbox mode

Use a kkn_test_... API key. The server will connect to the sandbox CA — same tools, no regulatory validity.

KAKUNIN_API_KEY=kkn_test_... KAKUNIN_AGENT_ID=agt_... npx @kakunin/mcp

Overview

Installation

Quickstart

MCP client configuration

Claude Desktop

Cursor / other MCP clients

Custom agent (via MCP SDK)

Tools

`verify_agent_scope`

`check_risk_score`

`audit_log_append`

Recommended usage pattern

Environment variables

Sandbox mode

On this page