Quickstart -- 60 Seconds to First Interception

From zero to cryptographically governed AI agent in four commands.

Prerequisites

Node.js 18+
An MCP-compatible AI client (Claude Desktop, Cursor, or any MCP host)

Step 1: Initialize the Project

npx @oconnector/mcp-gateway init

This creates (source: abs-core/packages/mcp-gateway/src/cli.ts:cmdInit()):

.abs/policy.json -- Your governance rules (deny lists, escalation policies, argument restrictions)
.abs/.env.example -- Environment variable template for key management
.abs/example-integration.ts -- Example integration code
Updates .gitignore to exclude .abs/ secrets

There is no abs-governance.config.ts file — that was a documentation error.

Step 2: Configure Your MCP Client

Add the ABS Core gateway to your MCP client configuration:

{
  "mcpServers": {
    "abs-governance": {
      "command": "npx",
      "args": ["@oconnector/mcp-gateway", "watch"],
      "env": {
        "ABS_POLICY_PATH": ".abs/policy.json"
      }
    }
  }
}

Step 3: Define Your First Policy

Edit .abs/policy.json:

{
  "$schema": "https://abscore.app/schemas/policy-v4.json",
  "version": "4.0.0",
  "deny_tools": ["system.exec", "system.shell", "fs.delete*", "fs.rmdir*", "db.drop*"],
  "escalate_tools": ["payment.*", "email.send", "api.deploy*"],
  "arg_restrictions": {
    "db.query": { "sql": "deny:DROP|DELETE|TRUNCATE|ALTER" },
    "fs.write": { "path": "deny:/etc|/sys|/proc|../" }
  },
  "fail_mode": "closed",
  "require_oid": true
}

arg_restrictions values are deny:-prefixed regex strings, not arrays like blocked_patterns — that format was a documentation error. See abs-core/packages/mcp-gateway/src/cli.ts (the real default policy generated by init) for the exact shape.

This policy:

Blocks system.exec, system.shell, and destructive filesystem/db tools unconditionally
Escalates payment, email, and deploy tools for review
Restricts SQL and filesystem-path arguments to prevent destructive operations

Step 4: Start the Gateway

npx @oconnector/mcp-gateway watch

Output:

[ABS CORE v4.3.3] Gateway active
[ABS CORE] Policy loaded: .abs/policy.json
[ABS CORE] Enforcement mode: FULL
[ABS CORE] Waiting for MCP tool calls...

What Happens Next

When the AI agent attempts a tool call:

OID -- Agent identity is verified via Ed25519 signature
HASH -- Request payload is hashed (SHA-256) for deterministic fingerprinting
ENGINE -- Policy engine evaluates the request: ALLOW, DENY, or ESCALATE
LEDGER -- Decision is recorded in the hash chain with cryptographic receipt
EXECUTE -- If allowed, the tool call proceeds. If denied, it is blocked with a reason.

Every decision produces a SovereignAuditRecord -- a frozen, immutable proof of governance.

Verify It Works

There is no status subcommand — the CLI only supports init, scan, watch, keys, and --help (abs-core/packages/mcp-gateway/src/cli.ts). To see interception activity, watch the watch command's own log output, or inspect the ledger directly (see Events API). The example output below is illustrative of the concept, not a real command's output.

Illustrative summary of what a governed session looks like once agent traffic flows through watch:

Governance Status:
  Total intercepted: 1
  Allowed: 1
  Denied: 0
  Escalated: 0
  Chain integrity: VALID
  License status: FULL

Next Steps

KeyProvider API -- Configure persistent Ed25519 key management
MCP Security Gateway -- Advanced proxy configuration
Air-Gapped Operations -- Deploy without internet access
NIST AI RMF Mapping -- Understand compliance coverage

Quickstart -- 60 Seconds to First Interception

On this page