REDHAND.

04 — AI AGENT SECURITY & CONTROL LAYER · MCP GATEWAY

Identify every agent, control what it touches, and stop it in seconds.

RedHand sits inline between your connected AI agents and their MCP tool servers. It enforces short-lived Ed25519 Agent Passports, read/write tool scopes, Organization Safe Mode, daily Budget Guard caps, planted honeytokens, SHA-256 tool pinning, and a tamper-evident Black Box timeline—with zero LLM calls on the hot path.

AGENTWALL MCP GATEWAY● FAIL-CLOSED
REVOCATION / FREEZE SLO≤ 5.0 s
HOT-PATH LLM CALLS0 (DETERMINISTIC)

Covers agents connected via the RedHand gateway (n8n, Claude, LangGraph, CrewAI, custom). Includes 10-min Ed25519 tokens, Telegram /freeze_all, SHA-256 Black Box, Honeytokens, and MCP Tool Pinning.

01 — INTERACTIVE GATEWAY & CONTROL SIMULATOR

Inspect how RedHand enforces Phases 1–4 inline.

FAIL-CLOSED MCP GATEWAY · 0 LLM CALLS ON HOT PATH
PROBE 01 · PHASE 1 PASSPORT & AST
REFUSED BY GATEWAY & SAFE MODE (0.8 ms)
1. INBOUND MCP EVENT / CONTROL TRIGGER
// JSON-RPC tools/call via AgentWall MCP Gateway (:8080)
{
  "method": "tools/call",
  "params": {
    "name": "postgres_prod.execute_sql",
    "arguments": {
      "query": "DROP TABLE retail_ledger_transactions CASCADE;"
    }
  }
}
2. REDHAND GATEWAY & CORE EVALUATION
gate: agentwall/src/gateway/service.py + ast_guardrails.py
passport_token: Ed25519 JWT (ttl <= 600s, status: "active")
org_safe_mode: true (read-only enforced across organization)
sqlglot_ast: DDL_DESTRUCTIVE ("DROP") + scope "write" not in passport
3. RUNTIME ENFORCEMENT & BLACK BOX OUTCOME

Refused fail-closed before reaching the upstream MCP server. Recorded in the organization's SHA-256 Black Box timeline with reason `safe_mode` / `out_of_scope`.

02 — BUILT & VERIFIED CAPABILITIES

Six deterministic control layers.

RedHand only controls and sees what flows through the RedHand gateway, or what an owner submits to the scanner.

RH-01PHASE 1 · IDENTITY

Ed25519 Agent Passports & Inline MCP Gateway

Registers every agent (n8n, Claude, LangGraph, CrewAI, custom) with owner, backup owner, server/tool read-write scopes, expiry, and 10-minute Ed25519 tokens enforced at the fail-closed AgentWall proxy.

Fail-Closed Gateway Enforcement✓
RH-02PHASE 2 · ≤ 5S SLO

Kill Switch, Safe Mode & Telegram /freeze_all

Freeze or permanently revoke a single agent, flip the organization into read-only Safe Mode, or send /freeze_all from Telegram on your phone to stop every connected agent within 5 seconds.

Fail-Closed Gateway Enforcement✓
RH-03PHASE 2 · METERING

Budget Guard & Daily Tool-Call Caps

Enforces per-agent and organization-default daily tool-call limits with non-blocking in-memory metering, instant Telegram breach alerts, and opt-in automatic agent freeze.

Fail-Closed Gateway Enforcement✓
RH-04PHASE 3 · AUDIT

SHA-256 Hash-Chained Black Box & DLP Capture

Per-organization append-only timeline of every allowed, refused, or failed tools/call with one-click chain Verify, JSONL/CSV export, and opt-in 2 KB payload capture that redacts secrets, emails, cards, and Malaysian NRICs.

Fail-Closed Gateway Enforcement✓
RH-05PHASE 3 · DECEPTION

Planted Honeytokens & Decoy MCP Tools

Issues bait credentials (sk_live_…, AKIA…, passwords stored only as SHA-256 digests) and per-agent decoy tools in tools/list. Any touch is refused before upstream, alerts Telegram in < 10s, and freezes the agent.

Fail-Closed Gateway Enforcement✓
RH-06PHASE 4 · INPUT CHECKS

Bilingual MCP Scanner & SHA-256 Tool Pinning

Scans pasted manifests or live registered MCP servers (English & Bahasa Malaysia rules, measured by redhand bench) and pins SHA-256 tool hashes on tools/list to stop post-approval rug pulls.

Fail-Closed Gateway Enforcement✓