Skip to main content
The kill switch lets you immediately block an agent from making any LLM requests through the MeshAI proxy. Blocks are cached in Redis for instant enforcement — the proxy returns 403 Forbidden for any request from a blocked agent.

​
How It Works

  1. You call POST /agents/{id}/block with a reason.
  2. The API writes the block to the database and publishes it to Redis.
  3. The proxy checks Redis on every incoming request — blocked agents get an immediate 403.
  4. An agent.blocked audit event is emitted automatically.
  5. To restore access, call POST /agents/{id}/unblock — an agent.unblocked event is emitted.

​
Block an Agent

curl -X POST https://api.meshai.dev/agents/agt_abc123/block \
  -H "Authorization: Bearer msh_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "reason": "Anomalous behavior detected — cost spike 10x above baseline"
  }'

{
  "success": true,
  "data": {
    "agent_id": "agt_abc123",
    "status": "blocked",
    "reason": "Anomalous behavior detected — cost spike 10x above baseline",
    "blocked_at": "2026-03-18T14:30:00Z",
    "blocked_by": "user@company.com"
  }
}

​
Unblock an Agent

curl -X POST https://api.meshai.dev/agents/agt_abc123/unblock \
  -H "Authorization: Bearer msh_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{}'

{
  "success": true,
  "data": {
    "agent_id": "agt_abc123",
    "status": "active",
    "unblocked_at": "2026-03-18T15:00:00Z",
    "unblocked_by": "user@company.com"
  }
}

​
What the Agent Sees

When a blocked agent sends a request through the proxy, it receives: 
{
  "error": "Agent blocked: Anomalous behavior detected — cost spike 10x above baseline",
  "agent_id": "agt_abc123",
  "status": 403
}
This is compatible with all LLM SDK error handling — OpenAI, Anthropic, and other clients will raise their standard authentication/permission errors.

​
Audit Trail

Every block and unblock action emits an audit event:
Event TypeDescription
agent.blockedAgent was blocked (includes reason and actor)
agent.unblockedAgent block was removed (includes actor)
Query block-related audit events: 
curl "https://api.meshai.dev/governance/audit-trail?event_type=agent.blocked" \
  -H "Authorization: Bearer msh_YOUR_API_KEY"

​
Enforcement Architecture

The kill switch uses Redis for sub-millisecond enforcement:
  • Write path: API writes block to PostgreSQL and publishes to Redis.
  • Read path: Proxy checks Redis on every request (before policy evaluation).
  • Fail-open: If Redis is unavailable, the proxy falls back to allowing the request (same as policy cache behavior).
  • Latency: Block takes effect within milliseconds of the API call.

​
Use Cases

  • Incident response — Immediately stop a compromised or malfunctioning agent.
  • Cost containment — Block an agent that is burning through budget unexpectedly.
  • Compliance — Emergency shutdown to satisfy regulatory requirements (EU AI Act Article 14).
  • Anomaly response — Automated block triggered by anomaly detection (coming soon).