Deployment Overview

Deployment Overview

Architecture

┌─────────────┐       ┌─────────────┐       ┌─────────────┐       ┌──────────┐
│   Clients   │──────▶│   Server    │◀──────│    Agent    │──────▶│ Database │
│ (CLI / MCP) │ HTTP  │ (REST API)  │ poll  │ (executor)  │  SQL  │(PG / MY) │
└─────────────┘       └─────────────┘       └─────────────┘       └──────────┘

• Clients never touch the database. They submit requests to the server and retrieve results.
• Server never touches the database. It manages approval state, audit logs, and routes requests.
• Agent polls the server. It runs in a network that can reach the database and fetches approved work via outbound HTTP.

Deployment models

Separated

Server on a central host, agent on a host with database access, CLI on developer machines.

Developer laptop  ──▶  Server (cloud VM)  ◀──  Agent (DB subnet)  ──▶  PostgreSQL

Multi-agent

One server, multiple agents for different databases or environments.

                        ┌── Agent (staging) ──▶ Staging DB
Server ◀────────────────┤
                        └── Agent (prod)    ──▶ Production DB

Agents register with capabilities; the server matches requests to the appropriate agent.

Choose a deployment method

Method	Page	Best for
Binary / systemd	server.md + agent.md	Bare metal / VM
Docker Compose	docker.md	Small teams, single host
ECS Fargate	ecs.md	AWS native
Kubernetes	kubernetes.md	K8s clusters
Helm	helm.md	GitOps / Helm-managed clusters

Binary / systemd: Deploy the server first, obtain bootstrap tokens, then deploy the agent.

server.md and agent.md are component references. All platform-specific pages (Docker, ECS, K8s, Helm) link back to them for configuration details.

Network requirements

From	To	Port	Protocol	Purpose
Client	Server	3000	HTTP/HTTPS	Submit requests, get results
Agent	Server	3000	HTTP/HTTPS	Poll for work, report results
Agent	Database	5432/3306	PostgreSQL/MySQL	Execute queries

The server needs no inbound access from the agent’s network — the agent initiates all connections outbound.

Transport security: The agent refuses to start if its [server].url is external HTTP (non-private IP, non-localhost). Use HTTPS for the agent→server connection, or set allow_insecure = true if TLS is handled at a network layer not visible in the URL.

Security model

1. No DB credentials on clients — credentials exist only in the agent’s config
2. Signed execution tokens — Ed25519 signatures prevent request tampering
3. RBAC — admin, developer, readonly roles enforced by the server
4. Audit log — every operation recorded with hash-chain integrity
5. Fail-closed workflows — if policy evaluation fails, the request is denied (not auto-approved)

See also

• Server configuration — all server settings and operations
• Agent configuration — agent settings, capabilities, resilience
• Troubleshooting — common deployment issues