Skip to Content
Agents & CLIOverview

Agents & CLI Overview

ShipMore is agent-native: every operation a human can perform via the admin UI is also exposed via the shipmore CLI and the MCP plugin, on top of the same service layer. Agents drive the box; the admin is the safety net.

Four interface layers

LayerWhen to use it
CLI (shipmore)Default for humans and agents driving the box. Every operation is a command. JSON output on non-TTY. Pipe it into cron, scripts, Claude Code, CI.
MCP (plugin-mcp)Conversational agent workflows. Same operations as the CLI, structured for chat. SKILL.md tells agents how to use it.
Admin (Payload)Human review, draft approval, overrides. Safety layer for the operate loop.
REST (Payload built-in)Canonical wire protocol the other three speak. Public for external integrations.

Custom Code (Next.js / API routes) stays for true edge cases; anything that repeats should graduate into the service layer (and therefore into every surface).

The agent boundary

ShipMore has no LLM. No model SDK, no API key, no inference calls inside the platform. The intelligence layer is the operator’s harness — Claude Code, Claude Desktop, any MCP-capable agent.

The split:

  • ShipMore ships deterministic primitives. Commands, endpoints, heuristic inference, storage. Produces structured facts.
  • The harness agent brings all judgment. Reads raw data, interprets heuristic output, decides what to change.

Concretely:

  • Ship rich heuristic output if it helps the agent frame its work — labels, types, roles — not per-field analytics the agent could derive itself.
  • SKILL.md teaches agents how to think, not what to decide. No decision rules (“if sentinelRatio > 0.5, drop the field”). Narrate reasoning patterns; trust the agent with the call.
  • Platform helpers for the agent should be pull-model (agent asks: shipmore data inspect --column X) rather than push-model (platform dumps pre-aggregated stats up front).

SKILL.md

SKILL.md lives at the repo root and ships with the CLI. Any MCP-capable agent (Claude, GPT, Gemini) can read it and operate the deployment. It defines:

  • The glossary (tenant, import batch, records, product, block, page).
  • The agent protocol — JSON output, no prompts, exit codes, error envelope.
  • The 4-phase workflow for end-to-end builds (Understand → Propose & ASK → Apply → Build).
  • The command overview and pointers to references for flags, errors, recipes.

When an agent picks up a ShipMore deployment, this is the only file it needs to start.

Draft / publish workflow

Agents create content as drafts by default:

await payload.create({
  collection: 'pages',
  data: { title: 'Doctor Landing Page', layout: [...] },
  draft: true,
})

Humans review and approve via the Payload admin. Drafts never go live without explicit publish. Every agent action is auditable through Payload’s version history.

Where to next

  • CLI — install, auth, command overview, JSON contract.
  • MCP Plugin — current config, what’s enabled, how to extend.
  • Patterns — adapter rule, tenant scoping, hook safety, two operating modes.
AI GenerationCLI