MCP setup

ERPSpan exposes your GP data to AI agents (Claude Desktop, Claude Code, any MCP client) at one endpoint: POST /mcp. It runs the same machinery as the REST API — the MCP tools call the identical internal services — so an agent reads and (with a human in the loop) writes GP exactly as the REST API does.

• Read tools — list_connections, list_<object> / get_<object> for every read object (customers, vendors, items, accounts, ar-invoices, …), and search_customers / search_vendors / search_items. Every list is capped at 50 rows; page with the returned nextCursor.
• Write tools — writes are always draft → commit (never a single call): draft_<object> validates and previews without executing, commit_draft executes it, discard_draft cancels.

Connect a client (Claude Desktop)

Add ERPSpan to your claude_desktop_config.json. It's a remote, streamable-HTTP server authenticated with an API key — nothing to install locally:

{
  "mcpServers": {
    "erpspan": {
      "type": "http",
      "url": "https://api.erpspan.com/mcp",
      "headers": { "Authorization": "Bearer lb_live_YOUR_KEY" }
    }
  }
}

• The key must have the mcp scope to reach /mcp, and additionally the write scope to use the draft/commit write tools.
• Writes are off by default per connection: a draft_/commit_ is refused with WRITE_NOT_ENABLED until you enable that object under Connections → Write access. Sandbox connections are always read-only.

A worked example

A typical hands-free session — the model picks each tool from its description:

Safety

• A draft_ call never touches GP — an unbalanced or malformed body is rejected in the draft and never reaches the tunnel.
• commit_draft runs the same eConnect pipeline as a REST write. The result lands in an unposted batch — a person still posts it in GP. Re-committing the same draftId replays the first result (idempotent); it never double-posts.
• Every write requires the write scope and the object on the connection's whitelist and the draft→commit two-step — none can be bypassed.