← All Docs

MCP Server

Overview

Webase exposes itself as a Model Context Protocol (MCP) server, so external AI agents — ChatGPT, Claude.ai, Claude Code, Codex, and anything else that speaks MCP — can list, read, build, refine, and deploy Webase apps directly from their own conversations or shells.

What you can do over MCP: create new applications, write source files, define data models, create / update / delete records, evaluate apps, deploy to Netlify, and (for desktop coding agents) upload pre-built static bundles.

Three endpoints, picked by agent capability

Webase ships three MCP endpoints that share the same backing services but expose different tool subsets. Pick the one that matches the agent's environment, not the application's structure.

Managed endpoint

https://www.webase.com/mcp/managed/messages

For chat-class agents (ChatGPT, Claude.ai). They write source files via MCP; Webase compiles automatically in the background. Use this when the agent has no local Node toolchain.

External endpoint

https://www.webase.com/mcp/external/messages

For desktop coding agents (Claude Code, Codex). They run a real toolchain locally and upload compiled assets via upload_build. Source-file mutation tools are removed from this surface.

Legacy endpoint

https://www.webase.com/mcp/messages

Exposes every tool. Kept for back-compat with existing integrations and for power users that want the full surface in one place. New integrations should use managed or external.

Protocol

JSON-RPC 2.0 over HTTPS. Standard MCP initialize, tools/list, and tools/call. Authentication is OAuth 2 (recommended) or an x-api-token header — see Authentication & Scopes.

First call any agent should make

Every MCP session should start by calling get_runtime_constraints. It returns the conventions every Webase app must follow (entry points, import map, Tailwind/HashRouter rules, the data API URL shape) plus a build-mode comparison. Without it, even good agents will produce non-bundleable apps. 

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": { "name": "get_runtime_constraints" }
}

Where to go next