> ## Documentation Index > Fetch the complete documentation index at: https://docs.metabind.ai/llms.txt > Use this file to discover all available pages before exploring further. # Draft and production > How Metabind separates editable drafts from stable production endpoints Every Metabind project has two MCP server endpoints: a **draft** that updates instantly as you edit and a **production** endpoint that only changes when you publish. This separation is the same shape web teams use for staging and prod, mapped to MCP. ## The two endpoints Metabind Studio edits flow into a draft endpoint that always reflects the latest save; promotion to the production endpoint only happens on publish. Both endpoints are project-scoped — connecting requires a project ID and an auth token (production needs no token when the project's **Access** setting is **Public**), and tools are isolated per project. ## When to use draft * Metabind Studio's test panel runs against draft. * Live device previews (iOS App Clip, Android Metabind app) connect to draft. * Connecting Claude Desktop or another MCP host to the draft endpoint lets you test the full chain — AI picks the tool, draft runs it, host renders the result — before you publish. The draft endpoint reflects every save within seconds. There is no build step and no staging deploy; the draft endpoint caches for 15 seconds. Save in Metabind Studio, refresh the host, the change is live. ## When to use production * Connect end users' Claude Desktop, ChatGPT, or other MCP hosts to your project's production URL. * Embed your tools in your own iOS or Android app via the Assistant SDK. * Anything user-facing. Production only updates when someone with publish permission publishes in Metabind Studio (**Publish Components**, then **Publish Server**). Edits in Metabind Studio do not affect production until that step. ## Connecting to the draft endpoint The draft endpoint lives at the same host as production, with `/draft` appended to the project path: ``` https://mcp.metabind.ai//projects//draft ``` For example, in Claude Desktop, add the draft URL as a custom connector. There is no separate draft token. A project API key works on both endpoints, and an OAuth token needs the `execute:mcp-draft` scope for draft. Anyone with a project API key can connect a host to your draft — share carefully. ## Connecting to the production endpoint ``` https://mcp.metabind.ai//projects/ ``` Same host, no `/draft` suffix. See [Connect to Claude Desktop](/guides/connecting-to-mcp-hosts/claude-desktop) for the full host setup. ## Why separate endpoints The separation answers a problem that isn't obvious until you've shipped tools to real users: a small bad change can break a connected AI experience as fundamentally as a bad deploy breaks a website. * A renamed property breaks every host that has the old schema cached. * A new component reference that fails the allowlist returns errors mid-conversation. * A typo in a Data Tool's `handler` throws on every call. Draft is where you make those changes without consequence. Production is the audited, versioned snapshot users actually see. ## What lives at each layer | Surface | Reads from | | - | - | | Metabind Studio test panel | Draft | | Live device previews | Draft | | MCP host pointed at draft URL | Draft | | MCP host pointed at production URL | Production (latest published package) | | Assistant SDK embedded in your app | Production | | End users via any of the above | Production | ## Switching hosts between draft and production A connected host doesn't auto-switch — it points at whichever URL you configured. Most teams maintain two host configurations (one named `-draft`, one named `-prod`) and switch by selecting the connection. Some hosts let you list both side by side. In that case, label them clearly so end users don't accidentally hit draft. ## Related The promote-to-production flow. How published packages are pinned and rolled forward. Configure draft and production connections. Preview against the draft endpoint on real hardware. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.