Connect via MCP
The Remote MCP Server lets an AI agent work with your Boost.space system in real time — read and search records, create and update data, build modules and fields, and run Boost.space's AI operations. It runs on Boost.space infrastructure, so there is nothing to host yourself.
An agent connected over MCP acts with exactly your Boost.space permissions — it can read, write, and query only the data you can, nothing more.
MCP is the surface for when an agent decides what to do. You describe an outcome and it works out the steps. If instead you already know the steps and want to write them down, you want the SDK — or the REST API if your language has no SDK yet. Same data, same permissions; the difference is who is authoring the sequence. Bulk writes and deletes exist only on those two.
Two ways to work with an agent
- In the app — open the Agent Chat (the Agent button in the sidebar): the built-in agent chat, scoped to one module or all of them. No setup at all.
- Your own client — connect Claude, ChatGPT, Cursor, v0, Lovable, or any MCP client to your data. That's what this guide covers.
The fastest way: the Connect AI window
Boost.space has a built-in Connect AI window — a guided control panel that connects an external agent to your data. It creates the MCP token for you, gives you the connection URL, and shows the exact steps for your chosen client.
Open it from the Connect AI button on a custom module, or from your user menu → API / MCP tokens → Create new. Then:
- Choose MCP token (the default) over an API token.
- Pick your AI model — or choose Other and select OAuth or API key.
- Copy the MCP URL and follow the on-screen steps for that client.
- Optionally try the example prompts, then finish — your token is created.
Prefer to wire it up by hand? Use the steps below.
What you need
Two things, both from your Boost.space instance:
-
Your connection URL — copy it from the Connect AI / MCP token screen (don't build it by
hand) — it looks like
https://mcp.boost.space/v2/{system}/sse, where{system}is your system name. - Either an MCP token or an OAuth sign-in, depending on how your client connects.
See API tokens and MCP tokens for how to create the token and find your URL.
Treat an MCP token like a password — anyone who has it can act on your data through the assistant, up to the token's permissions.
Two ways to connect
Modern clients support both; pick whichever the client offers.
| Method | How it works | Best for |
|---|---|---|
| URL + token | Paste the connection URL and your MCP token into the client's connector settings | Claude, ChatGPT, Cursor, VS Code — any client with a "custom connector / MCP server" field |
| OAuth | Paste the URL only; the client sends you to a Boost.space sign-in and receives its own token | Clients that support OAuth for remote MCP — no token to copy or store |
Both connect to the same server. OAuth avoids handling a token by hand; URL + token works everywhere. Client-specific steps are in Common MCP clients.
What you can do once connected
The agent works with your whole data model, not just reads. Within its token's permissions it can:
- Records — search, read, create, and update records across your spaces, one record at a time.
- Structure — create and change modules, fields, field groups, spaces, statuses, and labels.
- Files, users, teams — upload and manage files, and read or manage users and teams.
- AI operations — generate and enrich field values, transform or validate whole columns, and generate module definitions.
- Semantic search — find the records most relevant to a natural-language query.
The tool list is not fixed. Your server advertises the tools for the modules your system has active, so two systems can legitimately offer different sets — an agent is better off without warehouse tools on a system that has no warehouses. The live server is the only accurate list, so the way to see yours is to ask: "what Boost.space tools do you have?"
Where MCP stops. Deleting records and writing in bulk are not MCP operations — use the REST API or an SDK for those, and for anything where you want to push thousands of rows in one call rather than have an agent walk them one by one.
