How to host a remote MCP server without running one
Updated
Kaiva Bridge hosts the MCP server for you. Point it at an OpenAPI spec, a database, your documents, a spreadsheet or an existing remote MCP server, and you get an address on api.kaiv.ai that speaks MCP over Streamable HTTP. There is nothing to deploy, scale or keep running.
Step by step
- Choose a sourceIn the console choose New server and pick REST API, Database, Spreadsheet, Documents or MCP server (to put an existing remote MCP server behind your address).
- Choose the toolsBridge reads the source and lists its tools. Switch on the ones AI clients should see.
- PublishYour server is live at https://api.kaiv.ai/api/bridge/mcp/<name>.
- Connect clientsAdd the address to Claude, ChatGPT, Cursor or any MCP client. Clients sign in with OAuth 2.1, or use a key made for one server.
Wrap an existing remote MCP server from a terminal:
npx @kaiva/bridge login
npx @kaiva/bridge push vendor-tools --mcp https://mcp.example.com/mcp
npx @kaiva/bridge doctor vendor-toolsA no-code hosted MCP gateway
Bridge sits between AI clients and your systems. A client connects to one address, signs in and calls tools; Bridge sends each call to your API, database or MCP server with your credential and returns the answer. The tools are generated from the source, so there is no server code to write, and several servers can share one address as a bundle.
Local stdio MCP server or managed remote MCP server
An MCP server either runs on each person's machine, started by their AI client over stdio, or runs once at an address every client reaches over Streamable HTTP. Frameworks such as FastMCP can serve either way, and some platforms host servers you write; the table compares running a local server you write with a remote server Bridge generates and runs.
| Local server over stdio | Managed remote server (Kaiva Bridge) | |
|---|---|---|
| Where it runs | On each person's machine, started by their AI client | On Kaiva's managed cloud, at one address |
| Transport | stdio between the client and a local process | MCP over Streamable HTTP, stateless per request (MCP 2026-07-28; earlier versions accepted) |
| Which clients can use it | Desktop apps and coding tools on that machine | Any client that can reach a URL, including Claude and ChatGPT on the web |
| Code to write | A server in Python or TypeScript, with a function per tool | None: tools are generated from an OpenAPI spec, a GraphQL schema, a database, documents or an existing MCP server |
| Sign-in | Usually none: the process runs as the person on their machine | OAuth 2.1 with PKCE and dynamic client registration, so clients register themselves; or a key made for one server |
| Your API's credentials | In environment variables or a config file on every machine | Stored once, encrypted with AES-256-GCM, sent only to your API's host, never shown to an AI client |
| Client keys | None, or one per machine | Kept only as a SHA-256 hash and shown once; a key reaches the servers it names |
| Tool definitions the model reads | Every tool the code defines | Only the tools you expose, so a client's tool list holds what the task needs |
| When the source changes | Update the code and reinstall on every machine | Read the source again; a changed tool is held until you accept it |
| Record of calls | Whatever the server logs | Every tool call: the tool, when, the decision, how long it took and the answer's size |
| Best for | Tools only you use, on your own machine, or custom logic | Your API, database or documents for any client, your team or your customers |
What hosting includes
Read-only by default, with writes optional. Sign-in with OAuth 2.1 and dynamic client registration, keys scoped to one server, a log of every tool call, a live status page and changes to the source held for your review before AI clients see them.
Free includes 10 MCP servers, 200,000 tool calls a month and 200 pages of documents, with no card.
Other ways to do it
- Deploy your ownRun an MCP server on your own cloud or a serverless platform. You manage the runtime, authentication, scaling and updates.
- Self-host BridgeRunning Bridge on your own infrastructure is an Enterprise option.
Questions
- Which transport does it use?
- MCP over Streamable HTTP, at https://api.kaiv.ai/api/bridge/mcp/<name>.
- How do clients authenticate?
- OAuth 2.1 sign-in with dynamic client registration, or a key made for one server.
- Can I put several servers behind one address?
- Yes. Bundles combine servers behind one address.
- Is it free?
- Free includes 10 MCP servers, 200,000 tool calls a month and 200 pages of documents, with no card.