The Model Context Protocol lets an agent or editor call tools on a server you point it at.
vx-build.ai exposes a remote MCP endpoint, so any MCP client can reach open-weight models without
you writing a plugin.
What you get
Two tools, authenticated with the same API key as the REST API:
- list_models — the live catalog of open-weight models
- chat — run a completion against any of them
Usage is metered against your prepaid credits exactly like a REST call, so there is nothing separate
to configure for billing.
The catch worth knowing
Every client uses a different config schema, and one of them has a real gotcha: **Claude Desktop
custom connectors authenticate over OAuth**, so a static API key needs the mcp-remote bridge rather
than the connector UI. Cursor and VS Code accept headers directly, but they name the top-level key
differently — Cursor uses mcpServers, VS Code uses servers.
Rather than reproduce all five here, the documentation has a copy-paste block for Claude
Desktop, Cursor, VS Code, OpenCode and OpenClaw, each with the file path and the safe way to supply
the key for that client.
Do not commit your key
Each client supports some way of keeping the key out of the file:
- Cursor reads shell environment variables with
env:interpolation - VS Code can prompt for it once and store it securely, which makes the config file safe to commit
- Claude Desktop takes it from an
envblock rather than inline in the arguments
Use those. A key pasted into a config file that lands in version control is the most common way these
credentials leak.
Verifying it worked
Restart the client, then ask your agent to list the models available on vx-build. If the connection
is live it will call list_models and come back with the catalog rather than guessing.