Use Vinci as the reasoning model inside a coding assistant that already connects to MCP servers. This recipe is for power users who want their current host and tools to stay in place while Vinci decides which tool to call and when.
Vinci works with any MCP host that accepts a configurable OpenAI-compatible model. The host remains responsible for connecting to servers and executing tools; Vinci does not run the servers or act as the host.
- Check your assistant’s provider settings. Continue only if it lets you set an OpenAI-compatible endpoint and model. A host locked to its own model cannot use Vinci for tool selection.
- In the Vinci Platform, issue a
vinci_live_…API key and copy it when shown. Give it the scopes required by your host;inferencecovers chat-completion calls, whilemodelscovers model discovery. - Set the base URL to
https://vinci.getsimpledirect.com/api/v1. - Paste the key into the host’s API-key field and set the model to a class id —
mezzo,forte, orfortissimo. For code and file tasks, useforte; Mezzo is fast and light for everyday chat, while Fortissimo is the deepest class for the hardest work. Hosts that require the field to be set should use a class id rather thanauto. - Connect one or two focused MCP servers in the host. Their tool definitions are passed to Vinci as standard function schemas, so no Vinci-specific adapter is needed.
- Start with a narrowly scoped request that makes the expected tool obvious. Confirm that the host receives a tool call, executes it, and returns the result.
- Add more servers or multi-step tasks only after the focused workflow works reliably.
On tool-enabled requests, your MCP servers provide the grounding. Plain-chat features such as automatic web search, memory, and attachments do not apply on this path.