Use Vinci with MCP tools
Under maintenance. The managed Vinci API and Platform are under maintenance. These service-specific instructions are retained as reference for existing integrations when service is available. They do not promise current access or a return date. Downloadable models and direct/BYOK use of Vinci Code CLI are separate paths.
The Model Context Protocol (MCP) is the standard way an agent reaches external tools and data. In an MCP setup there are three roles: the servers (which expose tools), the host (which connects to those servers and routes calls), and the model (which decides which tool to call and when). The managed Vinci API serves the selected model; Vinci names SimpleDirect’s broader AI research, models, technologies and products.
A host collects the tool definitions from its MCP servers and hands them to the model as standard function/tool schemas — exactly the wire format Vinci already speaks (Tool calling). A compatible host can use this interface when the managed service is available. Check that the host supports the endpoint’s tool-call and streaming formats.
In this integration, your application is the MCP host and operates the server connections. The selected model proposes tool calls; your host executes them.
Hosted-service settings (reference)
| Setting | Value |
|---|---|
| Base URL | https://vinci.getsimpledirect.com/api/v1 |
| API key | a vinci_live_… key — issue one at the Vinci Platform ↗ |
| Model | auto (your account’s class), or pin mezzo / forte / fortissimo |
Wherever your host asks for the “model provider” or “OpenAI-compatible endpoint,” use those
three. The last documented service limits were a 1,000,000-token context window
and up to 131,072 output tokens; confirm applicable limits when service returns. Vinci forwards your tools / tool_choice straight through and returns real
tool_calls in OpenAI’s format, streaming or not, across multiple turns — see the
API reference for the exact shape.
Which hosts work
Any host that lets you choose an OpenAI-compatible model:
- Any MCP host with a configurable OpenAI-compatible model.
- Framework-based hosts whose MCP adapters call the model through a configurable OpenAI-compatible client.
- Any custom host built on an OpenAI-compatible client.
Hosts that are locked to their own model can’t use Vinci for tool selection. That’s a limit of the host, not of Vinci.
What to expect
The last documented class identifiers were mezzo, forte and fortissimo,
with auto following the account’s choice. piano was reserved. These are routing
references, not current availability claims; confirm the service’s model list after
maintenance. No release date for a reserved identifier is stated.
Start with a focused server or two and grow from there.
Notes
- This is the same direct pass-through coding agents use, so the plain-chat grounding
features (auto web search, memory, attachments) don’t apply on a
toolsrequest — your MCP servers are the grounding. - Every call still gets the Vinci character layer, credit-based billing, and Zero Data Retention by default.
See Getting started for the general API setup, or the Chat Completions reference for the tool-calling wire format.