MCP

Your private model, inside every tool

Fine-tune a model on Nucleus, then expose it behind MCP. The editors and agents your teams already use connect with one configuration block.

mcp.json
{
  "mcpServers": {
    "nucleus": {
      "url": "https://api.nucleus-ai.io/mcp",
      "headers": { "Authorization": "Bearer nk_live_8x…" }
    }
  }
}
ClientsStreamable HTTP · JSON-RPC
Claude Desktop
connected
Claude Code
connected
Any MCP client
same server entry
Key
nk_live_8x… · read, infer

One endpoint, every client

Your fine-tuned model is reachable from any tool that speaks MCP. The client sends a message; Nucleus handles the inference.

Scoped credentials

The key the client holds carries only the scopes it needs. Bind it to one project, and a leaked key reaches that project alone.

The same key, three routes

One API key authenticates MCP, the OpenAI-compatible route and the Anthropic-compatible route. No separate credentials per protocol.

Expose your fine-tuned model to MCP clients

A model trained on Nucleus is served through a private endpoint. The MCP server wraps that endpoint so any MCP-compatible tool can reach it. The client authenticates with your API key and addresses the model by its checkpoint path or base model ID.

Results

Results from LLMs powered by Nucleus

$177K

downtime prevented this month

47 vehicles

rerouted in a single fleet optimisation

$73.4K

monthly revenue re-attributed in marketing

94%

prediction accuracy on equipment failures

Manufacturing

Line 3's next failure, booked into the maintenance window

$177Kdowntime prevented
18.3 daysadvance notice
Fleet operations

47 trucks rerouted while they were still moving

$2.4Kdaily fuel saved
23delays resolved without a call
At a glance

What MCP on Nucleus includes

Protocol

  • MCP compliant
  • Streamable HTTP transport
  • JSON-RPC messages

Authentication

  • Scoped API keys (read, train, infer, admin)
  • Project-bound credentials
  • Environment prefixes: nk_live_, nk_test_

Compatibility

  • OpenAI-compatible route (/v1)
  • Anthropic-compatible route (/anthropic)
  • One key for all three protocols

Governance

  • Four access policy gates on every call
  • Append-only audit trail
  • Revocation on the next request
Pricing

Served tokens

Served tokens

$0.15 – $6.50per million tokens

Calls through MCP are metered the same way as any private endpoint: per million input and output tokens. Language models from $0.15/Mtok (Nemotron 3.5 Lightning) to $6.50/Mtok (Kimi K3).

Pick your path

Pick your path

Harness

Permissions, scoped keys and audit

Explore Harness

On-Premise Deployment

Our GPU servers, your site

Explore On-Premise Deployment

Tool use

Your private APIs as skills

Explore Tool use

RAG

Answers from your documents

Explore RAG

Webhooks

Signed events to your systems

Explore Webhooks

Put your private model in every tool your team uses

Fine-tune a model, configure MCP, and the editors and agents connect the same day.