Skip to main content

Local (Ollama, LM Studio) for NinjaTrader 8 Trading: Strengths, Costs, and Best Harnesses

You can run NinjaTrader trading workflows entirely against local LLMs via Ollama or LM Studio. The model never leaves your machine; only the MCP tool calls hit the CrossTrade hosted server. For traders with proprietary strategy IP or strict data policies, this is the right pattern.

A summary of A local model for trading work: running the model on your own hardware, where only the MCP tool calls leave the machine. Three columns cover what to reach for it for, what you have to compensate for, and where it runs. Beneath them, the price tier this page states and the harnesses it is known to route cleanly through. The connection path, the tool list and the OAuth scopes are identical whichever model you choose, so only the behaviour shown here changes when you swap one for another.

What this model changes, and what stays the same whichever one you pick.

Why local for trading​

  • Strategy IP stays local. Your NinjaScript and prompts never touch a cloud model provider.
  • No per-token cost. GPU electricity only.
  • Offline-capable for the model half; MCP calls still need internet.

Which model to pick​

ModelStrengthNotes
Qwen 3 CoderBest open-weight tool-callingDefault pick
DeepSeek V3 / R1 (distilled)Reasoning at low costR1 distilled fits on consumer GPUs
Llama 3.x / 4 70BGeneral purposeUse 70B+ for serious MCP work
gpt-ossOpenAI-style tool callingFamiliar prompts

For NinjaScript drafting, none of the local options match Claude. Use cloud Claude for the draft, local Qwen/DeepSeek for compile-repair iterations and journal review.

Stack​

Ollama (Llama / Qwen / DeepSeek / gpt-oss)
   │
   ▼
Open WebUI (or OpenCode / Cline / Continue / LibreChat / AnythingLLM)
   │  MCP
   ▼
CrossTrade hosted MCP server (https://app.crosstrade.io/v1/api/mcp)
   │
   ▼
CrossTrade NT8 Add-On in NT8

The MCP server is hosted; only the LLM is local.

Hardware floor​

  • 24GB VRAM (RTX 4090): Qwen 2.5 14B, DeepSeek R1 distilled, Llama 3 70B Q4.
  • 48GB VRAM: Qwen 3 Coder, Llama 3 70B at higher quants.
  • M-series Mac (64GB+ unified): runs most of the above acceptably.

Limitations​

  • Local models are slower than cloud.
  • Tool-call reliability is worse than cloud Claude / GPT. Pick a tool-aware harness (the recommended list above) and a tool-tuned model (Qwen 3 Coder, DeepSeek with structured output).
  • For mcp:trade workflows, local models are NOT a substitute for careful prompt design. Their refusal behavior is weaker than Claude's.

Pick your harness

This model works through any MCP-capable harness. Recommended pairings:

Related