# Extend Claude limits by offloading AI tasks to Neo

> Source: <https://heyneo.com/claude-code>
> Published: 2026-06-19 15:20:44+00:00

Add NEO's MCP server to any environment with Python 3.11+.

Open your NEO dashboard, create a key, and copy it. Keys look like sk-v1-….

Register NEO with one command, then just ask in a new chat to ship work.

```
claude mcp add --scope user neo \
  -e NEO_SECRET_KEY=sk-v1-your-key \
  -- python3 -m neo_mcp
```

Using Cursor, VS Code, or another MCP client? [See the neo-mcp setup](/neo-mcp).

Independent benchmark

The task: benchmark a speech-to-text model on a CPU-only Azure VM — 2 cores, 7.7 GB RAM, no GPU. Claude Code alone reached for the obvious HuggingFace + PyTorch path and iterated in real time. NEO spent two minutes researching first, then chose ONNX Runtime for its AVX2-optimized CPU kernels — same task, same machine.

Cost per task

$1.96$0.74

62% cheaper

Runtime (RTF)

0.5190.328

37% faster

Backend chosen

PyTorchONNX Runtime

CPU-optimized

Benchmark by Gaurav Vij · [Read the full writeup](https://medium.com/@gauravvij/claude-code-an-ai-agent-cut-my-ai-engineering-costs-by-62-claude-code-alone-couldnt-840c6e8502e9)

Install once, then delegate ML work to NEO from any Claude Code session.