Neo MCP: Your autonomous AI engineering agent

neo-mcp is the Model Context Protocol server that plugs Neo into Claude Code, Cursor, Codex, and the editors you already use. Neo is an autonomous AI engineer: describe any AI/ML task in plain English and it plans, builds, runs, and evaluates the full workflow.

Everything lands in your repo on your machine (code, models, metrics, reports). Nothing is stored remotely.

Neo is built for AI engineering, not general code chat, so it goes deeper on ML, LLM, and data workflows than a typical coding agent.

🌐 Neo · 📚 Docs · 🔑 Get an API key: Neo dashboard

See it in action

Click to watch the full demo with sound.

Related MCP server: Nexus-MCP

What MCP unlocks

• 🧩 Stay in your editor. Drive Neo from Claude Code, Cursor, VS Code (Copilot), Windsurf, Zed, Continue, or Codex. No new app, no context switching.

• 🔬 Go deeper on AI/ML. Autonomous planning, experiments, evaluation, and iteration tuned for real ML work, not just one-shot answers.

• 💾 Local-first. Every output file is written to your machine. Nothing is stored remotely.

What you can build with Neo

• 🤖 Generative AI & LLMs: RAG, semantic search, agents, chatbots, fine-tuning (Llama, Qwen, Gemma), document analysis

• 🧠 ML & deep learning: PyTorch, TensorFlow, scikit-learn training, architecture search, evaluation

• 📊 Data science & analytics: EDA, feature engineering, forecasting, segmentation, A/B testing, reporting

• 👁️ Computer vision: image classification, object detection, OCR

• 🎤 Speech & audio: speech-to-text, text-to-speech, audio classification

• 🔌 Bring your own keys: GitHub, HuggingFace, Anthropic, OpenRouter, OpenAI, AWS S3, Weights & Biases, Kaggle. Stored locally and injected as env vars.

Built for data scientists, ML engineers, analysts, researchers, and PMs who want results, not boilerplate.

Try it

Ask your agent to use Neo. For example:

Use Neo to fix the failing training run and re-run with logging

Benchmark these prompts on our eval set using Neo

Build or debug an end-to-end ML pipeline using Neo

Train a fraud detection model on fraud.csv, optimize for recall

Fine-tune a text classifier on my training data with 5-fold cross-validation

Neo runs the ML work. Your editor handles everything else.

Install

pip install neo-mcp

Requires Python 3.11+.

Tip: use pipx install neo-mcp to install in an isolated environment and avoid conflicts with your project's virtualenv.

Use Neo from your editor

Replace sk-v1-YOUR_KEY with your actual API key.

After setup, ask your agent: "What Neo tools do you have available?" You should see neo_submit_task, neo_task_status, neo_get_messages, and the rest.

Claude Code

claude mcp add --scope user neo \
  -e NEO_SECRET_KEY=sk-v1-YOUR_KEY \
  -- neo-mcp

Open a new Claude Code session after running this. Neo tools load at session start, not mid-session.

Scope options: --scope user (global, recommended) · --scope project (writes .mcp.json in current repo) · --scope local (this machine only)

Verify it registered:

claude mcp list

You should see neo with a green checkmark.

Cursor

Open the config:

• GUI: Ctrl+Shift+J (Windows/Linux) or Cmd+Shift+J (Mac) → Tools & MCP → New MCP Server

• Or edit the file directly: ~/.cursor/mcp.json

{
  "mcpServers": {
    "neo": {
      "command": "neo-mcp",
      "env": { "NEO_SECRET_KEY": "sk-v1-YOUR_KEY" }
    }
  }
}

Restart Cursor after editing the file directly. Changes via the GUI apply immediately.

OpenAI Codex CLI

Open the config:

• Run codex mcp to manage servers interactively via CLI

• Or edit the file directly: ~/.codex/config.json

{
  "mcpServers": {
    "neo": {
      "command": "neo-mcp",
      "env": { "NEO_SECRET_KEY": "sk-v1-YOUR_KEY" }
    }
  }
}

Also works with

Windsurf (~/.codeium/windsurf/mcp_config.json):

{
  "mcpServers": {
    "neo": {
      "command": "neo-mcp",
      "env": { "NEO_SECRET_KEY": "sk-v1-YOUR_KEY" }
    }
  }
}

VS Code (GitHub Copilot) (.vscode/mcp.json in your workspace root; requires VS Code 1.99+ and Agent mode):

{
  "servers": {
    "neo": {
      "type": "stdio",
      "command": "neo-mcp",
      "env": { "NEO_SECRET_KEY": "sk-v1-YOUR_KEY" }
    }
  }
}

Zed (~/.config/zed/settings.json):

{
  "context_servers": {
    "neo": {
      "source": "custom",
      "command": {
        "path": "neo-mcp",
        "args": [],
        "env": { "NEO_SECRET_KEY": "sk-v1-YOUR_KEY" }
      }
    }
  }
}

Continue.dev (~/.continue/config.yaml):

mcpServers:
  - name: neo
    command: neo-mcp
    env:
      NEO_SECRET_KEY: sk-v1-YOUR_KEY

More GUI paths and per-editor notes: docs/GUIDE.md

How it works

Your editor  ──MCP──▶  neo-mcp server  ──API──▶  Neo backend
                            │                          │
                            └──────────────────▶  Local daemon
                                                  (writes files,
                                                   runs scripts)

1. You describe a task: "Train a fraud detection model on data.csv"

2. The editor calls neo_submit_task via MCP

3. Neo's backend processes the task and sends commands to the local daemon

4. The daemon runs on your machine, writing files, running scripts, and installing packages

5. Output files appear directly in your workspace

Files are always written to your machine, never stored remotely.

Neo can also store third-party API keys locally (GitHub, HuggingFace, Anthropic, OpenRouter, OpenAI, AWS S3, Weights & Biases, Kaggle) so tasks can use them without asking every time. Keys stay on your machine and are never sent to Neo's backend. Full guide: docs/INTEGRATIONS.md.

Tools

Integration tools store credentials locally (file mode 0o600 under ~/.neo/integrations/, or native tool files like ~/.aws/credentials, ~/.netrc, ~/.kaggle/kaggle.json), or your OS keyring if NEO_INTEGRATIONS_BACKEND=keyring. Keys never leave your machine. Full guide: docs/INTEGRATIONS.md.

Workflow

Standard (tasks over a few minutes):

neo_submit_task  →  returns thread_id
      ↓
neo_task_status  →  poll until COMPLETED or WAITING_FOR_FEEDBACK
      ↓
neo_get_messages →  read full output

Quick task: Pass wait_for_completion: true to neo_submit_task. It blocks until done and returns output directly. No polling needed.

Mid-task question: When status is WAITING_FOR_FEEDBACK, call neo_send_feedback with your reply. Neo resumes automatically.

Reconnecting after closing your editor:

neo_list_tasks   →  all tasks with live status + thread IDs
neo_task_status  →  check the specific task you care about
neo_get_messages →  read output of any COMPLETED task

Environment variables

Diagnostics

neo-mcp status      # daemon and key status
neo-mcp doctor      # full health check; identifies common issues
neo-mcp list        # list known threads
neo-mcp logs --source neo-mcp --lines 100   # MCP server logs
neo-mcp logs --source daemon --lines 100    # daemon logs

# JSON output
neo-mcp status --json
neo-mcp doctor --json

Claude Code logs:

claude mcp logs neo

Troubleshooting

Full setup guide (all editors, GUI paths): docs/GUIDE.md · Docs