Neo-mcp

Neo là tác nhân kỹ thuật AI tự động đầu tiên có khả năng lập kế hoạch, nghiên cứu và thực hiện công việc đa bước cho các tác vụ như xây dựng tác nhân AI, tinh chỉnh và đánh giá mô hình AI, cũng như các pipeline học máy; sử dụng mã nguồn, dữ liệu và thí nghiệm của bạn để đẩy nhanh tiến độ với các tạo phẩm có thể kiểm tra. Nó có thể suy luận trên kho lưu trữ của bạn, kết nối công cụ và truy xuất, gỡ lỗi các lần chạy huấn luyện, và giúp bạn phát triển các quy trình AI sẵn sàng cho sản xuất.

Xem trên heyneo.com

Tài liệu

NEO MCP

NEO MCP connects your AI editor to NEO - an autonomous AI engineering agent - via the Model Context Protocol. Submit long-running tasks (fine-tuning, RAG pipelines, benchmarking, evals) directly from Claude Code, Cursor, Windsurf, VS Code, or any MCP-compatible editor, and get results synced back to your project folder.

Homepage: https://heyneo.com
Docs: https://docs.heyneo.com/neo-mcp
Install: pip install neo-mcp
PyPI: https://pypi.org/project/neo-mcp/
Tags: AI Agents, Developer Tools, Machine Learning, Task Automation, LLM, Fine-tuning

About

NEO is an autonomous AI engineering agent that plans, researches, and executes multi-step ML and AI engineering work, including fine-tuning, quantization, benchmarking, evals, and agentic pipeline construction, using your codebase, data, and experiments.

The neo-mcp package runs a local MCP server that bridges any MCP-compatible editor to the NEO backend. Tasks run autonomously and results are written back to your workspace.

Supported editors: Claude Code, Cursor, Windsurf, VS Code, Zed, Continue, Codex CLI, Antigravity

Installation

pip install neo-mcp

If you get error: externally-managed-environment, run:

python3 -m pip install --user --break-system-packages neo-mcp

Quick Start (Claude Code)

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

Get your API key at: https://heyneo.com/dashboard

MCP Tools

Tool	What it does
`neo_submit_task`	Start a NEO task; returns a `thread_id`. Optional `wait_for_completion` blocks until done.
`neo_task_status`	Poll task status (`RUNNING`, `COMPLETED`, `WAITING_FOR_FEEDBACK`, etc.).
`neo_get_messages`	Read the task transcript after completion.
`neo_list_tasks`	List recent or active tasks.
`neo_send_feedback`	Reply when NEO is waiting for user input.
`neo_pause_task` / `neo_resume_task`	Pause or resume a running task.
`neo_stop_task`	Stop and tear down a task.
`neo_add_integration`	Store a provider key (GitHub, HuggingFace, Anthropic, OpenRouter, or any custom secret) locally for use in tasks.
`neo_list_integrations`	List configured providers (names only, never the secret).
`neo_test_integration`	Validate a stored key against the provider's API.
`neo_remove_integration`	Delete a stored key.

Typical flow: neo_submit_task -> poll neo_task_status -> neo_get_messages (or use wait_for_completion to skip polling).

Integrations

Give NEO access to external services by storing API keys locally once. Keys are injected as environment variables on every task run and never leave your machine.

Provider	Credential	What NEO can do with it
`github`	Personal Access Token (`ghp_...`)	Clone private repos, push, open PRs
`huggingface`	Token (`hf_...`)	Download private models and datasets
`anthropic`	API key (`sk-ant-...`)	Run Claude models inside tasks
`openrouter`	API key (`sk-or-...`)	Route through any model OpenRouter supports

Security

Keys are stored locally at ~/.neo/integrations/<provider>.env with 0o600 permissions.
Keys are never sent to the NEO backend, written to logs, or committed to git.
neo_list_integrations returns provider names only, never the secret itself.

NEO Research Inc. - https://heyneo.com