NeuralDeepCode — AI skill for Claude Code

NeuralDeepCode (`ndcode`)

**Terminal AI coding agent for the [NeuralDeep](https://neuraldeep.ru) hub.**

`ndcode` is a terminal coding agent that talks to the self-hosted NeuralDeep LLM hub — RU-hosted Qwen3 / gpt-oss models on your own GPUs. Log in with your browser, see your tier and budget right inside the TUI, and start coding — no other accounts, no provider juggling.

The idea

Most terminal coding agents are tuned for Claude / GPT. `ndcode` takes a mature open-source agent loop ([opencode](https://github.com/sst/opencode), MIT) and **adapts it for working with Qwen models** served by the NeuralDeep hub — Qwen3-Coder / Qwen3 (`qwen3.6-35b-a3b`) and `gpt-oss-120b`. The goal is a coding agent that feels native to *these* models: model-specific system prompts and tool-calling tuned for Qwen, the hub as the default (and only) provider, browser SSO login, and budget/limits visible in the TUI — instead of a generic client pointed at someone else's cloud.

Benchmark

A small agentic coding suite ([`bench/`](bench)) — each task runs the agent headless in a clean repo and is scored by its own tests passing. On the hub's default model:

model	suite	pass-rate
qwen3.6-35b-a3b	bench/ (6 tasks: implement + bug-fix, Python/JS)	6 / 6

Run it yourself (needs a working `ndcode` + `/login`, plus `python3` and `node`):

bench/run.sh                                 # default qwen3.6-35b-a3b
MODEL=neuraldeep/gpt-oss-120b bench/run.sh   # compare another hub model

Note the hub's `max_parallel_requests` limit (tier-dependent) — the runner clears stray runs and spaces tasks so a correct rate-limit back-off isn't counted as a failure.

Hub integration

command	what it does
/login	Browser SSO (localhost-callback, like gh auth login). Open