nemocode 0.1.25

Terminal-first control plane for NVIDIA Nemotron 3 — agentic coding, RAG, doc-ops, and multi-model formations.

NeMoCode

Agentic coding CLI for NVIDIA NIM. Reads your code, makes edits, runs commands — powered by any model on the NIM API or your own GPU via vLLM, SGLang, or TensorRT-LLM.

Community project — not affiliated with or endorsed by NVIDIA.

Install

From PyPI:

pip install nemocode

Recommended (isolated environment):

pipx install nemocode

Or from source (editable):

pip install -e .

Quick Start

Run the guided setup wizard:

nemo setup

The wizard defaults to hosted NVIDIA NIM, prompts for NVIDIA_API_KEY, and can also configure a local vllm, sglang, or trt-llm backend for you.

Hosted NVIDIA NIM (default)

Get a free API key from build.nvidia.com:

export NVIDIA_API_KEY="nvapi-..."
nemo code

Hosted Nemotron endpoints use NVIDIA_API_KEY by default. The setup wizard can store it in your system keyring.

Local vLLM, SGLang, or TensorRT-LLM

Serve a model locally on any NVIDIA GPU:

# vLLM
vllm serve nvidia/NVIDIA-Nemotron-Nano-9B-v2 \
  --host 0.0.0.0 --port 8000
nemo code -e local-vllm-nano9b

# SGLang (best for Nemotron 3 Super long context on DGX Spark)
python -m sglang.launch_server \
  --model nvidia/nemotron-3-super-120b-a12b \
  --host 0.0.0.0 --port 8000
nemo code -e local-sglang-super

# TensorRT-LLM
docker run --rm -it --gpus all --ipc host --network host \
  -e HF_TOKEN=$HF_TOKEN \
  -v $HOME/.cache/huggingface/:/root/.cache/huggingface/ \
  nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc6 \
  trtllm-serve nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8 \
  --trust_remote_code --port 8000
nemo code -e local-trt-llm-nano4b

TensorRT-LLM launch flags can vary by image release and whether a model needs a prebuilt engine. The bundled NeMoCode presets assume the OpenAI-compatible trtllm-serve path with nvidia/nemotron-3-super-120b-a12b and nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8.

No GPU? Rent one via Brev:

nemo setup brev

Usage

nemo code                              # interactive REPL
nemo code "fix the bug in auth.py" -y  # one-shot, auto-approve tools
nemo chat "explain this error"         # chat, no tools
cat log.txt | nemo code "diagnose"     # pipe input
nemo code -f super-nano "refactor"     # multi-model formation
nemo code --tui                        # full-screen TUI

Plan Mode

Plan mode is a read-only planning phase with an approval gate before execution.

• Read-only: Plan mode only reads files, searches code, and explores — no writes, shell commands, or commits.
• Approval gate: The planner proposes a concrete plan. You review and approve, revise with feedback, or cancel.
• Execution: Once approved, a build agent executes the plan with full tool access.

Switch modes in the REPL with Tab or /mode:

Mode	Behavior
code	Ask before tool calls (default)
plan	Read-only planning + approval gate
auto	Auto-approve everything

Launch directly in plan mode:

nemo code --agent plan "implement user auth"

The plan agent can also spawn read-only research subagents to help with exploration.

Endpoints

Works with any OpenAI-compatible API. Pre-configured:

Endpoint	Model	Access
nim-super	Nemotron 3 Super (12B/120B MoE, 1M ctx)	NIM API key
nim-nano	Nemotron 3 Nano (3B/30B MoE, 1M ctx)	NIM API key
nim-nano-9b	Nemotron Nano 9B v2	NIM API key
nim-nano-4b	Nemotron Nano 4B v1.1	NIM API key
nim-vlm	Nemotron Nano 12B VL (vision)	NIM API key
nim-embed	Nemotron Embed 1B v2	NIM API key
nim-rerank	Nemotron Rerank 1B v2	NIM API key
openrouter-super	Super via OpenRouter	OpenRouter key
together-super	Super via Together AI	Together key
local-trt-llm-super	Nemotron 3 Super 120B via TensorRT-LLM	GPU + Docker + TensorRT-LLM
local-trt-llm-nano4b	Nemotron 3 Nano 4B FP8 via TensorRT-LLM	GPU + Docker + TensorRT-LLM
local-vllm-*	Any model on local vLLM	GPU + vLLM
local-sglang-*	Any model on local SGLang	GPU + SGLang
local-nim-*	Local NIM container	GPU + Docker

Formations

Multi-model pipelines — Super plans, Nano executes, Super reviews:

nemo code -f super-nano "implement caching"

Formation	Pipeline
solo	Super does everything (default)
super-nano	Super plans + reviews, Nano executes
spark	All-local on DGX Spark (Super + Nano 9B)
spark-sglang	Super via SGLang on Spark (best long context)
spark-trt-llm	Super 120B + Nano 4B via TensorRT-LLM on Spark
vision	VLM reads screenshots, Super writes code
local	Nano on local GPU, no internet needed

Agents & Sub-agent Orchestration

NeMoCode supports named agent profiles for top-level sessions and delegated sub-agents.

• Primary agents: build (default full-access), plan (read-only planning)
• Sub-agents: general, explore, review, debug, test, doc, code-search, fast
• Inspect them with nemo agent ls and nemo agent show <name>
• Switch primary agents with nemo code --agent <name> or /agent <name> in the REPL/TUI

Sub-agent tools

In coding sessions, these orchestration tools are available:

Tool	Purpose
delegate	Spawn a sub-agent and wait for the result
spawn_agent	Spawn a background sub-agent for parallel work
wait_agent	Wait for a spawned sub-agent to finish
close_agent	Close or cancel a sub-agent handle
resume_agent	Reopen a previously closed sub-agent handle

Sub-agents inherit read-only mode when delegated from plan mode.

Custom agents

Define custom agents in .nemocode.yaml under agents: or as markdown files under .nemocode/agents/*.md:

---
description: Review code for bugs and regressions
mode: subagent
role: reviewer
prefer_tiers:
  - super
tools:
  - fs_read
  - git_read
  - rg
---

Review the requested changes. Focus on correctness, regressions, and missing tests.

Setup Commands

nemo setup          # guided wizard
nemo setup --list   # show all setup topics
nemo setup wizard   # force the interactive wizard
nemo setup trt-llm  # TensorRT-LLM serving guide
nemo setup vllm     # vLLM serving guide
nemo setup sglang   # SGLang serving guide
nemo setup nim      # NIM container guide
nemo setup brev     # rent a cloud GPU

More Commands

nemo endpoint ls / test     # manage endpoints
nemo model ls / show        # inspect model manifests
nemo formation ls / show    # inspect formations
nemo agent ls / show        # inspect agent profiles
nemo hardware recommend     # GPU-based recommendations
nemo doctor                 # run diagnostics to check setup
nemo session ls             # past conversations
nemo obs pricing            # token pricing
nemo init                   # create .nemocode.yaml without overriding user defaults

Contributing

See CONTRIBUTING.md for setup, code style, and PR guidelines.

pip install -e ".[dev]"
ruff check src/ tests/ && ruff format --check src/ tests/
pytest tests/ -v

License

MIT. NVIDIA, Nemotron, and NIM are trademarks of NVIDIA Corporation.

Architecture

  User Input (CLI / TUI)
         |
  CodeAgent (orchestrator)
    |  agent profiles, project context, git state, memories
    v
  Scheduler (formation pipeline driver)
    |  single-model or plan/execute/review formation
    |  stagnation detection, auto-compaction, permission engine
    v
  Registry (endpoint / manifest / formation resolver)
    |
    v
  Providers (NIM Chat, Embeddings, Rerank)
    |  OpenAI-compatible API, SSE streaming, retry with backoff
    v
  ToolRegistry (18+ tools)
    - fs: read, write, edit, multi_edit
    - git: status, diff, log, commit, snapshot
    - search: rg, glob
    - bash: shell command execution
    - agent: delegate, spawn, wait, close, resume
    - memory: save, recall
    - web, parse, test, ask_user, clarify, LSP, MCP

Comparison

Feature	NeMoCode	Claude Code	Cursor	OpenCode
Open source	MIT	No	No	MIT
Terminal-first	Yes	Yes	IDE	Yes
Multi-model formations	Yes	No	No	No
Local GPU serving	vLLM, SGLang, TRT-LLM	No	No	No
Hardware detection	Yes (GPU/RAM/Spark)	No	No	No
1M token context	Yes (Nemotron 3)	200K	128K	Varies
Sub-agent orchestration	Yes	No	No	Yes
LSP integration	Yes	No	Yes	Yes
Vision (screenshots)	Yes (VLM)	Yes	Yes	Yes
Plugin system	Yes	No	No	Yes
NVIDIA NIM native	Yes	No	No	No