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:

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:

Formations

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

nemo code -f super-nano "implement caching"

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:

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