LLaMACode

A multi-model coding agent powered by NVIDIA Llama that runs in your terminal. Reads, writes, and edits files, runs shell commands, searches code, browses the web, and more — all through a natural-language chat interface.

Features

• Two model tiers — Llama 3.2 3B (balanced, default) and Llama 3.3 70B (most powerful)
• Autonomous tool loop — the agent plans, searches, edits, and verifies code on its own
• File operations — read, write, edit files, run shell commands, glob, grep
• Web search — search the internet and fetch page content (no API key needed)
• Session persistence — save/load conversations, undo exchanges, compact context
• Multi-agent architecture — planner, searcher, coder, reviewer, summarizer roles
• Animated spinner — shows which phase/agent is working in real time

Installation

pip install llamacode

Quick Start

You need an NVIDIA API key to use Llama models. The easiest way to get one:

llamacode --generate-key

If llamacode is not on your PATH (common on Windows with user installs), use:

python -m llamacode --generate-key

This opens build.nvidia.com in a browser. Select the model (3B or 70B), log in to your NVIDIA account, and the key is saved automatically.

Then just run:

llamacode

You'll see an interactive model picker:

  ═══════════════════════════════════════════════════════════════
     LLaMACode - Choose a Model
  ═══════════════════════════════════════════════════════════════

  [1] Llama 3.2 3B        Balanced, default model             Key: ✓
  [2] Llama 3.3 70B       Most powerful (needs your own key)  Key: ✗

  Select [1-2] (Enter to keep current):

API Keys

Every user must provide their own NVIDIA API key. There is no bundled key.

Option 1 — Generate via browser (easiest)

llamacode --generate-key

Prompts you to pick 3B or 70B, opens build.nvidia.com, detects and saves the key automatically. Requires Playwright (installed on first use).

Option 2 — .env file

Create a .env file in one of these locations:

Location	Path
Home directory (recommended)	%USERPROFILE%\.coding_agent\.env
Current working directory	.env

Contents:

NVIDIA_API_KEY_LLAMA_3_2_3B=nvapi-your-3b-key-here
NVIDIA_API_KEY_LLAMA_3_3_70B=nvapi-your-70b-key-here    # optional

Option 3 — Environment variable

set NVIDIA_API_KEY_LLAMA_3_2_3B=nvapi-your-key      # cmd
$env:NVIDIA_API_KEY_LLAMA_3_2_3B="nvapi-your-key"   # PowerShell
export NVIDIA_API_KEY_LLAMA_3_2_3B=nvapi-your-key   # bash

llamacode checks these env var names (in order):

Model	Env vars checked
3B	NVIDIA_API_KEY_LLAMA_3_2_3B, NVIDIA_LLAMA_3_2_3B_API_KEY, NVIDIA_API_KEY_3B
70B	NVIDIA_API_KEY_LLAMA_3_3_70B, NVIDIA_LLAMA_3_3_70B_API_KEY, NVIDIA_API_KEY_70B, NVIDIA_API_KEY

Usage

llamacode                            # interactive CLI with model picker
python -m llamacode                   # same as above (fallback if not on PATH)
llamacode --model llama-3.2-3b       # skip picker, use 3B
llamacode --model llama-3.3-70b      # skip picker, use 70B
llamacode --generate-key             # generate API key via browser
python -m llamacode --generate-key     # same, fallback

CLI Commands

Command	Description
/exit	Exit the CLI
/new	Start a new session
/clear	Clear conversation history
/status	Show session info, model, workdir, key status
/model	Open interactive model picker
/model <name>	Switch model (llama-3.2-3b, llama-3.3-70b)
/workdir	Show working directory
/workdir <path>	Change working directory
/index	Build .agent/project_index.json
/save	Save conversation to timestamped file
/save <file>	Save conversation to a specific file
/load <file>	Load a saved conversation
/undo	Remove the last exchange
/compact	Summarize conversation to save context
/help	Show help

Switching Models at Runtime

/model llama-3.2-3b    → Llama 3.2 3B (balanced, default)
/model llama-3.3-70b   → Llama 3.3 70B (most powerful)

The 70B model switches immediately only if you have an NVIDIA API key configured (see API Keys).

Available Models

Alias	Model ID	Size	Key Required
llama-3.2-3b	meta/llama-3.2-3b-instruct	3B	User-provided
llama-3.3-70b	meta/llama-3.3-70b-instruct	70B	User-provided

• 3B — handles coding, reviewing, and summarizing (balanced, default).
• 70B — most powerful; best for complex reasoning tasks. Requires your own NVIDIA API key.

Agent Capabilities

LLaMACode has a built-in tool loop that lets the agent autonomously work through tasks:

Tool	Description
read_file	Read a file chunk by line range
write_file	Write content to a file (creates directories if needed)
edit_file	Replace specific text in an existing file
bash	Run a shell command with timeout
glob	Find files matching a pattern
grep	Search file contents and indexed symbols
web_search	Search the internet via DuckDuckGo
web_fetch	Fetch and extract text from a URL
think	Internal reasoning (invisible to the user)

Project Structure

llamacode/
├── coding_agent.py        Main CLI entry point
├── key_generator.py       Browser-based API key generation
├── api_server.py          Optional FastAPI server
├── pyproject.toml         Package config
├── MANIFEST.in            Build exclusions
├── agents/                Multi-agent wrappers
│   ├── planner_agent.py
│   ├── search_agent.py
│   ├── coder_agent.py
│   ├── reviewer_agent.py
│   └── summary_agent.py
└── core/                  Core modules
    ├── model_manager.py    Model & key resolution
    ├── context_manager.py  Message trimming
    ├── file_manager.py     Read/write/edit files
    ├── project_index.py    Symbol indexing
    └── summary_cache.py    Summary caching

Development

# Clone the repo
git clone https://github.com/anomalyco/third_party_connect.git
cd third_party_connect

# Editable install
pip install -e .

# Set up API keys
echo NVIDIA_API_KEY_LLAMA_3_2_3B=nvapi-your-key > .env

# Run
llamacode

Publishing to PyPI

Push a version tag to trigger the automated workflow:

git tag v1.2.1
git push origin v1.2.1

The GitHub Actions workflow (.github/workflows/publish.yml) builds, checks, and publishes to PyPI using trusted publishing (OIDC).

Manual publish:

pip install build twine
python -m build
python -m twine upload dist/*