Raw LLM

The simplest way to context engineer.

Minimal, streaming CLI clients for Claude and Gemini that keep your conversations in plain JSON files.

What is this?

Raw LLM is a pair of thin Python scripts that talk to the Anthropic and Google GenAI APIs. No frameworks, no agents, no abstractions you don't need. Just a prompt, a streaming response, and a JSON file you can version, diff, edit, and pipe.

The entire idea: your conversation is a file. You build context by editing that file. That's it. That's the context engineering.

Features

• Streaming output — responses print token-by-token as they arrive
• Conversation persistence — every exchange is saved to a plain JSON file you own
• Resume any conversation — pass the JSON file back in to continue where you left off
• Pipe-friendly — reads from stdin, writes content to stdout, writes diagnostics to stderr
• Colored output — reasoning in gray (stderr), content in cyan (stdout), auto-disabled when piped
• Conflict detection — refuses to overwrite a conversation file modified by another process
• Symlink to switch models — symlink claude.py as opus or haiku to change the default model

Installation

From PyPI

pip install raw-llm

This installs the claude, sonnet, opus, haiku, and gemini commands globally.

From source

git clone https://github.com/rodolfovillaruz/raw-llm.git
cd raw-llm
pip install .

Development install

git clone https://github.com/rodolfovillaruz/raw-llm.git
cd raw-llm
pip install -e ".[dev]"

Set your API keys:

export ANTHROPIC_API_KEY="sk-ant-..."
export GEMINI_API_KEY="..."       # or GOOGLE_API_KEY, per google-genai docs

Usage

Start a new conversation

claude
# Type your prompt, then press Ctrl+D to submit

echo "Explain monads in one paragraph" | claude

gemini

Resume an existing conversation

claude .prompt/some-conversation.json

The JSON file contains the full message history. Edit it with any text editor to reshape context before your next turn.

Pipe a file as context

cat code.py | claude conversation.json

Switch models

# By flag
claude -m claude-opus-4-6

# By command name
opus
haiku
sonnet

Command	Default model
claude / sonnet	claude-sonnet-4-6
opus	claude-opus-4-6
haiku	claude-haiku-4-5
gemini	gemini-3.1-pro-preview

Options

usage: claude [-h] [-n] [-v] [-m MODEL] [-t MAX_TOKENS] [-i] [conversation_file]

positional arguments:
  conversation_file         JSON file to resume (omit to start fresh)

options:
  -n, --dry-run             Build the prompt but don't send it
  -v, --verbose             Show model name and prompt preview
  -m, --model MODEL         Override the default model
  -t, --max-tokens TOKENS   Cap the response length
  -i, --interactive         Interactive REPL mode

Conversation format

Conversations are stored as a JSON array of message objects. Each message carries the standard API fields plus metadata that Raw LLM writes automatically:

[
  {
    "role": "user",
    "content": "What is context engineering?",
    "timestamp": "2025-01-15T10:23:45.123456+00:00"
  },
  {
    "role": "assistant",
    "content": "Context engineering is the practice of ...",
    "timestamp": "2025-01-15T10:23:47.654321+00:00",
    "usage": { "input": 18, "output": 312 },
    "model": "claude-sonnet-4-6"
  }
]

Field	Added to	Description
timestamp	every turn	UTC time the message was appended (ISO-8601)
usage	assistant	input and output token counts for that turn
model	assistant	Model that produced the response

Metadata fields are stripped automatically before the conversation is sent to the API, so you can safely edit or add them without breaking future turns.

You can create these files by hand, merge them, truncate them, or generate them with other tools. Raw LLM doesn't care. It reads the array, appends your new message, streams the response, and appends that too.

Project structure

.
├── src/
│   └── raw_llm/
│       ├── claude.py       # Claude CLI client
│       ├── gemini.py       # Gemini CLI client
│       └── common.py       # Shared utilities (streaming, I/O, conversation management)
├── pyproject.toml          # Package configuration and entry points
├── Makefile                # Formatting, linting, typing
└── .prompt/                # Default directory for conversation files (auto-used if present)

Development

make fmt      # Format with black/isort
make lint     # Lint with pylint/flake8
make type     # Type-check with mypy
make all      # All of the above

Why?

Most LLM tools add layers between you and the model. Raw LLM removes them. The conversation is a file. The prompt is stdin. The response is stdout. Everything else is up to you.

License

MIT