monet.py

A simple DSPy based coding agent with a Textual TUI.

Install

pip install monet-code

Or for development:

git clone https://github.com/wallyqs/monet.py.git
cd monet.py
uv sync

Run

# Set API keys in .env (loaded automatically)
echo 'export ANTHROPIC_API_KEY=sk-...' >> .env

uv run monet

Features

ReAct-based chat

Chat uses ReAct for multi-step reasoning with tool calling:

class ChatSignature(dspy.Signature):  # type: ignore[misc,valid-type]
    """Respond to the user. Use tools when helpful. Maintain context from history.
	...
    Use Rich markup in log.write(): [bold], [dim], [#F4A300]amber[/], etc.
    """

    history: dspy.History = dspy.InputField(
        desc="Prior turns of the conversation."
    )
    message: str = dspy.InputField(desc="The user's current message.")
    response: str = dspy.OutputField(desc="Assistant reply.")

self._react = dspy.ReAct(ChatSignature, tools=self._tools)
return self._react

The agent thinks, calls tools, observes results, and iterates until it has an answer. The trajectory (thinking, tool calls, results) is displayed in the content pane. By pressing ctrl-o you can toggle on/off thinking.

Multi-provider model switching

Switch models at runtime with /model or /model <name>. The current model is shown in the bottom-right corner of the status bar.

Built-in coding tools

The agent has access to these tools:

Self-extending

The agent can add new slash commands to itself at runtime via the register_command. Ask it to "add a /greet monet command" and it will:

1. Create a plugin file at .monet/commands/greet.py
2. Call register_command() to activate it immediately
3. The command persists across restarts

Plugin files follow a simple convention:

NAME = "/greet"
DESCRIPTION = "Greet someone"

def handler(app, args: str) -> None:
    log = app.query_one("#content")
    log.write(f"Hi {args.strip() or 'world'}!")

Commands can call app.predict(signature, **kwargs) to query the LLM using DSPy Predict:

def handler(app, args: str) -> None:
    log = app.query_one("#content")
    result = app.predict("topic -> summary", topic=args.strip())
    log.write(result.get("summary", result.get("error", "")))

Commands also have full access to the Textual framework via the app object. For example, you can ask the agent: "add a /screenshot monet command which uses app.save_screenshot()" and it will generate:

NAME = "/screenshot"
DESCRIPTION = "Take a screenshot of the current application"

def handler(app, args: str) -> None:
    log = app.query_one("#content")
    try:
        app.save_screenshot()
        log.write("Screenshot saved successfully.")
    except Exception as e:
        log.write(f"Error saving screenshot: {e}")

Slash command completion

Type / to see available commands. Tab autocompletes from the filtered list. ctrl+n / ctrl+p navigate the picker.

Input history

ctrl+p / ctrl+n navigate through previously submitted chat messages (when the command picker is not visible). Slash commands are not added to history. In-progress text is stashed and restored.

Keyboard shortcuts

Response rendering

• Markdown with syntax-highlighted code blocks
• ReAct trajectory display (thinking, tool calls, tool results)
• Unified diffs with syntax highlighting after edits
• Animated spinner during LLM calls
• Text selection and auto-copy to clipboard
• Command output is kept in chat context for follow-up conversation

UI behavior

• Banner auto-hides when content overflows
• Typing anywhere auto-focuses the input
• ctrl+o retroactively shows/hides all thinking blocks
• Current model shown in status bar

Testing

uv run pytest