Sloppy
Sloppy is a terminal app where two AI models discuss an initial prompt. AI 1 replies first, then AI 2, alternating for a chosen number of replies. It supports OpenAI-compatible chat-completions APIs, including local servers.
Setup
Use Python 3.10 or newer and configure two reachable models. The same endpoint can serve both.
python -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
Copy-Item .env.example .env
Edit .env for your endpoints, keys, model IDs, and topic, then run python run_example.py. On macOS or Linux, activate with source .venv/bin/activate and copy the example with cp .env.example .env. python ai_chat.py is an equivalent entry point.
python run_example.py --prompt "Can AI be creative?" --turns 6
python run_example.py --help
Configuration
.env accepts AI1_NAME, AI1_BASE_URL, AI1_API_KEY, AI1_MODEL, and AI1_SYSTEM_PROMPT, plus matching AI2_* variables. INITIAL_PROMPT is required. MAX_TURNS defaults to 20 and must be positive. Set each base URL to an API root such as http://localhost:11434/v1; Sloppy appends /chat/completions.
Command-line options --ai1-name, --ai1-url, --ai1-key, --ai1-model, --ai1-prompt and matching --ai2-* options override .env values for a run. When any connection option is supplied, both connections need a base URL and API key after applying defaults. Prefer environment variables over command-line arguments for hosted-provider secrets, since command-line values can enter shell history.
Press Ctrl+M to add, edit, delete, and select saved connections. They are stored in sloppy.json in the current working directory:
{
"models": {
"local": {
"name": "Local Llama",
"base_url": "http://localhost:11434/v1",
"api_key": "env:LOCAL_API_KEY",
"model": "llama3",
"system_prompt": "You are a helpful assistant."
}
},
"selected": {"ai1": "local", "ai2": "local"}
}
Set LOCAL_API_KEY in the process environment when using env:LOCAL_API_KEY. The app resolves it when selecting the connection. You can select the current .env connection for either speaker. .env values are reread on launch and are never copied into sloppy.json. With complete .env connections and no sloppy.json, the conversation starts directly; otherwise a picker opens. Changes to speaker selection during a conversation apply on the next run.
Controls and exports
| Key | Action |
|---|---|
| Ctrl+M | Connections and speaker selection |
| Ctrl+P | Command palette |
| F1 | Reading controls |
| S or F2 | Sharing menu |
| Mouse wheel, Up / Down, Page Up / Page Down | Scroll |
| Home / End | Oldest / latest message |
| Ctrl+C | Quit |
Scrolling up pauses automatic following; End resumes it. The UI shows elapsed time, reply count, word counts, and token usage when the provider reports it. Token and speed figures may be estimates when usage is unavailable.
The sharing menu saves the full conversation or visible screen as SVG in screenshots/, or a Markdown transcript in transcripts/. The transcript includes the initial prompt and replies in order. Both directories are created in the current working directory and ignored by Git.
Troubleshooting
- No connection in the picker: Complete the base URL and key in
.env, or add a saved connection with Ctrl+M. - Missing API key: Set the environment variable named by a saved
env:NAMEreference before launch. - API error: Check the API root, model ID, server availability, and credentials. A failed or incomplete stream falls back to a non-streaming request; transport and empty-response failures receive up to two retries.
- Invalid settings: Fix malformed
sloppy.jsonin the current working directory.modelsandselectedmust be objects; connection fields and selections must be strings. The error identifies the invalid field.
See CONTRIBUTING.md for code structure and tests.