papote

Fully vibe-coded. Every line of this was written by an LLM, reviewed only as far as a conversation allows. Use at your own risk.

A terminal chat with a local LLM whose answers are formatted while they stream in — headings, tables, boxed code with syntax highlighting, all re-wrapped live as the answer grows.

❯ Explain TCP briefly, with a code block and a table.

TCP (Transmission Control Protocol) is a transport-layer protocol that
guarantees reliable, ordered delivery between two hosts.

╭─ rust ───────────────────────────────────────────────────────────────────────╮
│ use std::net::TcpStream;                                                     │
│ let mut stream = TcpStream::connect("127.0.0.1:8080")                        │
│     .expect("cannot connect");                                               │
╰──────────────────────────────────────────────────────────────────────────────╯

┌────────────────┬───────────────────────────┬────────────────────────────┐
│ Property       │ TCP                       │ UDP                        │
├────────────────┼───────────────────────────┼────────────────────────────┤
│ Reliability    │ Guaranteed (retransmits)  │ Best-effort                │
│ Mode           │ Connection-oriented       │ Connectionless             │
└────────────────┴───────────────────────────┴────────────────────────────┘

Getting started

You need an OpenAI-compatible server that streams SSE — typically llama-server. http:// and https:// both work, and endpoints behind HTTP basic auth are supported.

cargo install --git https://git.tforgione.fr/tforgione/papote
papote

Nothing is configured out of the box. The first run asks for an endpoint, then lists the models it offers and lets you pick one, and writes both down.

What it does

• Renders markdown as it arrives: headings, lists, quotes, tables aligned to the terminal, boxed code blocks coloured for a dozen languages, clickable links.
• Full-screen: the conversation scrolls, the input box grows with your message, and resizing re-wraps everything mid-answer.
• Saves every conversation, names it by asking the model, and /resume picks one back up where you left it — or papote --private for one that is never written down.
• Talks to several endpoints — /provider and /model switch between them without restarting.
• /system sets the system prompt, shown at the top of the conversation and saved with it.
• /image ~/Images/moi.jpg hangs a picture on the input line, to ask your question beside it; /model says which models can see. A picture copied in a file manager can be pasted in, and Ctrl-V attaches a screenshot straight off the clipboard.
• The model can ask papote for what it cannot know: get_datetime gives it the local time, and the call and its answer show in the conversation.
• A model that thinks before answering has its thinking streamed, shown and counted, rather than a still screen and a silent bar.
• Tab completes the commands and the path of an image, then walks the candidates one press at a time, the one in the line highlighted on the bar.
• Ctrl-C stops an answer that has gone off the rails and keeps what arrived.

Commands

/resume	carry on with a saved conversation
/image <path>	attach a picture to your next message
/system	set the system prompt, or drop it
/rename	name it, or let the model find one
/model, /provider	switch model or endpoint
/clear	start a new conversation
/help	the rest of them

Documentation

The details live in the wiki:

• Installing — requirements and the first run
• Configuration — the three TOML files, every key
• Commands and keys — the full keyboard
• Saved conversations — where they live, how /resume works
• How it works — the screen, the streaming, the rendering
• Limitations — what it deliberately does not do
• Development — tests, the file map, the dependencies

Development

cargo test && cargo clippy --all-targets && cargo fmt

See CLAUDE.md for the invariants worth knowing before changing the renderer or the display.