A native AI research platform built with Rust. Send the same prompt to multiple models simultaneously, compare responses side-by-side, diff where they agree or disagree, and track costs — all from a single keyboard-driven interface. Zero Electron, zero JavaScript.

Most AI chat apps let you talk to one model at a time. Stoa lets you interrogate multiple models in parallel and see where they converge or diverge. Built for researchers, engineers, and anyone who wants to think harder with AI.

• Native performance — Pure Rust + GPU rendering via iced. Starts instantly, uses minimal resources.
• Local-first — All data stays on your machine. SQLite database, no cloud sync, no telemetry.
• Keyboard-driven — Quick Switcher, Command Palette, and shortcuts for everything.

• Parallel streaming — Send to 2-7+ models at once, watch all responses stream simultaneously
• Side-by-side comparison — Toggle comparison mode to view responses in columns
• Response diffing — Word-level diff highlights agreements vs. differences with percentage
• Per-conversation system prompts — Different personas for different research threads
• Conversation forking — Branch at any message to explore alternative directions

• OpenAI — GPT-5, GPT-4.1, o3, o4-mini
• Anthropic — Claude Opus, Sonnet, Haiku
• Ollama — Auto-discovers locally installed models. No API key needed.
• OpenRouter — 200+ models via a single API key (Gemini, Llama, Mistral, DeepSeek, Qwen, and more)

• Cost tracking — Estimated token counts and USD cost per message, conversation, and session
• Response ratings — Thumbs up/down on any response, tracked per model
• Analytics dashboard — Per-model stats: response count, tokens, cost, latency, approval rate

• Tags & Pins — Tag conversations, pin important ones to the top
• Quick Switcher (Cmd+K) — Fuzzy search across all conversations
• Command Palette (Cmd+P) — Every action in one searchable list
• Full-text search — FTS5-powered search across titles and message content
• Markdown / HTML / JSON export — Copy any conversation to clipboard in multiple formats
• Auto-titling — AI generates meaningful titles after your first exchange
• ChatGPT import — Import your full ChatGPT history from conversations.json

• Image/Vision — Attach images and send them to vision-capable models
• Text file attachment — Native file dialog for code, text, and config files
• Web search — Inject DuckDuckGo search context into your prompt

• OS keychain — API keys stored in macOS Keychain / Linux Secret Service / Windows Credential Manager
• No telemetry — Zero data collection, all conversations stored locally in SQLite

All keybindings are configurable in Settings.

• Rust (2024 edition) — rustup.rs
• GPU-capable system — iced uses wgpu for rendering
• Ollama (optional) — ollama.com for local models

git clone https://github.com/unbalancedparentheses/stoa
cd stoa
cargo run --release

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull a model
ollama pull llama3.2

# Stoa auto-discovers local models on startup

Type a message, press Enter. Goes to the model shown in the picker chip.

1. Click the model chip to open the picker
2. Check multiple models
3. Type your prompt, click Run N (or Cmd+Shift+Enter for all)
4. All responses stream in simultaneously

1. Send to multiple models
2. Click Compare in the chat header
3. Responses appear side-by-side in columns
4. Click Show Diff to see word-level differences

• Set a per-conversation system prompt via the Sys button
• Fork at any message to branch the conversation
• Rate responses with thumbs up/down to track model quality
• View aggregated stats in the Analytics tab
• Tag conversations for organization, pin important ones

Config path: ~/.config/stoa/config.json Data path: ~/.config/stoa/chat.db

Enable RUST_LOG=info (or debug) to see structured log output.

• Response diffing — Word-level LCS diff with agreement percentage. No competitor has this.
• Command palette + Quick switcher — Obsidian-style keyboard-driven navigation.
• Native Rust/GPU — Not Electron, not Tauri, not web. Pure Rust + wgpu rendering.
• OS keychain — API keys in your system's secure credential store, not plaintext config files.

src/
  main.rs              Entry point
  lib.rs               Library root
  app.rs               State, Message enum, update loop, view
  model.rs             Conversation, ChatMessage, Provider
  config.rs            AppConfig, model routing, keychain integration
  db.rs                SQLite + FTS5 persistence, migrations, search
  cost.rs              Pricing table, token estimation
  diff.rs              Word-level LCS diff
  export.rs            Markdown / HTML / JSON export
  import.rs            ChatGPT import parser
  shortcuts.rs         Shortcut specs + key matching
  commands.rs          Command palette entries
  web_search.rs        DuckDuckGo search integration
  theme.rs             Color palette
  handlers/
    streaming.rs       Stream lifecycle, auto-titling
    send.rs            Message sending, retry, review, analyze
  api/
    mod.rs             LlmEvent, stream dispatch, shared HTTP client
    openai.rs          OpenAI + Ollama + OpenRouter streaming
    anthropic.rs       Anthropic streaming
    ollama.rs          Model discovery via /api/tags
  ui/
    chat_view.rs       Messages, comparison mode, diff panel
    input_bar.rs       Input, model picker, file/image attach
    sidebar.rs         Nav, search, history, tags, pins
    right_panel.rs     Streams, system info, cost, shortcuts
    analytics.rs       Per-model stats dashboard
    settings.rs        Provider config, keybindings
    quick_switcher.rs  Cmd+K overlay
    command_palette.rs Cmd+P overlay
    shortcut_help.rs   Shortcut cheat sheet
    markdown.rs        Markdown renderer (pulldown-cmark)
    diagnostics.rs     Debug diagnostics view
    bottom_bar.rs      Status bar

• MCP tool support
• RAG / Knowledge base
• PDF/DOCX file attachment
• Conversation folders
• Plugin system
• Voice input/output
• Cross-platform testing (Windows, Linux)

MIT