whiskr

whiskr is a private, self-hosted web chat interface for interacting with AI models via OpenRouter.

Features

• Private and self-hosted (data stored in localStorage)
• Supports any model available on your OpenRouter account
• Real-time streaming responses
• Edit, delete, or copy any message
• Persistent settings for model, temperature, and other parameters
• Full conversation control including clearing and modifying messages
• Title generation (and refresh)
• Smooth UI updates with morphdom, selections, images, and other state are preserved during updates
• Easy model selection:
  • Tags indicate if a model supports tools, vision, or reasoning
  • Search field with fuzzy matching to quickly find models
  • Models are listed newest -> oldest
• Web search tools (set the tokens.exa to enable):
  • search_web: search via Exa in auto mode; returns up to 10 results with short summaries
  • fetch_contents: fetch page contents for one or more URLs via Exa /contents
  • github_repository: get a quick overview of a GitHub repository (repo info, up to 20 branches, top-level files/dirs, and the README) without cloning (optionally set tokens.github for higher rate limits and private repos)
• Images attachments for vision models using simple markdown image tags
• Text/Code file attachments
• Reasoning effort control
• Structured JSON output
• Statistics for messages (provider, ttft, tps, token count and cost)
• Import and export of chats as JSON files
• Authentication (optional)

TODO

• improved custom prompts
• settings
  • auto-retry on edit
  • ctrl+enter vs enter for sending
• multiple chats

Built With

Frontend

• Vanilla JavaScript and CSS
• morphdom for DOM diffing without losing state
• marked for Markdown rendering
• highlight.js for syntax highlighting
• Fonts: Inter (UI), Comic Code (code)
• Icons: SVGRepo
• Color palette: Catppuccin Macchiato

Backend

• Go
• chi/v5 for the http routing/server
• OpenRouter for model list and completions
• Exa for web search and content retrieval (/search, /contents)

Getting Started

1. Copy example.config.yml to config.yml and set tokens.openrouter:

cp example.config.yml config.yml

2. Build and run:

go build -o whiskr
./whiskr

3. Open http://localhost:3443 in your browser.

Authentication (optional)

whiskr supports simple, stateless authentication. If enabled, users must log in with a username and password before accessing the chat. Passwords are hashed using bcrypt (12 rounds). If authentication.enabled is set to false, whiskr will not prompt for authentication at all.

authentication:
  enabled: true
  users:
    - username: laura
      password: "$2a$12$cIvFwVDqzn18wyk37l4b2OA0UyjLYP1GdRIMYbNqvm1uPlQjC/j6e"
    - username: admin
      password: "$2a$12$mhImN70h05wnqPxWTci8I.RzomQt9vyLrjWN9ilaV1.GIghcGq.Iy"

After a successful login, whiskr issues a signed (HMAC-SHA256) token, using the server secret (tokens.secret in config.yml). This is stored as a cookie and re-used for future authentications.

Nginx (optional)

When running behind a reverse proxy like nginx, you can have the proxy serve static files.

server {
    listen 443 ssl;
    server_name chat.example.com;
    http2 on;

    root /path/to/whiskr/static;

    location / {
        index index.html index.htm;

        etag on;
        add_header Cache-Control "public, max-age=2592000, must-revalidate";
        expires 30d;
    }

    location ~ ^/- {
        proxy_pass       http://127.0.0.1:3443;
        proxy_set_header X-Forwarded-For $remote_addr;
        proxy_set_header Host            $host;
    }

    ssl_certificate /path/to/cert.pem;
    ssl_certificate_key /path/to/key.pem;
}

Usage

• Send a message with Ctrl+Enter or the send button
• Just add a message with the add button
• Hover over a message to edit, delete, or copy it
• Adjust model, temperature, prompt, or message role from the controls in the bottom-left
• Use the model search field to quickly find models (supports fuzzy matching)
• Look for tags in the model list to see if a model supports tools, vision, or reasoning
• Use ![alt](url) in your message to display an image inline. If the model supports vision, the same image URL is passed to the model for multimodal input.

License

GPL-3.0 see LICENSE for details.