whai 0.12.1

A lightweight, Python-based CLI tool that integrates LLMs directly into your terminal

Whai - A Terminal tool to stay in the flow

Whai is a simple way to call an LLM from terminal. It is not an agent that wipes your hard drive while you are in the kitchen.

I built this to lend me a hand when I get stuck or confused when ssh'd into my NAS. The goal is to keep you in your flow and save you from an alt tab to a browser and copy pasting context to get help.

https://github.com/user-attachments/assets/cbe834f0-2437-405b-9c95-88f02f6f69d9

Table of Contents

• What is it
• Core Features
• Quick Examples
• Installation
• Quick Start
• Supported Providers
• MCP (Model Context Protocol) Support
• Key Features
• FAQ
• Contributing
• Acknowledgments

What is it

whai is a lightweight AI terminal assistant that integrates directly into your native shell and workflow. The philosophy of whai is to supplement your normal terminal usage without taking control. You don't have to start a chat loop, alt-tab or copy paste. When you get stuck, need a command, or encounter an error, just call whai for help on demand.

Core Features

• Analyze Previous Errors: If a command fails, you can call whai (no arguments needed!) or ask whai why did that fail?. It reads the failed command and its error output from your terminal history to provide a diagnosis and solution. Note: Command output is available when running inside tmux or a whai shell session. Otherwise, the model will only see your commands but not their outputs.

• Persistent Roles (Memory): whai uses simple, file-based "Roles" to provide persistent memory. You define your context once, what machine you are on, what tools are available, your personal preferences, and how you like to work, and whai retains this context for all future interactions.

• Session Context: When running inside tmux or a whai shell session, whai reads your command history and outputs to understand both what you ran and what happened.

• Help On Demand: Get help as you're using your shell:

  > whai check my docker containers logs for errors

  > whai "Is this resource usage normal?"

• Requires Approval: Every whai command requires your explicit [a]pprove / [r]eject confirmation.

• MCP Tool Integration: Connect local MCP servers to extend whai with additional tools like file operations, database queries, or API integrations, all with the same approval workflow.

• Model-Agnostic: Use models from OpenAI, Gemini, Mistral, Anthropic, local Ollama models, and more.

• Insert-Command Mode (Optional): Turn natural language at your prompt into a single shell command with a keybinding that replaces your current line without auto-executing anything.

Quick Examples

Example 1: Running a command

$ whai give me the biggest folder here
Info: Model: gpt-5-mini | Role: default
I'll list all top-level directories (including hidden ones), 
compute their disk usage, sort them, and show the largest 
few with human-readable sizes.

I'll run this command:
find . -maxdepth 1 -mindepth 1 -type d -print0 | \
  xargs -0 du -sk 2>/dev/null | \
  sort -n | tail -n 5 | \
  awk '{print $1 "\t" $2}' | \
  numfmt --to=iec-i --suffix=B --field=1,1

Breakdown of what each part does:
- find . -maxdepth 1 -mindepth 1 -type d: Find directories at 
  current level only (excludes . and ..)
- -print0: Use null separator to safely handle spaces in 
  directory names
- xargs -0 du -sk: Pass each directory to du to get disk usage 
  in kilobytes, 2>/dev/null suppresses errors
- sort -n: Sort numerically by size
- tail -n 5: Show the 5 largest entries
- awk '{print $1 "\t" $2}': Format output as size (tab) path
- numfmt --to=iec-i --suffix=B --field=1,1: Convert sizes to 
  human-readable format (KiB, MiB, etc.)

╭─────────────────────── Proposed command ────────────────────────╮
│ find . -maxdepth 1 -mindepth 1 -type d -print0 |                │
│   xargs -0 du -sk 2>/dev/null |                                 │
│   sort -n | tail -n 5 |                                         │
│   awk '{print $1 "\t" $2}' |                                    │
│   numfmt --to=iec-i --suffix=B --field=1,1                      │
╰─────────────────────────────────────────────────────────────────╯
[a]pprove / [r]eject / [m]odify: a

╭─────────────────── Output ────────────────────╮
│ 440B ./tests                                  │
│ 1.4KiB ./.git                                 │
│ 87KiB ./.venv_tmp                             │
│ 108KiB ./.venv                                │
│ 123KiB ./.venv-linux                          │
╰───────────────────────────────────────────────╯

Result (interpretation)
- The largest top-level directory here is ./.venv-linux 
  with ~123KiB.

Next steps (options)
- Show the top N largest items anywhere under the current 
  tree (files+dirs):
  - du -ah . 2>/dev/null | sort -rh | head -n 30
- Show only directories, recursively, largest first:
  - du -h --max-depth=3 2>/dev/null | sort -hr | head -n 30

Example 2: Troubleshooting terminal output

$ uv pip install requirements.txt
Using Python 3.11.9 environment at: .venv-linux
  × No solution found when resolving dependencies:
  ╰─▶ Because requirements-txt was not found in the package registry...

$ whai what is the problem
Info: Model: gpt-5-mini | Role: default
What's happening
- You ran: uv pip install requirements.txt
- pip treated "requirements.txt" as the name of a package to 
  download from the package index, not as a file of requirements.
- The resolver tried to find a package called (normalized) 
  "requirements-txt" on PyPI and failed, so it reports the 
  dependencies are unsatisfiable.

How to fix it
- Use the -r (or --requirement) flag so pip reads the file 
  instead of treating it as a package name.

Commands you can run
1) Check the requirements file exists:
ls -l requirements.txt

2) Install from the file (match how you invoked pip):
uv pip install -r requirements.txt

Do you want me to:
- run the ls to confirm the file exists, or
- run the install command for you now?

Example 3: Calling whai without arguments

$ git push origin main
To https://github.com/user/repo.git
 ! [rejected]        main -> main (non-fast-forward)
error: failed to push some refs to 'https://github.com/user/repo.git'
hint: Updates were rejected because the tip of your current branch is behind
hint: its remote counterpart. Integrate the remote changes (e.g.
hint: 'git pull ...') before pushing again.

$ whai
Info: Model: gpt-5-mini | Role: default
What's happening
- You tried to push your local branch "main" to the remote, 
  but git rejected it.
- Your local branch is behind the remote branch—someone else 
  pushed changes to the remote that you don't have locally.
- Git won't let you push because it would overwrite those 
  remote changes.

How to fix it
- Pull the remote changes first, then push. This integrates 
  the remote commits with your local ones.

Commands you can run
1) See what commits are on the remote but not locally:
git fetch
git log HEAD..origin/main

2) Pull and merge the remote changes:
git pull origin main

3) After resolving any merge conflicts (if they occur), push again:
git push origin main

Do you want me to run git pull for you now?

Example 4: Asking questions

$ whai "Quick how do I exit Vim??"
Info: Model: gpt-5-mini | Role: default
I'll show the quick ways to exit Vim and what each one does.

1) Get to Normal mode
- Press Esc (or Ctrl-[) until you stop inserting text. If Esc doesn't respond, try Ctrl-C.

2) Common exit commands (type them in Normal mode, then press Enter)
- :q
  - Quit (works only if no changes since last save).
- :q!
  - Quit and discard any unsaved changes.
- :wq
  - Write (save) and quit.
- :x
  - Write (only if changed) and quit (similar to :wq).

3) Shortcuts
- ZZ (Shift+Z twice) — save and quit (same as :wq if there are changes).
- ZQ — quit without saving (same as :q!).

Installation

Option 1: uv (Recommended)

You can install uv here

Then:

uv tool install whai

Or even without installing it!

uvx whai "your command"

Option 2: pipx

pipx install whai

Option 3: pip

pip install whai

Option 4: From source

git clone https://github.com/gael-vanderlee/whai.git
cd whai
pip install -e .

Quick Start

1. Configure your API key

On first run, whai launches an interactive configuration wizard:

whai --interactive-config

Or edit ~/.config/whai/config.toml directly:

[llm]
default_provider = "openai"

[llm.openai]
api_key = "sk-proj-your-key-here"
default_model = "gpt-5-mini"

Make sure to use a model that is capable of calling tools. Most frontier models have this functionality. For local models, you can look at the LMStudio models tagged for tool use here. Get API keys from:

• OpenAI Platform
• Anthropic Console
• Azure Portal (for Azure OpenAI)
• Google AI Studio (for Gemini)

Using Local Models with LM Studio

To use a local model with LM Studio:

1. Enable the server in LM Studio:

   • Open LM Studio
   • Go to the Developer menu
   • Enable the server toggle

2. Configure whai:

   whai --interactive-config
   

   • Select lm_studio as the provider
   • Leave the API key blank (not required for local models)
   • Enter the API base URL: http://localhost:1234/v1
   • Enter the model name without prefix (e.g., llama-3-8b-instruct)

   Note: Model names are stored without provider prefixes in the config file. Prefixes are automatically added at runtime when needed.

Using OpenAI-Compatible Local Endpoints

Use openai_api for llama.cpp, llama-swap, vLLM, Tabby, and any other server that exposes an OpenAI-compatible /v1 API.

1. Start your local server and confirm its OpenAI API base URL.

2. Configure whai:

   whai --interactive-config
   

   • Select openai_api as the provider
   • Leave the API key blank if the server does not require one
   • Enter the API base URL for your server
   • Enter the model name without provider prefix (for example: llama3, qwen2.5, or your server’s exposed model id)

   Note: whai stores model names without provider prefixes and automatically adds the LiteLLM routing prefix at runtime.

2. Start using whai

Getting Help: For a complete list of command-line options and flags, run whai --help.

whai "your question"

That's it! whai will:

• Read your terminal context (commands + output if in tmux or whai shell, commands only otherwise)
• Send your question to the configured LLM
• Suggest commands with [a]pprove / [r]eject / [m]odify prompts
• Execute approved commands and continue the conversation

Tip: Quotes are not necessary, but do use them if you use special characters like ' or ?

whai show me the biggest file here
whai "what's the biggest file?"

Supported Providers

whai supports the following LLM providers:

• OpenAI - GPT models (e.g., gpt-5-mini, gpt-4, gpt-4-turbo)
• Anthropic - Claude models (e.g., claude-3-5-sonnet-20241022, claude-3-opus)
• Google Gemini - Gemini models (e.g., gemini-2.5-flash, gemini-2.5-pro)
• Mistral AI - Mistral models (e.g., mistral-small-latest, mistral-medium-latest, mistral-large-latest)
• Azure OpenAI - Azure-hosted OpenAI models
• Ollama - Running Ollama instance
• LM Studio - Local models via LM Studio
• OpenAI-compatible local endpoints - llama.cpp, llama-swap, vLLM, Tabby, and similar servers

Each provider must be configured in your ~/.config/whai/config.toml file. You can configure multiple providers and switch between them using roles or the --provider flag.

Note: whai uses LiteLLM for multi-provider support. Additional providers from LiteLLM's supported providers list can be added upon request.

Key Features

Roles

Roles allow you to customize whai's behavior and responses, and let you save information you don't have to repeat yourself in every conversation.

• Your system information (OS, available tools, paths)
• Your preferences (shell style, preferred commands, workflows)
• Environment constraints (what you can/can't do, security policies)
• Project-specific context (tools in use, conventions, setup)

Let's create a toy role that tells whai to respond only in emoji:

$ whai role create emoji # "Answer using only emojis"
$ whai can you tell me the plot of the first Shrek movie --role emoji
Info: Model: gpt-5-mini | Role: emoji
👑👸💤🐉🏰
👹🏞️🕳️➡️🏰🐴😂
⚔️🐉🔥💨👸
👹❤️👸💚
🌅💋✨💚💚
🎉🎶🧅

# Create a new role
whai role create my-workflow

# Use it
whai "help me with this task" -r my-workflow

# List all roles
whai role list

For a complete list of role management commands, run whai role --help. Roles are stored in ~/.config/whai/roles/ as Markdown files with YAML frontmatter, like so:

---
provider: openai
model: gpt-5-mini
# Optional parameters you can add:
# temperature: 0.3                # Only used when supported by the selected model
---
You are a helpful terminal assistant.
Describe context, behaviors, tone, and constraints here.

Available Providers:

You can specify any provider in the provider field. See the Supported Providers section for the list of available providers.

The provider must be configured in your ~/.config/whai/config.toml file before it can be used. If no provider is specified in the role, whai uses the default provider from your configuration.

The default role is defined in the config.

Context Awareness

whai automatically captures context from:

• tmux scrollback (recommended): Full commands + output context
• Recorded shell sessions: Full commands + output when using whai shell
• Shell history (fallback): Recent commands only when not in tmux

MCP (Model Context Protocol) Support

whai supports connecting to local MCP servers to extend functionality with additional tools. MCP servers can provide tools for file operations, database queries, API integrations, and more.

Setting Up MCP Servers

1. Create ~/.config/whai/mcp.json with your MCP server configuration:

{
  "mcpServers": {
    "server-name": {
      "command": "command-to-run",
      "args": ["arg1", "arg2"],
      "env": {"KEY": "value"},
      "name": "Display Name (optional)",
      "requires_approval": true
    }
  }
}

2. Example configuration for the time server:

{
  "mcpServers": {
    "time-server": {
      "command": "uvx",
      "args": ["mcp-server-time"],
      "env": {},
      "name": "Time Server",
      "requires_approval": false
    }
  }
}

Configuration fields:

• command (required): Command to run the MCP server
• args (optional): Arguments to pass to the command
• env (optional): Environment variables for the server
• name (optional): Display name shown in tool approval prompts
• requires_approval (optional, default: true): Whether to prompt for approval before executing tools from this server

MCP support is opt-in: if mcp.json doesn't exist, MCP is disabled. Tools from MCP servers are automatically discovered and made available to the LLM alongside the built-in execute_shell tool.

To disable MCP for a single run, use the --no-mcp flag:

whai "my question" --no-mcp

To disable MCP persistently, change in config.toml:

[mcp]
enabled = false

Targeting another pane (tmux)

In a tmux session with multiple panes, you can run whai in one pane and have it use context from, and run approved commands in, another pane (e.g. one pane SSH'd to a server). Use --target or -T with the pane number (see pane numbers with Ctrl+b q):

whai -T 1 "check disk space"

Set WHAI_TARGET=1 in your environment to use a default target pane so you can omit the flag.

Recorded Shell Sessions

For output context without tmux, use whai shell to launch an interactive shell with session recording:

whai shell

This command:

• Opens your normal shell (bash, zsh, fish, or PowerShell) with identical behavior
• Records all commands and outputs to a session log
• Provides commands + outputs to whai without requiring tmux

The recorded session behaves exactly like your normal shell. The only difference is that whai can now access full command outputs for better assistance.

To exit: Type exit in the shell to return to your previous terminal.

Options:

# Launch with a specific shell
whai shell --shell zsh

# Specify a custom log path
whai shell --log ~/my-session.log

Session logs are stored temporarily during the session and are deleted when you exit the shell. When you run whai from within a recorded shell session, it automatically uses the in-session log.

3. Insert-command keybinding

whai can also be used as a “turn this line into a command” helper.

• Interactive keybinding (bash/zsh):
  • In the interactive config wizard (whai --interactive-config) offers to add a Ctrl+G (can be changed in the snipped) keybinding snippet to your shell config file (e.g. ~/.bashrc or ~/.zshrc).

After enabling it:

• Type a natural-language request on the command line, e.g. list largest folders here.
• Press Ctrl+G.
• Your line is replaced with a real shell command proposed by whai --command-only.
• You can edit the command and press Enter when you are ready to run it.

Safety First

• Every command requires explicit approval
• Modify commands before execution
• Commands run in isolated subprocess (won't affect your main shell)
• Press Ctrl+C to interrupt anytime

FAQ

How is this different from [insert app here] ?

whai is integrated into your terminal with context awareness. It sees your command history and can execute commands. Most terminal assistants either require you to explicitely start a chat loop which takes you out of your usual workflow, don't allow for roles, or don't allow to mix natural language conversation and shell execution. You don't need AI to cd into a directory, you keep full agency over simple actions. The goal was really to replace alt-tabbing to google with a CLI command.

Does it send my terminal history to the LLM?

Only when you run whai. It captures recent history (50 last commands), tmux scrollback (commands + output), or recorded shell session content (commands + output while the session is active) and includes it in the request. If you use a remote API model, it will see your recent terminal history. You can disable this with the --no-context flag.

Can I use it with local models?

Yes! Configure Ollama or LMStudio for local models. See the configuration section above.

Contributing

Contributions are welcome! If you'd like to contribute to whai, please:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Make your changes
4. Run the test suite (uv run pytest)
5. Commit your changes (git commit -m 'Add some amazing feature')
6. Push to the branch (git push origin feature/amazing-feature)
7. Open a Pull Request

Every Pull Request triggers the Pull Request CI workflow, which runs the full nox test matrix across Python 3.10–3.14 plus ruff linting. Keep your local uv run pytest runs green to avoid CI surprises.

For detailed development setup and guidelines, see DEV.md.

Acknowledgments

Built with LiteLLM for multi-provider support, Typer for the CLI, and Rich for pretty terminal output.