Janito — AI Coding Agent

Janito is an AI Coding Agent with tools, MCP, and skills.

📖 Full documentation available at https://joaompinto.github.io/janito/

⚠️ Disclaimer: The code on this repo has been developed mostly using AI.

Features

• 🔧 Function Calling - Built-in tools for file operations, web search, and more
• 🔌 Plugins - Gmail, code search, OneDrive, and more via the plugin system
• 🔌 MCP Support - Connect to Model Context Protocol servers
• 🧩 Skills - Install and use task-specific skills from GitHub
• 🌐 Web UI (Alpha) - Chat through a browser instead of the terminal with --web
• 🔌 Agent Client Protocol (ACP) - Run as an ACP v1 subprocess agent for editors like Zed with --acp
• ↪️ Session Resume - Exit a terminal chat and reopen it with janito -C (full conversation restored)
• 📊 Real-time Progress - Watch tool execution progress as it happens
• 🚀 Easy Setup - Interactive configuration with --config or quick setup with --set flags
• 🔗 OpenAI-Compatible & Native APIs - Works with any OpenAI-compatible endpoint (OpenAI, LM Studio, Ollama, custom) and native SDKs (Anthropic, DashScope)

NOTE: Janito should be used on development systems on which data loss is tolerable and there are no critical secrets

Quick Start

# Install
pip install janito

# Or, with uv (recommended)
uv tool install janito

# Configure interactively
janito --config

# Or set options directly (two steps: config, then API key or ChatGPT login)
janito --set provider=openai --set model=gpt-6.1-sol
janito --set-api-key="sk-your-key" --provider openai
# ...or sign in with your ChatGPT plan instead of an API key:
# janito --provider openai --login

# Start chatting
janito "Hello!"

Installation

From PyPI

pip install janito

Or, with uv:

uv tool install janito

For development setup, see README_DEV.md.

Configuration

Interactive Setup

janito --config

You’ll be prompted for:

• Provider - openai or custom
• API Key - Masked for security
• Model - e.g., gpt-6.1-sol
• Max Output Tokens - Maximum output tokens (default: 65536)
• Max Input Tokens - Maximum input tokens / context window (default: 128000)

Quick Configuration with --set

Set options directly from the command line:

# Single key-value
janito --set model=gpt-6.1-sol

You can also use --get, --unset, and --set-secret with multiple values.

View Configuration

janito --show-config

Costs per provider

For custom endpoints (base-url), see README_custom.md.

Usage

Single Prompt

janito "What is the capital of France?"

Pipe Input

echo "Tell me a joke" | janito

Interactive Chat

janito

Commands in chat mode:

• /exit - End the session
• clear - Clear conversation and start a new one
• Ctrl+D / Ctrl+Z then Enter - Exit (EOF)
• /help - List all commands (with descriptions) and shortcuts

Web UI (Alpha)

⚠️ Alpha — The --web mode is currently in alpha. It works, but the interface, CLI flags, and API may change between releases.

Janito can serve a browser-based chat interface instead of the terminal. This requires optional dependencies that are not part of the core install:

# Install with the web extra
pip install janito[web]
# or
uv tool install janito[web]

Then start the server (opens your browser automatically):

janito --web

# All normal Janito flags still apply, plus web-specific ones:
janito --web -r -w --web-port 9090    # privileges + web, custom port
janito --web --no-web-open                      # don't auto-open the browser

The server binds to 127.0.0.1 by default (localhost only). For full details (features, security, API endpoints, architecture), see the Web UI documentation.

System Prompt Options

Control how the system prompt is handled:

# Use default system prompt (with tools enabled)
janito "What can you do?"

# Disable system prompt entirely (no tools at all)
janito -Z "Simple prompt without system context"

# Override with a custom system prompt (tools stay enabled)
janito -S "You are a concise coding assistant" "Explain recursion"

# Keep the system prompt but skip tool loading (skill tools stay enabled)
janito --no-tools "Explain recursion"

# Skip only the parallel-tasks toolset (StartTask/StopTask/WaitForTask/ListTasks)
janito --no-tasks "Refactor this file"

Note: When using -Z, built-in tools (file operations, MCP) are disabled. Use the default mode or -S flag when you need tool access.

Note: --no-tools disables loading of all non-skill tools (file operations, MCP) while keeping the skill tools (load_skill, read_skill_resource) available, so the model can still load installed skills on demand.

Note: --no-tasks disables only the tasks toolset (the StartTask / StopTask / WaitForTask / ListTasks tools that spawn parallel child processes). File, system, net and skill tools stay available.

Logging

janito --log=info "Your prompt"      # Info level
janito --log=debug "Your prompt"     # Debug level
janito --log=info,debug "Your prompt" # Multiple levels

Examples

OpenAI

# Step 1: Set provider and model
janito --set provider=openai --set model=gpt-6.1-sol
# Step 2a: Store API key (pay as you go)
janito --set-api-key="sk-your-key" --provider openai
# ...or step 2b: sign in with ChatGPT (OAuth, no API key)
# janito --provider openai --login

# Then run any prompt
janito "Explain quantum computing"

Alibaba (Qwen)

# Step 1: Set provider and model
janito --set provider=alibaba --set model=qwen3.8-flash
# Step 2: Store API key
janito --set-api-key="your-dashscope-api-key" --provider alibaba

# Then run any prompt
janito "Explain quantum computing"

Custom Endpoint

janito --set provider=custom --set base-url=http://localhost:8000/v1

Built-in Tools

janito includes tools for common tasks:

File Operations

# List files
python -m janito.tools.files.list_files . --recursive --pattern "*.py"

# Read file
python -m janito.tools.files.read_file README.md --max-lines 20

MCP Tools

Connect to MCP servers using the /mcp command inside the interactive shell:

# Add a stdio-based MCP server
/mcp add myserver stdio python -m mcp.server

# Add an HTTP-based MCP server
/mcp add remote http https://api.example.com/mcp

# List configured servers
/mcp list

For full MCP documentation, see README_MCP.md.

Tool Progress Reporting

Tools report progress in real-time:

🔄 Reading files...
📊 Processing: 50/100 files
✅ Completed: 100 files (2.3MB)

Progress messages go to stderr so they don’t interfere with tool output.

Error Handling

Dependencies

• Python 3.10+
• openai>=1.0.0
• rich>=10.0.0
• prompt-toolkit>=3.0.0
• requests>=2.28.0 (for MCP support)
• pathspec>=0.11.0 (for .gitignore-aware file listing)
• questionary>=2.1.1 (for the interactive --config wizard)

Optional extras: janito[web] adds fastapi/uvicorn (for --web); the native API types need anthropic (--api-type Anthropic), dashscope (DashScope) or google-genai (Gemini).

License

MIT License