Skip to main content

MahanAI: terminal AI agent for NVIDIA NIM, Claude, Codex, tools, streaming, /api-key, safety prompts.

Project description

(M MahanAI)

MahanAI

A terminal AI agent with a plugin ecosystem, gateway server, and full multi-model support.

PyPI version Python License: MIT Downloads

📖 Docs · 🐛 Issues · 🔌 Plugin Store


preview

What is MahanAI?

MahanAI is a terminal AI agent you install with pip. It gives you a powerful chat interface, agentic tool use (run commands, read/write files, search the web), and a local gateway server that makes all your configured AI providers available as a single unified API endpoint — so any tool that speaks OpenAI or Anthropic format can point at it.

On top of that: a plugin system (.mmd files) with a GitHub-backed store, custom themes (.mai files), conversation branching, effort levels, plan mode, and 60+ slash commands.

New in Max 3.0

  • Modern terminal composer — a three-line input bar, cyan prompt marker, and a clean model/effort/workspace status line
  • Clipboard image input — paste multiple images directly into the prompt; image tokens can be removed with one Backspace
  • Compact GitHub repository tokens — pasted repository references render as ◉ owner/repo, remain atomic while editing, and send the original URL to the model
  • Persistent default model/set-def-model opens the model selector and saves the choice for future sessions
  • Expanded OpenAI models — GPT-5.5 plus GPT-5.6 Sol, Terra, and Luna in Direct and Indirect modes
  • Improved Codex login — successful browser authentication opens the bundled MahanAI confirmation page

Install

pip install mahanai
mahanai

That's it. An onboarding wizard runs on first launch to help you pick a model, set up API keys, and connect GitHub for the plugin marketplace.

If you want a guided bootstrap that installs uv, installs Python through uv, installs mahanai, and launches it, use one of the included get-started scripts:

./scripts/get-started.sh
powershell -ExecutionPolicy Bypass -File .\scripts\get-started.ps1

If you use PowerShell Core, swap powershell for pwsh.

On macOS, you can also double-click scripts/get-started.command from Finder.


Highlights

  • Gateway server — expose all your providers behind one OpenAI/Anthropic-compatible endpoint; works with Cursor, Continue, LM Studio, Claude Code, and more
  • Plugin store — install community .mmd plugins with /store install, publish your own with /store upload
  • Custom themes — write .mai theme files with gradients, color aliases, and display name overrides
  • Multi-model — Claude (via Claude Code), NVIDIA NIM, OpenAI Codex, Ollama, and any OpenAI-compatible endpoint
  • Multimodal prompts — paste multiple clipboard images alongside text in the terminal composer
  • Repository-aware input — GitHub links collapse into editable ◉ owner/repo tokens without changing what the model receives
  • Agentic tools — shell commands, file read/write/edit, web search, Python REPL, URL fetch — all with approval prompts and inline diffs
  • Interact mode — enable once with /interact to let MahanAI control your computer with screenshots, mouse, and keyboard; Wayland helpers are used first, with X11 as a fallback; remove it later with /interact remove
  • Plan mode & effort levels/plan on to outline before acting; /effort high for deeper reasoning
  • Conversation branching — save and restore conversation states to explore multiple paths
  • Searchable chat history — search saved chats by keyword with /history search <query>
  • Shell history awareness — inject your recent bash/zsh history into context with /shell-history inject
  • Cost tracking/cost shows session token usage and estimated spend
  • Desktop notifications — get pinged when a long generation finishes

Gateway Server

Start a local HTTP server that routes requests to the right backend automatically:

mahanai --server                          # OpenAI-compatible on port 8080
mahanai --server --type anthropic         # Anthropic-compatible
mahanai --server --port 9000 --api-key sk-gaming

Point any OpenAI-compatible client at http://localhost:8080 and use model IDs like claude-sonnet-4-6, meta/llama-3.3-70b-instruct, gpt-5.6-sol-indirect, gpt-5.6-terra-indirect, gpt-5.6-luna-indirect, or your Ollama model name — MahanAI routes and converts formats automatically, including SSE streaming end-to-end.

A browser-based chat UI is also available at http://localhost:8080 when the server is running.


Models

Switch models interactively with /models or quick-switch with /mode claude / /mode default.

Provider How to connect
Claude (Opus, Sonnet, Haiku) Install Claude Code and sign in
NVIDIA NIM (Llama 3.3 70B) /api-key-nvidia your-key
OpenAI Codex (GPT-5.5; GPT-5.6 Sol, Terra, Luna; earlier GPT-5 models) /codex-login (browser OAuth, no API key needed)
GitHub plugin store /store login (OAuth when configured, PAT fallback otherwise)
Ollama /add-ollama name localhost 11434
Any OpenAI-compatible API /custom http://your-server/v1 model-name

Default model on first launch: Claude Haiku 4.5


Plugins

Plugins are .mmd files that register new slash commands. They can delegate to Claude Code, MahanAI itself, or the shell, and they can now also embed default .mai themes or launch Tk windows from a trigger.

# Install from the store
/store install mahancreate/maifoundation.example.mahmod

# Load a local plugin
/plugin-load path/to/my-plugin.mmd

# Browse the store
/store browse
/store search compact

Publishing your own plugin is one command:

/store upload path/to/my-plugin.mmd

This creates a public GitHub repo, pushes the .mmd file, and tags it so it shows up in /store browse.

GitHub auth for the store works in two modes:

  • Set MAHANAI_GITHUB_CLIENT_ID and run /store login for browser/device-flow OAuth
  • Leave it unset and /store login falls back to a GitHub personal access token prompt

Plugin Extras

You can embed a default theme directly inside a plugin with newdeftheme(...):

newdeftheme(
theme.name = example-neon
theme.pretty.name = Example Neon

sky = #5B8DEF
mint = #34D399

ascii-art.default.color = gradient("sky -> mint")
message.ai.color        = color("sky")
message.user.color      = color("mint")
)

When the plugin is loaded, that theme is registered into /themes like any other .mai theme.

You can also open a Tk window from a plugin command with pytknwd(...):

add command("/example-window"){
    pytknwd(
root = tk.Tk()
root.title("Hello")
tk.Label(root, text="Window from MMD").pack()
root.mainloop()
    )
}

See example-mahanai-mahmod.mmd for a full example that includes both features.


Themes

Four built-in themes (midnight, light, midnight-cb, light-cb) plus full custom theme support via .mai files:

/themes midnight          # switch theme
/theme-load mytheme.mai   # load a custom theme

A .mai theme file looks like this:

theme.name        = my-theme
theme.pretty.name = My Custom Theme

blue = #5B8DEF
gold = #F5C842

ascii-art.default.color = gradient("blue -> gold")
message.ai.color        = color("blue")
message.user.color      = color("gold")
message.ai.name         = text("assistant")

Project Config (.mahanairc)

Place a .mahanairc in any project directory and MahanAI loads it automatically — pre-loading context files, auto-installing plugins, and activating dev kit extras scoped to that workspace.

load(location="context.md" type=context)
load(location="plugins/my-plugin.mmd" type=mmd)
load(python-dev-kit)

Key Commands

Command Description
/models Interactive model picker
/set-def-model Select and persist the default model for future sessions
/effort <low|medium|high|very-high> Set reasoning depth
/plan on|off Outline approach before every response
/auto on|off Autonomous mode (skip approval prompts)
/interact Enable computer control with screenshots, mouse, and keyboard; Wayland first, X11 fallback
/interact remove Disable computer control and remove the feature
/branch save <name> Snapshot conversation state
/branch load <name> Restore a snapshot
/cost Show session token usage and cost
/memory add <text> Save a persistent memory
/history search <query> Search saved chat sessions
/shell-history inject Add recent shell history to context
/store browse Browse the plugin store
/cmd Fuzzy-search all 60+ commands
/init Generate a MAHANAI.md workspace context file

Full command reference: mahancreate.github.io/mahanai


API Keys

Provider How
NVIDIA NIM /api-key your-key or MAHANAI_API_KEY=...
NVIDIA direct /api-key-nvidia your-key
Claude Handled by Claude Code — no extra config
OpenAI Codex /codex-login (browser OAuth)
GitHub plugin store /store login (OAuth if MAHANAI_GITHUB_CLIENT_ID is set, otherwise PAT)

Keys are stored in ~/.config/mahanai/config.json (Linux/macOS) or %APPDATA%\MahanAI\config.json (Windows).


Environment Variables

Variable Purpose
MAHANAI_API_KEY Override saved server API key
MAHANAI_MODEL Override default model
MAHANAI_STREAM Set to 0 to disable streaming
MAHANAI_CONFIG_DIR Override config directory
MAHANAI_GITHUB_CLIENT_ID Enable GitHub OAuth for /store login
NO_COLOR Disable terminal colors

License

MIT © The MahanAI Foundation

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

mahanai-9.0.1.tar.gz (119.3 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

mahanai-9.0.1-py3-none-any.whl (112.5 kB view details)

Uploaded Python 3

File details

Details for the file mahanai-9.0.1.tar.gz.

File metadata

  • Download URL: mahanai-9.0.1.tar.gz
  • Upload date:
  • Size: 119.3 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.13.13

File hashes

Hashes for mahanai-9.0.1.tar.gz
Algorithm Hash digest
SHA256 ba3dd9c6d23d7d3094e337708142270b3eb5569a6e4cfad262796b9fab116f07
MD5 70eea5b96162eff4aeed8746746f2597
BLAKE2b-256 945a72bfa6a96eaef3b59ed72b2f4d15b8f858ee57620d057738043a5541fec8

See more details on using hashes here.

File details

Details for the file mahanai-9.0.1-py3-none-any.whl.

File metadata

  • Download URL: mahanai-9.0.1-py3-none-any.whl
  • Upload date:
  • Size: 112.5 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.13.13

File hashes

Hashes for mahanai-9.0.1-py3-none-any.whl
Algorithm Hash digest
SHA256 e38c7dfac960bc59de8dfa9b2a5a6c36acb621c97f63cbf272de6a6858820bba
MD5 e0238e4d35c726bb98f552c98091fe09
BLAKE2b-256 2bf3e97a9b76180355b0795a905d2e7736785a41cfec6f2756db540afc89cb54

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page