Last updated: March 21, 2026
Overview
MAUDE is a multi-model AI operating environment built around a central Python gateway. All clients — the server TUI, Mac/PC CLI, mobile app, and Telegram bot — connect through a single gateway that handles model routing, tool execution, and streaming.
Client (TUI / CLI / Mobile / Telegram)
│
▼
Gateway (port 30000 HTTPS / 30080 HTTP)
│
├── Local models (llama-server, port 30010)
├── Cloud models (Mistral API, Anthropic API, OpenRouter)
├── Tool execution (100+ tools)
├── Browser automation (Playwright)
├── Google Workspace (OAuth 2.0)
└── File server (port 30002)The gateway makes non-streaming LLM calls in a loop: send the conversation to the model, check for tool calls, execute tools, feed results back, repeat. The final text response is streamed to the client as SSE with real-time tool trace events.
Models
mistralDefault. Best all-round with tools.
codestralOptimized for code generation.
devstralFrontier code agent (123B, 256K context).
claudeDeep reasoning, long context.
sonnetFast, capable, cost-effective.
nemotronFree. No API key needed.
nemotron-superCloud inference, no cost.
Switch models in the TUI with /model switch <name>. The gateway resolves short names to full model IDs and routes to the correct provider.
Tools
read, write, edit, search, list, change directory
run_command (git, pip, python, system commands)
web_search (DuckDuckGo), web_browse, web_view
view_image (LLaVA analysis of photos, screenshots)
list, read, send emails
list, search, read, upload, create docs/sheets/folders
read, write, append, create
list, create, update, delete events
get, create presentations, add slides/text
list, get, create, update, delete, search
search, get video/channel, playlists, comments
PRs, issues, repos, branches, commits, CI/CD, releases
open, navigate, click, type, fill forms, screenshot, extract
post to X, LinkedIn, Facebook, Instagram (browser-based)
FLUX with LoRA support via ComfyUI
create/edit drafts, publish, list posts, stats
save, recall, list, forget (persistent across sessions)
mesh status, dispatch tasks, project management
cron-based task scheduling with natural language
CPU, GPU, RAM, disk, processes, sessions, activity
Docker container for autonomous builds (exec, read, write)
Autonomous builder: plan, execute, verify, fix loop
ask_frontier (escalate to Claude/Gemini)
Tools are dynamically filtered per message based on keywords. Only relevant tools are sent to the model, reducing token usage by 30–40%.
Reference
/model switch <name> Switch model (mistral, codestral, sonnet, claude, etc.)
/clear Clear conversation history
/copy Copy last response to clipboard
/keys set <provider> <key> Set an API key
/keys show Show configured providers
/help Show all commands
/quit Exit MAUDEReference
300003008030010300028998Reference
variables.env API keys (MISTRAL, CODESTRAL, CLAUDE, OPEN_ROUTER)
~/.config/maude/keys.json API keys (managed via /keys command)
credentials.json Google OAuth 2.0 client credentials
certs/cert.pem SSL certificate (enables HTTPS on gateway)
certs/key.pem SSL private key
~/.config/maude/browser_data/ Persistent browser session data
~/.config/maude/memory.db Persistent memory (SQLite)
data/conversations/ Conversation history (JSON)Reference
./maude Start all services + open TUI
./maude -n Start services in background only
./maude stop Stop all services
./maude status Show what's runningBackground services run in named tmux sessions. The TUI runs in the foreground so terminal copy/paste works normally. Press Ctrl+C to exit the TUI — services keep running.
Knowledge
MAUDE includes 10 markdown-based reference guides that are automatically injected into the system prompt when the user’s message matches relevant keywords. This gives the model domain expertise on demand without permanently bloating the context window — max 2 guides are loaded per turn.
Architecture, code quality, error handling, security, testing, Python patterns
Layout, typography, color, responsive, performance, accessibility, forms
Modals, toasts, loading states, navigation, tables, dark mode, CSS architecture
Composition, color theory, typography, imagery, icons, brand consistency
Color wheel, harmony systems, psychology, 60-30-10 rule, palette building
Clarity, structure, tone, technical writing, editing, grammar
REST conventions, status codes, pagination, auth, versioning, documentation
System prompts, few-shot, chain-of-thought, tool descriptions, image prompts
Prompt structure, composition, lighting, styles, quality modifiers, aspect ratios
Authentication, injection prevention, secrets, network hardening, containers
Plugins
Skills are Python-based plugins registered via a @skill decorator. Each skill is exposed to the LLM as an OpenAI-compatible tool definition with typed parameters and JSON schema. Built-in skills include weather, calculator, stocks, screenshots, image generation, datetime utilities, notes, and system info. User skills can be added to ~/.config/maude/skills/.
/skills List all installed skills
/skills enable <n> Enable a skill
/skills disable <n> Disable a skill
/skills info <n> Show skill details
/skills run <n> ... Run a skill directly
/skills reload Reload all skillsClients
Rich Console chat with prompt_toolkit input, tool traces, thinking spinner. Runs on the server via ./maude.
pip install maude-client. Connects over Tailscale. Braille spinner, trace visualization, conversation sync.
React + Capacitor native app. Voice chat, SSH terminal, file manager, browser, Telegram viewer.
Full tool access via Telegram. Photo analysis, tool execution, cross-channel sync.
Command Center with system monitoring, conversation history, and scheduled task management.