AI Council · Computer Use · Universal MCP · Memory Vault · Voice-Out

Every Feature in
RAPR AI

AI Council debates, computer use automation, multi-AI models, an editable memory vault, Telegram remote control, scheduling, Zapier and Composio MCP connectors, voice-out responses, release signing support, and 200+ skills in one desktop app.

Explore RAPR AI features

Computer Use

Give any AI complete control over your desktop — mouse, keyboard, windows, and apps. Works with Claude, Gemini, Codex, or any connected AI.

  • 16 desktop control tools — screenshot, click, type, scroll, key press, window focus, resize, app launch, semantic control, vision detection, and more
  • Works with any AI — Claude, Gemini, Codex, Ollama — via universal tool injection
  • Tier 1 (CLI path): any AI emits tool calls; RAPR executes and returns results as text
  • Tier 2 (Native API): Claude, Gemini, and Codex use direct vision APIs — screenshot auto-attached after every action, ~5x faster
  • Tier 3 (Semantic / pywinauto): computer_get_window_tree reads the Windows accessibility tree; computer_type_in_control types by control name with no pixel coordinates — sub-second, works in any native Win32 app
  • OmniParser (Bonus tier): computer_parse_screen runs YOLOv8 icon detection and Florence-2 captioning to label every clickable element on screen — lets cheap LLMs drive a UI without vision
  • Claude uses the official computer_20250124 native tool with full pixel-coordinate control
  • Serialization lock prevents overlapping mouse and keyboard actions across parallel sessions
  • DPI-aware on Windows — clicks land correctly on HiDPI (125–150% scaled) displays
  • Safety failsafe — move mouse to screen corner at any time to immediately abort automation
  • Enable with one toggle in Settings — off by default, no restart needed

Demo Guide

A self-contained 3-act recording script for showcasing RAPR AI computer use on camera. Covers pixel control, semantic Windows accessibility, and OmniParser vision detection.

  • Act 1 — "AI as your hands": AI opens Notepad and types via real keyboard injection; PyAutoGUI failsafe live on camera
  • Act 2 — "AI does things a script can't": AI navigates and interacts inside a logged-in browser session with no API or OAuth token
  • Act 3 — "Semantic control, no pixels needed": computer_get_window_tree reads the Windows accessibility tree; computer_type_in_control targets controls by name — sub-second, zero vision round-trip
  • Bonus Act — OmniParser: computer_parse_screen runs YOLOv8 + Florence-2 to label every clickable element; cheap LLMs drive the UI from a labeled bounding-box list
  • Pre-flight checklist: server restart, Computer Use toggle, OBS capture setup, Tier 2 Claude recommended for filming
  • Per-act talking points, on-camera beats, and a 6-minute timing template for final cuts
  • Failure-recovery cheat sheet: DPI drift, UAC mismatch, PyAutoGUI failsafe, address-bar focus, OmniParser missing deps

Multi-AI Integration

Connect 16 AI CLIs and providers reaching hundreds of models, then switch between them in the same workspace. Use the best model for each task.

  • Claude Code, Gemini CLI, and Codex CLI for agentic coding
  • Cursor, Aider, Goose, Kilo Code, and Open Interpreter coding agents
  • Ollama and local OpenAI-compatible servers (LM Studio, vLLM) run fully offline
  • OpenRouter, Groq, GitHub Models, Amazon Q, and Antigravity for hosted models
  • NemoClaw sandboxed NVIDIA agent and OpenClaw coding agent, or add your own
  • Switch AI mid-conversation with a single click
  • Run 4+ sessions in parallel, each with its own AI and project folder

AI Council

Pit multiple AIs against each other on any topic. A moderator AI drives the debate, judges each round, and declares a winner — or calls consensus when key angles are covered.

  • Pick 2+ active AI sessions as participants and one as moderator
  • Moderator AI picks the next speaker and poses a focused question each turn
  • Round ends when all participants speak once; moderator declares round winner with detailed reasoning
  • Cumulative scoring tracks round wins across all participants
  • Overall winner declared after all rounds with a 4–6 sentence final verdict
  • Consensus detection ends the debate early when key angles are covered
  • Each participant streams token-by-token to the UI in real time
  • Inject messages mid-debate to steer arguments or introduce new evidence
  • Stop any council at any time — partial results and scores are preserved

Shared Memory Vault

Start a project with Claude Code today, continue it with Gemini tomorrow, and inspect what RAPR AI remembers in a browsable vault.

  • 6 memory categories — facts, preferences, instructions, context, goals, people
  • One shared memory store for Claude, Gemini, Ollama, Codex, and other connected AIs
  • Auto memory extraction at session end and graceful shutdown
  • Decay scoring — old memories fade, pinned ones stay forever
  • Pin important memories so the right context stays available
  • Search, browse, edit, archive, and manage memories from the sidebar
  • Export an Obsidian-compatible Markdown vault with YAML frontmatter and an index of wikilinks
  • Edit memories in place from the web UI when the assistant learned something wrong

Voice-Out Responses

New

Listen to assistant replies from the web UI with offline text-to-speech by default and ElevenLabs when configured.

  • Speaker button on assistant messages with play and stop behavior
  • Optional auto-speak setting saved in local storage
  • Offline pyttsx3 backend keeps voice-out available without API keys
  • ElevenLabs backend activates automatically when ELEVENLABS_API_KEY is set
  • Code blocks are stripped before speech so responses remain listenable
  • Voice ID validation protects the ElevenLabs URL path before requests are made

Telegram Remote Control

Control RAPR AI from your phone. Send prompts, trigger sessions, and receive responses — all from Telegram.

  • Send messages to your AI from anywhere using Telegram
  • Tap inline buttons to trigger predefined AI sessions
  • Receive formatted responses with code blocks and images
  • Get heartbeat notifications about pending tasks and reminders
  • Secure authentication with your Telegram User ID

Task Scheduling

Schedule AI tasks to run automatically — daily standups, weekly reports, periodic checks, and more.

  • Cron-based scheduling with natural language input ("every morning at 9am")
  • One-time reminders and future tasks
  • Repeat intervals — hourly, daily, weekly, monthly
  • Each scheduled task runs in its own session with full AI access
  • View, pause, resume, and delete schedules from the sidebar

Heartbeat Monitor

RAPR AI watches your pending tasks and sends proactive reminders via Telegram before deadlines slip.

  • Automatic periodic check of your pending tasks and memories
  • Telegram push notifications for upcoming deadlines
  • Configurable check interval and notification preferences
  • Works even when the UI is closed — runs in the background

Universal MCP Connectors

Core

Connect hosted MCP servers from Zapier and Composio, then expose the app actions you choose to every AI runner in RAPR AI.

  • Zapier MCP exposes enabled actions from 9,000+ apps and 30,000+ actions
  • Composio MCP adds a second universal connector path for 100+ OAuth-backed toolkits
  • Both connectors bridge through npx -y mcp-remote as standard MCP servers
  • Personal MCP server URLs stay in the encrypted token vault
  • Discovered tools are routed to Claude, Gemini, Codex, Ollama, and other runners
  • Connectors are opt-in and disabled until the user pastes a trusted MCP URL

Composio MCP Connector

New

Paste a Composio MCP URL once and RAPR AI can discover connected app toolkits through the existing MCP manager.

  • Validates Composio-hosted MCP URLs before storing them
  • Stores COMPOSIO_MCP_URL in the encrypted token vault
  • Connect, disconnect, and status endpoints are built into the MCP routes
  • UI renders a first-class Composio connect card alongside Zapier
  • Keeps the no-API-keys-per-tool story: one MCP server URL unlocks connected toolkits

Knowledge Transfer

Import your knowledge from ChatGPT, Claude web, Gemini, or any other AI into RAPR AI's shared memory.

  • 2-step process: export a structured prompt, paste the response back
  • AI parses the response into categorized memories automatically
  • Deduplication prevents importing facts you already have
  • Works with any AI — ChatGPT, Claude, Gemini, Copilot, or others
  • Your accumulated AI knowledge in one place instead of scattered across apps

Custom Instructions

Write instructions once in Settings, and every AI provider follows them — Claude, Gemini, Ollama, all of them.

  • Persistent instructions stored in your preferences
  • Injected into every AI system prompt automatically
  • Works across all providers — no need to repeat yourself
  • Set coding style, tone, output format, domain rules, and more

Personalization

Name your AI assistant, tell it your name, and it will address you personally in every conversation.

  • Set your name and your AI's name in the Setup Wizard or Settings
  • AI addresses you by name and introduces itself by its custom name
  • Personalization injected into all provider prompts

Extensible Marketplace

Browse and install skills and integrations from the RAPR Packages marketplace — one click to install.

  • One-click install of community skills and integrations
  • Skills for coding, writing, research, automation, and more
  • Enable or disable skills with a single toggle
  • Publish your own skills to the marketplace

Local-First, Cross-Platform Runtime

Run local models with Ollama or connect the AI accounts you already use. The terminal runtime now supports Windows shells and POSIX shells.

  • Ollama integration with auto-detection of installed models
  • Full file system access — AI reads and writes your project files directly
  • Offline-capable workflows when every selected step uses a local model
  • All API keys and tokens encrypted with Fernet at rest
  • Cloud backup with AES-256 encryption when you choose to sync
  • Terminal sessions branch between cmd.exe on Windows and bash, zsh, or sh on macOS/Linux
  • Interrupt handling uses CTRL_C_EVENT on Windows and process-group SIGINT on POSIX

AI Pipelines

Chain multiple AI steps into automated workflows. Build complex automations with simple natural language.

  • Sequential step execution with output piping
  • Mix different AIs in a single pipeline
  • Save and reuse pipelines as templates
  • Error handling and retry logic built in

Session Templates

Save session configurations as reusable templates. One click to spin up a pre-configured AI session.

  • Save any session config — AI model, project folder, initial prompt
  • Share templates across devices with cloud backup
  • Quick-launch from the sidebar template panel

Security & Release Trust

Fernet encryption for credentials, explicit approval gates, budget controls, audit logs, and a cert-aware signing pipeline for release builds.

  • All API keys encrypted at rest with Fernet symmetric encryption
  • Daily budget caps per AI provider to control spending
  • Auto-retry with fallback models on failure
  • Full chat history with searchable logs
  • Cloud backup encrypted with AES-256
  • Approval queue gates destructive or sensitive actions before execution
  • Build pipeline can sign both the app executable and installer with OV, EV, or dev certificates
  • Signing skips cleanly when no certificate is configured, so local builds do not break

200+ Built-in Skills

Every skill auto-injects best practices into the AI prompt. Your AI becomes an instant expert — no configuration needed.

  • Playwright — Automate browsers, fill forms, take screenshots, run end-to-end tests
  • Docker CLI — Manage containers, compose stacks, build and deploy images
  • AWS, Azure, Google Cloud — Manage cloud resources via CLI from chat
  • GitHub CLI — Create PRs, manage issues, trigger Actions, review code
  • Terraform — Deploy infrastructure as code via natural language
  • Firebase, Supabase — Manage databases, auth, storage, and functions
  • Kubernetes — Manage pods, deployments, services via kubectl
  • Stripe — Test webhooks, manage payments and subscriptions
  • Plus: Redis, MongoDB, Vercel, Cloudflare, Twilio, and 50+ more

18 integrations, hundreds of models.

OpenRouter and Ollama alone reach hundreds of hosted and local models. Plus built-in Claude Code. Mix and switch any of them in one workspace, with no lock-in.

Google Gemini CLI

Google's official Gemini CLI

OpenAI Codex CLI

Codex and GPT models

Cursor CLI

Cursor coding agent

OpenClaw

Claude Pro/Max OAuth coding agent

Aider

Terminal pair programming

Goose

Block's open-source agent

Kilo Code

Open-source coding agent

Open Interpreter

Natural-language computer control

Antigravity

Google's agentic platform

Amazon Q Developer

AWS coding assistant

GitHub Models

Models through GitHub

OpenRouter

100+ models, one API

Groq

Ultra-fast LPU inference

Ollama

Run open-source models locally

Local AI

LM Studio, Jan.ai, llama.cpp servers

NemoClaw

NVIDIA sandboxed agent (Nemotron)

Zapier MCP

9,000+ apps and 30,000+ enabled actions

Composio MCP

100+ OAuth-backed app toolkits through one MCP URL

Connect via subscriptions, OAuth, local servers, or your own API keys. No per-token markup from RAPR.

200+ Skills + Universal MCP

RAPR AI auto-picks from your installed skills for each task. Zapier MCP and Composio MCP add broad app access through the tools you explicitly enable.

Playwright
Docker
AWS CLI
Azure
Google Cloud
GitHub
Terraform
Kubernetes
Firebase
Supabase
Stripe
Redis
MongoDB
Vercel
Cloudflare
Word Docs
Excel
PowerPoint
PDF
Twilio
Prisma
FireCrawl
Web Search
File Organizer
M365 CLI

Plus 175+ more skills for coding, DevOps, data, finance, education, and communication.

Ready to Get Started?

Download RAPR AI and unlock every feature — free and open source.