#mcp

Intermediate tutorial: have your AI agent install REA, an open-source reverse-engineering tool. Point it at an app and a feature, and it takes the feature apart without the source code, explaining the workflow with evidence. Once you understand the principle, it writes its own version from scratch in your project and adds one improvement. Includes 3 copy-paste prompts.

Reinstalling is easy. What's hard to get back on a new computer is your MCP servers, Skills, rules files, env vars and project dependencies. Ask your AI agent to inventory them, set up a private backup repo, list your keys, and write a restore script. Copy-paste prompt included.

Use when you want to know how a company gets customers, what ads it is running, how several competitors' channels compare, or you want a promotion plan modeled on a rival's playbook. Your AI agent runs the open-source tool REM to collect public evidence on ad libraries, tracking tags, SEO, social media, founders, pricing and past selling points, then writes a report where every claim cites an evidence number and anything it could not find is listed separately. Public sources only. It does not invent spend and does not copy the rival's creatives. Use case: Have your AI agent map out a rival's customer acquisition playbook, with every judgment backed by evidence.

Google's A2UI 1.0 lets an agent say what belongs on screen while your app paints it with its own native parts: forms, charts and confirmation cards. Jira and Alation already use it, and it runs in Claude and Gemini through MCP Apps. Includes a prompt that walks your AI agent through the quickstart.

Beginner tutorial: install the open-source Muse4AI app on a Windows PC, log in to your own Meta Muse account, and hand the connection command to Claude Code. After that, one sentence in chat makes images and short videos. Covers checking credit use first and splitting shots to build longer videos. Includes 3 prompts you can copy and paste.

The open-source Poki project has released a preview iPhone app. An AI agent reads the screen, taps, and types on the phone locally. It takes commands from the app, iMessage, Telegram, or a terminal, and it can plug into Claude Code and Codex as an MCP server. Payments, deletions, and messages need your approval in the Dynamic Island first.

Use when the user wants you to work directly in their ElevenLabs account: generating voiceovers, sound effects, music, images and video, or creating, editing and debugging voice and chat agents, or adding ElevenLabs to a project. There are two official ways in. The hosted MCP (https://api.elevenlabs.io/v1/mcp, signs in with OAuth, no API key) fits chatting in Claude, ChatGPT or Cursor. The open-source CLI @elevenlabs/cli (MIT) with the official ElevenLabs Skills fits the terminal, scripts, and managing agent configs as code. Confirm credit use before generating, and confirm and dry-run before any delete or push. Use case: Connect your ElevenLabs account so the agent can make voiceovers, music and sound effects in chat or the terminal, and create and debug voice support agents.

Meta has released Astryx, a React design system it used internally for eight years and that powers over 13,000 apps, on GitHub under the MIT license, with themes, templates, and a hosted MCP and command-line tool for AI coding assistants. It is currently in Beta and requires React 19.

Use when a user is building a React UI and the project uses Astryx (@astryxdesign/core), or when they name Astryx or ask about its components, themes, or upgrades. Check component docs and templates first via the hosted MCP (search, get) or @astryxdesign/cli before writing code; get consent before running commands that write files; never run the unscoped npx astryx. Use case: Have the agent write UIs following Meta's Astryx component conventions instead of guessing from memory

The agent completes tasks on your real phone, key steps such as payments wait for you to click allow on your computer, sensitive apps are invisible, and every step comes with an explanation and a screen recording.

Use when the user asks you to do things on their own iPhone or Android phone (look up information, reply to messages, place orders, organize photos, move files). Operates through the Pinky desktop app and phone_* MCP tools, with the user approving key steps such as payments; never enters passwords or verification codes, and never bypasses any block. Use case: Let an Agent operate your phone while you watch and approve

Developer deep dive: how the three agent protocols divide the work. MCP lets models connect to tools and data, A2A lets different agents collaborate with each other, and NLWeb lets websites be accessed in natural language.

Use when a user wants to understand how an app, binary, or Electron program implements a feature and has no source code. Based on the open-source REA (MIT): connects analysis tools for native binaries, JavaScript/Electron, .NET, and websites via MCP. Analysis runs locally, and conclusions come with evidence and limitations. Only for your own targets or targets you are authorized to analyze. Use case: Have the agent work out how a feature of an app without source code is implemented, and provide evidence and limitations.

Use when a user wants a ready-to-run reverse engineering workspace: analyzing samples with tools like Ghidra and Frida, doing CTF work, and producing evidence reports. Based on the open-source Open-ReverseLab (GPL-3.0): 180+ knowledge base articles plus 100+ MCP tools, with tools installed by board. Only for your own samples, lab targets, and authorized testing. Use case: Let an agent analyze samples and produce evidence reports in one workspace, using the knowledge base and over a hundred tools.

Use when a user needs to locate front-end encryption functions (login passwords, request signing, response decryption) and wants the results turned into reusable JSRPC, a Flask service and Burp autoDecoder configuration. Based on the open-source js-reverse-automation--skill (Apache-2.0), connecting to a real browser via chrome-devtools-mcp. Only for your own targets or targets you are authorized to test; shut down the services when done. Use case: Let the agent find front-end encryption entry points in a real browser and deliver callable services and Burp configuration.

Google AI Ultra includes monthly Colab compute that, at A100 rates, runs for over 200 hours, which is a good deal for running open-source models. Compares the official Colab CLI and the remote Colab MCP: the CLI suits complex deployments, while MCP lets you turn GPUs on and off from within a chat at any time. Includes a quota table, hands-on steps, cost-saving principles and FAQs.

Use when a user wants an agent to start a GPU or TPU on Google Colab, run code, deploy or fine-tune open-source models, fetch results, and release the machine. Combines the official google-colab-cli (Apache-2.0) and the Colab remote MCP into one workflow: the CLI handles file transfer, SSH, Drive, and one-off jobs, while MCP handles lightweight start/stop from a chat client. Enforces releasing runtimes after use, asking before expensive GPUs, and never paying on the user's behalf. Use case: Let an agent start a GPU on Colab for you, run open-source models, fetch the results, and shut down automatically. Use the CLI from a local terminal, or the remote MCP from a chat client.

Use when a user wants an agent to generate images, video, audio, 3D assets and game art, or to model, rig, build levels and create effects in Blender, Maya, ZBrush, Unreal, Unity or Godot. Built on the open-source scenario-labs/skills (MIT): run dry_run to estimate cost before generating, and confirm before spending; expert tools only need the matching software installed locally. Use case: Have an agent generate images, video, audio and 3D assets, estimate cost with a dry run before generating, and drive software such as Blender, Maya, ZBrush, Unreal and Unity.

Intermediate tutorial: copy the prompt from this site's BlenderMCP Skill page and give it to your Agent, which installs everything for you; after a restart, open Blender and model, change materials, and arrange scenes one sentence at a time, then save and run a pre-delivery check. Suited to 3D prototypes, asset sketches, and scene layouts.

Use this when a user wants AI to work in Blender through natural language for 3D modeling, scene building, or object editing, or asks how to connect Blender to MCP. BlenderMCP exposes Blender as an interactive environment via the Model Context Protocol, so clients like Claude can talk to Blender and the result is an editable scene. Before using it in a real production pipeline, still check the Blender version, plugin permissions, file paths, and output quality. Use case: Use natural language to have AI model, build scenes, and edit 3D objects in Blender, producing editable scenes.