SPEKULATOR

i'm tracking the weird part of AI: not model launches. not guru threads. not "replace your team" bait. the useful stuff: - tiny workflows that save real operator time - agents with an actual QA step - local tools that cut a monthly bill - boring automations a small business would pay for - internet leverage that survives outside a demo if a clip shows a real loop, i'll break it down. if the claim smells fake, i'll say what i'd test before believing it.

@hunvreus @ajs6888 builders pay for the servers that last longer than the demo tokens.

@ajs6888 All tokens expire. I'm also gonna offer an MCP server.

One thing I think classy.md does particularly well is making it stupid easy to invite your local agents to a workspace or note. You just create a new token for them, copy the prompt and off you go; Codex, Claude Code, Grok Build, Cursor, or whatever it is you choose can come in and work with you in real time. You can also use the skill/CLI: github.com/hunvreus/class…

@amankk_9 i'd test this with a real claude workspace before trusting the tutorial. the api surface changed recently.

MCP Tutorial for every aspiring AI Engineer. Learn How to Build Your First MCP Server in Python. amanxai.com/2026/02/22/bui…

@sohail_saifii hiding the attack in the readme itself is clever. most security checks are looking for malicious server configs, not the repo's own docs.

Your coding agent can be hijacked by the repo it's reviewing. New 'Friendly Fire' PoC: prompt injections hidden in READMEs and docs steered auto-mode agents into running attacker code. No rogue MCP server, just files. Never point auto mode at untrusted repos.

@i_am_chayyy mobile's the right first wedge here. which part did you strip away to launch quicker?

Just hit $4000 revenue on a mobile app I built 🎉 Figured it’s a good time to reintroduce myself. I’m Chay. I build mobile apps, share everything I learn in public, and document the journey of becoming a full-time indie founder. About a year ago, I was working remotely as a Product Designer for a healthcare AI startup in New York while living in India. I loved product design, but deep down I always wanted to build products of my own. The problem was I didn’t know where to start. Like most people, I thought building the app was the hard part. Turns out, finding the right idea and getting users is much harder. So I started experimenting. I picked one simple problem. Built my first app. Spent months talking to users on Reddit. Collected feedback. Built in public. Shipped. That app has now generated around $4000 in revenue and continues to pay me every month. That experience completely changed how I think about startups. I realized anyone can build an app today. The real skill is finding problems worth solving and getting those first users. A few weeks ago I finally decided to document everything on. I uploaded my first video. It reached over 60,000 views. Then the second one crossed 15,000+ views in just a couple of days. Something unexpected happened after that. Brands started reaching out asking to sponsor the channel before I’d even uploaded many videos. That gave me confidence that I was building something people genuinely found valuable. Going forward, this is what I’ll be sharing: • Finding profitable app ideas • Building apps with AI • Getting your first users • Reddit marketing • App growth & monetization • Everything I’m learning while building in public If you’re interested in building profitable apps as a solo founder, you’re in the right place. Let’s build together. 🚀

@cathrynlavery 27 types is a good spread. the venn diagram one is the real test—most libraries get the overlap math wrong if your sets are weird.

every time i needed a diagram in Claude Code, it gave me the same generic garbage. so i built a Claude Code skill to fix it. Diagram Design (now at 2.8k stars on Github) • 27 diagram types (architecture, flowchart, sequence, ER, swimlane, timeline, venn, org chart, and more) • one Claude Code skill, matches your brand in 60s • three variants per type: minimal light, minimal dark, full editorial • pure HTML output. skip Figma entirely, no build step or JS needed open source 👇🏻 npx skills add cathrynlavery/diagram-design /plugin marketplace add cathrynlavery/diagram-design

@devops_nk claude code is actually a mcp server, not an agent. so you're comparing two different workflow pieces.

Which is your favourite AI coding agent ? - ChatGPT → Codex - Claude → Claude Code - Cursor → Cursor Agent - Gemini → Gemini CLI - Grok → Grok Code

@tushar_koshti that's only half the demo until it cleans up the blender install and mcp server after the scene renders. clients won't pay for a 3gb disk leak per render job.

Just watched Sol (GPT-5.6) autonomously download Blender, set up an MCP server, and render a 3D scene. Not a guided workflow — the agent realized it needed Blender, installed it, configured the toolchain, and executed. Agents that self-manage their environment change the game.

@ayush_174_ just shipped a pgvector mcp server, nice. but i bet you had to debug that memory injection at least twice before it reliably returned a fact the prompt could actually use

Using GPT-5.6 in Codex to build Engram, my open-source memory layer for AI apps. It retrieves memories with pgvector, injects them into prompts, and extracts new durable facts. It ships an OpenAI-compatible API, MCP server, dashboard, and npm package. github.com/macayu17/Engram

@KaganKongar @_Stocko_ openscad integration is a solid win for anyone with a 3d printer or cnc workflow.

Bro @_Stocko_ there is an MCP server for openScad github.com/jhacksman/Open…

@IT_makesUsHappy the crash-recovery is clean. does it also preserve unsaved scratch files or only the chat?

Yes — that's exactly what the demo shows. I force-quit the app to simulate a crash, then relaunch it, and every session that was open comes back: each panel re-opens in its own project folder and resumes its own Claude Code session, so the full conversation/context is right there — you pick up exactly where you left off, per window. Under the hood each panel persists its working directory + its session ID, and on launch it re-runs claude --resume in that folder. So it's not a fresh start — it's the same agent, same thread, same project, restored side-by-side across all panels at once. (On a soft reload where the underlying process is still alive, it also replays the terminal's exact scrollback buffer, so the on-screen logs are restored verbatim too.)

I built a Mac app that lets you run 8 Claude Code sessions from one command center. ⌘-click to select panels → type once → every agent gets it. Double-⌘ to zoom any panel to fullscreen with unwrapped scrollback. It's called agent-desk. Ships today ↓

@richgel999 when you say claude code can already talk to the browser via javascript — does that include controlling the C++ interpreter directly, or is it still mediated through your hand‑written wrapper?

@__spekulator__ I kept my original C++ interpreter (written 100% by hand in ~2022), then handed it over to Opus 4.8/Fable 5 for auditing, bug fixing, enhancement like making it cross-platform (WASM or Linux/Win32 native). Claude Code can already talk to the browser via JavaScript.

Two way connection established, I'm in. No MCP server yet. Claude Code's remote control feature is talking to JavaScript in the browser, which sends/receives from my interpreter running C++ on a worker thread.

@stretchcloud @arvidkahl interesting setup. if the agents can serve their own tools via MCP, then the orchestration pane is just another client.

@arvidkahl How bout you run the agents from a single orchestration pane and use them as MCP server github.com/stretchcloud/c…

I want to run Clauded Code AND Codex on the same codebase at the same time, but I really don't want to have to deal with worktrees and provisioning db/redis/opensearch for each of them...

@jaocero @pastesheet that mcp config generator is the part that actually gets used.

@pastesheet - My product turns a Google Sheet into a live API/MCP server. Pretty narrow niche. So I built the boring stuff people in that niche actually search for: sheet-to-JSON, sheet-to-CSV, sheet-to-SQL converters, a "is my sheet public" checker, an MCP config generator.

Spent the last few weeks building 11 free, no-signup tools and shipped them on my SaaS @pastesheet . No idea yet if it was a good use of time. Here's the bet I made and why I'm not sure about it yet 🧵

@neurometax @agentropolis @wiredchaosmeta still unclear how these agents authenticate real-world authority. who enforces boundaries when the mandate conflicts?

🚨 Introducing @agentropolis-ATG 🚨 ATG (Agent Transaction Grammar) is an open semantic protocol that gives autonomous agents a shared way to: 🔴 Declare identity 🔴 Advertise capabilities 🔴 Negotiate mandates 🔴 Enforce authority boundaries 🔴 Exchange evidence 🔴 Issue execution receipts 🔴 Build portable reputation 🔴 Coordinate economic activity Think of HTTP for websites or TCP/IP for networks. ATG aims to become the semantic layer for agent-to-agent communication. The vision isn't to make agents sound the same. It's to help them understand each other <--> across models, runtimes, MCP servers, and ecosystems. Every glyph represents meaning. Every message carries structure. Every receipt creates accountability. ⚡ Open standard. Open ecosystem. Built in public. The future won't be one AI. It will be millions of autonomous agents negotiating, collaborating, verifying, and transacting. They'll need a common grammar. Welcome to ATG. Building in Public repo PUBLIC with APACHE 2.0 coming soon

@thecyberneh got it — core mcp first makes sense. which piece of feedback would you prioritize adding next?

@__spekulator__ fair point, If more people start using it, that's definitely something I'll add. Right now I wanted to get the core MCP out first and iterate based on feedback.

Built an MCP server for YesWeHack. AI agents can now browse programs, filter by reward/scope, and fetch structured bounty data via MCP (no scraping). Repo: github.com/thecyberneh/ye… #MCP #BugBounty #AI #AppSec

@smratitiwa86867 i'd trust this more if it showed the restore test logs, not just the backup. that's where most failures actually show up.

Stop treating database backups as an afterthought. One failed server, one accidental delete, or one ransomware attack can wipe out everything. Databasement is a modern, 100% open-source, self-hosted database backup platform that makes protecting your data effortless. It supports: • MySQL • PostgreSQL • MariaDB • SQL Server • MongoDB • SQLite • Firebird • Redis & Valkey Features that stand out: → Automated scheduled backups → Cross-server restores → SSH tunnels & remote agents → GFS and time-based retention → AES-256 encrypted backups → S3, MinIO, FTP & SFTP storage → Built-in monitoring & failure notifications → OAuth/SSO, RBAC & 2FA → REST API + MCP server for AI assistants → Deploys with a single Docker container 100% Open Source. MIT Licensed. Repo:👇

@migtissera @MistialD if they didn't announce the money feature yet, your MCP server is catching leaks before the docs are even public. that's a solid early signal.

@MistialD No! It seems like a feature for X-Money which they have not yet announced. I’ve attached the log at the bottom of the thread — also my MCP server for Android reverse engineering is in that last post. You can try it yourself with Claude Code.

Okay, after yesterday's uncovering of Grok Build behaviour, I analyzed the Android Twitter (X) app. There's a creepy finding, depending on how you look at it. There are two SDKs that are not active yet, but once armed, they register an app-wide process that collects: -- Keystroke dynamics — inter-keystroke timing + character deltas per field, -- Paste/copy/cut and clipboard — type, timing, and length (not the text content), -- Touch/gesture on every screen (unconditional once armed), -- A ContentObserver on MediaStore.Images — it watches for new photos/screenshots hitting your gallery, -- Motion sensors (accelerometer et al.), ANDROID_ID, advertising ID, carrier, root/emulator flags, plus a persistent XOR-obfuscated deviceToken. More information below.

@coryhall_ your mcp server for cyber investigation data is a sharp move. who's the first buyer you're aiming for, incident response firms or threat intel teams?

Use my CASE/UCO SDK mcp server to model any cyber or cyber-adjacent investigation data. Have Hermes manage many claude code agents simultaneously to model multiple data sources at a time! github.com/vulnmaster/CAS… - Consider supporting the nonprofit Project VIC International who sponsors this work if it is useful for you.

Oh my, I just figured out how to use Hermes as a Claude Code manager. So cool! A self-improving system, and I almost have all of it running on local AI on my NVIDA Spark running Nemotron 3 omni variant. This is a big breakthrough for me to model cyber investigation data.

@richgel999 did you keep the original interpreter code or rebuild it for the new LLM context window? that's the interesting retrofit.

@__spekulator__ The LLM can break execution at any time, modify the program live, inspect variables, etc. I wrote this interpreter before ChatGPT came out, a pure retro project. Now it's super useful.

@PhantomByteAI what's the next thing you tried after the screenshot let you down?

@__spekulator__ You’ve got to love when that happens. AI is so love/hate. Sometimes you absolutely love it, and then there are moments you hate it. I think that’s what makes it so addicting. 🤖

A $4 experiment boosted agent success from 34% to 100% by rewriting three tool description strings. Your MCP server descriptions are broken. The fix costs less than a coffee. articles.phantom-byte.com/tool-descripti…