next level nerdingRas Mic's stack

Ras Mic's stack

What Ras Mic builds with right now: five harnesses, two models, a six-part tech stack and seven tools. He says the list keeps changing as models and tools do. The useful part is why he picked each one, and the reason is almost always the same: the agent is good at it.

Jump to what he reaches for, the tech stack, or his workflows on this site.

Original sources

Ras Mic · My Agentic Development Stack (19 min) youtu.be His skills on GitHub github.com

The video

19 minutes, posted Oct 1, 2026. YouTube also shows it as “The Tools & Tech Stack I Use To Build”. Every timestamp on this page opens it at that moment. 4:00 to 6:13 is a sponsor read.

0:00 What this is: an update on everything he uses
0:35 Harnesses · T3 Code, his daily driver
2:49 Codex, for computer use and research
4:00 Sponsor (Grokbot), skip to 6:13
6:13 Claude Code: the CLI, never the app
6:42 Cursor and Devin: cloud agents, fired from Linear
7:41 Which harness? Whichever is cheapest right now
8:09 Models · Opus 5.5, and Astra for deep work
9:10 Tech stack · SvelteKit
10:08 Convex and Convex components
11:20 Effect for every service
12:37 PlanetScale, Vercel, Cloudflare
14:14 Tools · Greptile, Orgo, Paper, Mobbin, Wispr Flow, Linear
17:01 His skills: the factory in ninety seconds

The short version

One main pick per group, and what fills in around it.

HarnessT3 CodeCodex for computer use. Claude Code in the terminal. Cursor and Devin in the cloud.
ModelOpus 5.5GPT-6 Astra when the job is deep research or understanding a codebase.
TechSvelteKit + ConvexEffect for services. PlanetScale for Postgres. Vercel, or all of Cloudflare.
ToolsGreptileOrgo, Paper, Mobbin, Wispr Flow, Linear, and his own skills.

Words on this page

Twelve terms, in plain words. Skip this if you already build with coding agents.

Thread: One conversation with an agent, working on one task. He runs one thread per feature, many at once.
CLI: Command-line interface. A tool you run by typing commands in a terminal instead of clicking in an app. He uses Claude Code this way.
MCP: Model Context Protocol. A standard plug that lets an agent use another tool. Paper, Mobbin and Linear each have one, which is how his agents reach them.
Cloud agent: An agent that runs on someone else's servers instead of your computer. Cursor and Devin offer them; he starts them from Linear tickets.
Computer use: An AI driving a computer the way a person would: opening apps, clicking, typing. He uses Codex for this, for example to drive Blender.
PR: Pull request. A proposed code change on GitHub, waiting to be reviewed and merged. Greptile scores each one out of five.
Skill, AGENTS.md: Markdown files of instructions an agent reads before it works. His seven skills and one AGENTS.md make any harness follow his four steps.
SSH: A way to log into another machine over the network and type commands on it. He does this to set up machines with Claude Code.
Postgres: A widely used database. PlanetScale hosts it for him, for services that don't run on Convex.
Typed: In TypeScript, types let the code check that a function is called with the right kind of data. Effect extends that to errors, so a failure can't be quietly ignored.
Real time, websockets: Real time means the screen updates the moment the data changes. Websockets are the usual plumbing for that. Convex gives him real time without managing them.
Workers, R2, Hyperdrive, Durable Objects: Parts of Cloudflare. Workers run code, R2 stores files, Hyperdrive connects to a Postgres database like PlanetScale, and Durable Objects keep state for a running service.

What he reaches for

He lists tools by category. This turns the list around: pick the job, see what he uses for it and why.

Build a feature

Watch at 1:04

T3 Code, running Opus 5.5, with his skills loaded

T3 Code
Opus 5.5
his skills repo

One thread per feature. T3 Code runs the real Claude Code, Cursor or Codex harness underneath, and can send the thread to any machine on his network. The sidebar tells him which threads are done.

Why: He runs many at once. He needs to see what's finished without clicking into each one.

The four steps each thread follows

Harnesses

A harness is the app or command-line tool an AI model works inside. He uses all of them, partly because reviewing them is his job. 0:35

Daily driver
1:04
T3 Code
He calls it “a harnesses harness”: a desktop app that runs the real Claude Code, Cursor or OpenCode underneath. He connects his Linux box, a Mac and his MacBook Pro, and hands agents to whichever machine is free.
The sidebar works like an inbox. A thread settles when its PR is merged.
Computer use
2:49
Codex
The ChatGPT app. Computer use, research and anything 3D. He explored a hardware startup by having it drive Blender.
Not what he builds apps with.
Terminal
6:13
Claude Code
The CLI only, with Opus 5.5. He reaches for it when he's logged into another machine and needs something configured.
He says he almost never uses the app.
Cloud agents
6:42
Cursor
Mostly the web app, for cloud agents. He also uses its CLI. Wired to Linear so a ticket can start an agent.
“The cloud agent product is really impressive.”
Cloud agents
6:42
Devin
Same job as Cursor's cloud agents: work that should run somewhere other than his own machines.
Also assignable from Linear.

Which one should you pick?

His answer: the one that gives you the most for your money this month. It was OpenAI for a while; with Opus 5.5 he says it's now Anthropic. “I think you should be very disloyal.” Don't marry a subscription. 7:41

Models

Two. He expects this section to be wrong in two or three months, when the next model lands. 8:09

Main

8:18

Claude Opus 5.5

Building applications. Nearly everything.

Not because it's a giant leap in intelligence. It's well rounded: it follows instructions and holds up on long-running tasks. Since it came out he hasn't needed another model.

Second

8:40

GPT-6 Astra

Deep work, research, understanding a codebase in depth.

Still strong where the job is thinking more than building.

This already moved once. In his Sept 14 episode with Greg Isenberg, Astra was the workhorse. His software factory didn't change at all, which is the point of keeping the workflow in files.

Tech stack

“I still think there are smart decisions to be made on the tech that you use.” He doesn't use all six on every app, but every app uses at least one. 9:10

Framework
9:32
SvelteKit
His default for web apps. Sometimes Next.js instead. Both deploy cleanly to Vercel.
Agents are good at it, and the apps are fast by default.
Backend
10:08
Convex
Database and backend functions in one. Typed end to end, real time without websockets, and a library that fits Svelte.
He works at Convex and says so.
Integrations
10:27
Convex components
Building blocks for third-party services, shipped as code: Stripe (he built that one), WorkOS for login, PostHog, Browserbase, sandboxes.
Give the agent the link. It installs the component.
Services
11:20
Effect
A TypeScript library he uses for every backend service. Its selling point is typed errors: failures are part of a function's signature.
He found it too verbose to learn by hand. AI removed that problem.
Database
12:37
PlanetScale
Postgres, for services that don't sit on Convex. Works well with Effect. Reachable from Cloudflare through Hyperdrive.
“The big dog Postgres provider.”
Deploy
13:00
Vercel or Cloudflare
Vercel by default. If any part needs Cloudflare he goes all in: Workers, R2 storage, Hyperdrive, Durable Objects. His email service for agents runs entirely there.
Usually one or the other, not a mix.

Path one · a web app

SvelteKit → Convex → Vercel

The site talks to Convex, Convex holds the data, components bolt on payments and login. Sometimes Next.js in place of SvelteKit. His Stripe component is the example he shows.

Path two · a service

Effect → PlanetScale → Cloudflare or Vercel

No Convex. Bezalel is one big MCP server written in Effect, with a Svelte site in front. On Cloudflare he uses the whole set: Workers, R2, Hyperdrive, Durable Objects.

Tools

Seven, in the order he gives them. Paper, Mobbin and Linear each have an MCP server, which is how his agents reach them. 14:14

Code review
14:20
Greptile
Reviews every PR and gives a confidence score out of five. His greploop skill loops the agent on the feedback until it's five.
“You need a code reviewer.”
Agent computers
15:00
Orgo
A separate computer for an agent to use. Every Pluto instance has one.
He's an investor, and says to take his view with a grain of salt.
Design canvas
15:28
Paper
A Figma alternative built for agents, with an MCP so the agent can draw on the canvas. Pluto and his portfolio were redesigned in it.
Covered on the design page.
Inspiration
16:05
Mobbin
A library of well-designed sites and apps, also with an MCP. Where he collects looks to borrow.
Covered on the design page.
Voice to text
16:22
Wispr Flow
Dictation for prompts. He'd rather say it than type it.
He types when he needs to be specific.
Issue tracker
16:37
Linear
Tickets for his team, and the trigger for cloud agents: assign a ticket to Cursor or Devin and it starts.
He says it may be too much if you work alone.
Workflow
17:01
His skills
Seven skills and an AGENTS.md that make any of the harnesses above follow his four steps.
Free and open. The factory page takes them apart.

The Ras Mic set

Three pages. This one is the parts list. The other two are how he uses the parts.

How he designsDesign like Ras Mic Paper, Mobbin, T3 Code and Opus 5.5 in a six-step loop. Borrow a design system, try it on a page, write the fixes back.
How he shipsRas Mic's software factory Seven skills and an AGENTS.md. Isolate, build, prove, ship, and loop until Greptile says five out of five.
What he uses · you are hereRas Mic's stackThe harnesses, models, tech and tools those two workflows run on, as of October 2026.

He closes the video with a ninety-second recap of the factory. Nothing in it has changed since the page was written. 17:01

FakeNerd's take

This page has a date on it

Oct 1, 2026. Two weeks earlier he called Astra his workhorse; now it's Opus 5.5. He expects the model picks to flip again in two or three months. Read the reasons, not the names.

Notice who pays him

He works at Convex and invested in Orgo, and he tells you both. The picks can still be right. Just weigh them like you would any recommendation from someone on the payroll.

The real rule is “what is the agent good at”

Svelte, Convex and Effect all get the same argument: AI writes them well, and they make mistakes loud. That's the test worth copying, whatever it picks for you.

You need three of these, not twenty

One harness, one model, one code reviewer. He uses five harnesses partly because testing them is his job. Start with the subscription you already have.

The Linear trigger is the upgrade

On the factory page, every job starts with him typing a prompt. Here a ticket starts the agent. That's his factory taking one step toward the kind that runs without you.

Be disloyal

His advice on harnesses is to follow the subsidies and switch whenever the deal changes. Because his workflow lives in markdown files, switching costs him nothing. That's the quiet payoff of the factory.

Attribution

My Agentic Development Stack, Ras Mic (Oct 1, 2026). Every pick and reason on this page is his, from the video's captions. Tool descriptions are in our words.
His disclosures, as he gives them: he works at Convex, and he is an investor in Orgo. The video is sponsored by Grokbot, which isn't part of the stack.
rasmic.xyz, Michael Shimeles (Ras Mic). His skills: michaelshimeles/skills.
The by-job picker, the two build paths and the takes are FakeNerd's. Links go to each tool's own site, checked Oct 1, 2026. Not sponsored by anyone on this page.