Open HuggingChat → MCP Servers panel → Custom Servers → "Add Your First Server."
First Break AI
What's New
For builders Install the First Break AI MCP server Model Context Protocol · HuggingChat · Claude · Cursor · Codex
First Break AI runs a live MCP server (Model Context Protocol) — a standard way for AI assistants to call tools over HTTP.
Add it once in HuggingChat, Claude Desktop, Claude Code, Cursor, or OpenAI Codex, and your AI can use cohort tools (ask, find, validate, and more) without leaving the chat.
Same URL on every client card below — or copy the config snippet that matches your tool.
Settings → Developer → Edit Config → paste under mcpServers (uses mcp-remote shim). Restart Claude Desktop after saving.
Terminal: claude mcp add --transport http fba-cohort <url> — or add to ~/.claude.json / project .mcp.json. Run /mcp in a session to verify.
Settings → Features → Model Context Protocol → add server, or edit ~/.cursor/mcp.json / .cursor/mcp.json with a remote url entry.
Add to ~/.codex/config.toml under [mcp_servers.fba-cohort] — or run codex mcp in the CLI. Then /mcp in a session to verify.
What can my AI actually do with FBA installed?
ask— answer any cohort question using the FBA roadmap, lessons, FAQ.do— enroll you, send the Discord invite + welcome email.find— locate specific lessons, steps, or office-hours topics.validate— judge your submitted work against the rubric and unlock the next step.next— tell your AI what to ask you next so you keep moving.
Or, if you'd rather stay on this page, .
Start here
Two ways into First Break AI
Pick a path — they're built to work together. Most learners ping-pong between the map and the lessons.
Cohort: 1 May 2026 — 30 June 2026 (2 months) Live
First Break AI is a free, community-driven cohort for anyone who wants their first break in AI. It doesn’t matter what you studied or where you work — what matters is that you’re ready to learn by doing. We focus on what matters: running and training models, understanding inference, and shipping AI-powered products. Most learning is self-directed and peer-supported; the roadmap, checklist, and resources live in the open so you can contribute and others can follow. The goal is simple: upskill, build, showcase — and get that first role or first break in AI.
Who it’s for
You don’t need a specific degree. You need a starting point. This cohort is for anyone — students, professionals, career switchers, the simply curious — who wants their first real break in AI. No applications, no prerequisites. Follow the roadmap, build in the open, and let your work speak for itself.
Cohort lead: FireHacker (blog, TIL) · GitHub @thefirehacker · About
Ready to start your journey?
This is a free, community-driven cohort. To join, sign up on Discord — that’s where the cohort lives, where you get help, and where you connect with fellow learners.
Join Discord and start learning
Office Hours: Every Friday, 9:00 — 10:00 PM IST. Meeting link shared on Discord.
Recommended for learners aged 16+. If you are under 18, please join only with parent or guardian permission. By joining you agree to our Community Guidelines, Terms, and Privacy Policy.
Bonus · The Anatomy
Want the full picture?
The cohort drawn as one inference flow — six steps, every detail on a single poster. Click to enlarge.
Get started
| Section | Description |
|---|---|
| Lesson 0: Welcome to First Break AI | Cohort intro video with interactive transcript — start here. Also on YouTube (full playlist) |
| Roadmap | Learning path: Quarto blog, local inference, training, and beyond |
| Checklist | Accounts to create (Hugging Face, GitHub, Fetchlens.ai, Colab) and who to follow |
| AI Setup | AI-based IDE (Cursor / Claude Code), ChatGPT, Open Router |
Frequently asked
Do I need to know math to take this course?
No. We deliberately go in reverse: start by running a model on your laptop, then add math (attention, RoPE, softmax) only when you have already seen the thing it describes. That is the same teaching approach Andrej Karpathy uses in Let's reproduce GPT-2.
Lesson 1's section "Do you need deep math for this?" goes deeper. If you have the math, great — it speeds you up. If you don't, what you can't skip is willingness to read code and run experiments.
Do I need a CS or tech background?
You need comfort with reading code (any language) and running terminal commands. That's the floor.
Past learners have come from career switches, design, math/econ/physics, and self-taught backgrounds. There are no applications and no prerequisites — the roadmap and checklist tell you exactly what to install on day one.
Is this really free? What's the catch?
Yes, free. No paywall, no upsell, no certificate-for-a-fee. For comparison, three current AI-engineering cohorts that audience-overlapping learners are already paying for:
| Length | Time / wk | Price | |
|---|---|---|---|
| Paid cohort A | 4 weeks | ~6 hrs | $2,000 |
| Paid cohort B | 8 weeks | 10–12 hrs | $2,000 |
| Paid cohort C | 6 weeks | 5+ hrs | (sold out) |
| First Break AI | 8 weeks | 5–8 hrs | $0 |
Same scope, free, longer cohort window. The catch is you have to show up — the cohort works because you build in public, not because we chase you.
How much time per week do I need to commit?
Plan for 5–8 hours a week to comfortably keep up — one lesson, Friday office hours, and the homework. Comparable paid cohorts ask for 5–12 hours per week, so this is in the same range.
Many learners spend more by choice; the roadmap goes deeper than the weekly minimum and the code is open. Cohort 01 runs 1 May 2026 – 30 June 2026, but everything stays online afterward, so you can also move at your own pace.
What hardware do I need? Mac, Windows, or do I need a GPU?
Any modern laptop is enough to start. We design the lessons CPU-first — for example, Lesson 1 runs Qwen3-0.6B in pure C on your laptop's CPU with no GPU required.
Mac (Apple Silicon or Intel), Windows (with WSL), and Linux all work. When the cohort hits larger training experiments in Step 4, the AI Setup page covers free-tier cloud GPU options.
Will I get a certificate or job placement?
No certificate, no placement service. What you walk away with is a portfolio: a Quarto blog of everything you learned, a model you trained from scratch, a deployed AI product, and optionally a merged open-source contribution.
In 2026, that portfolio is a much stronger hiring signal than a course PDF. Recruiters can see your blog and your GitHub commits; they can't see a certificate badge.
What if I miss a session, or join mid-cohort?
Everything is asynchronous-friendly. Office hours are recorded; AI-generated podcast summaries and lesson recaps go up on The Journey within a few days; transcripts are interactive and searchable.
You can join Cohort 01 today, in three weeks, or after Cohort 01 ends — the materials stay online and the Discord stays active. Nobody is gated.
How is this different from fast.ai, Karpathy's lectures, or HuggingFace's free course?
First Break AI is openly inspired by Andrej Karpathy's Let's reproduce GPT-2 (124M) and llm.c — the from-scratch, no-magic, run-it-in-pure-C approach to teaching how LLMs actually work. Those are still the gold standard for a single lecturer building a model in front of you, and they are the philosophical foundation under this entire cohort. The pure-C Qwen3 demo in Lesson 1 (code by William Song) is the direct intellectual descendant of llm.c.
We sit on top of that work and extend it in four directions:
- Cohort-based. Live Friday office hours, 9–10 PM IST, where you can ask anything. Karpathy's lectures are one-way; this is two-way.
- 2026-current stack. Qwen3 instead of GPT-2,
transformersv5, vLLM, agent tooling, the HuggingFace supply-chain lens — but the teaching philosophy (run the model first, study the math after) is Karpathy's, applied to today's models. - Shipping a product is in scope. Step 5 of the roadmap is building a deployed AI product, not just training a model in a notebook.
- Agent-readable, by design. The site ships structured data (
Course,FAQPage) for AI answer engines, and Lesson 1 walks you through Croissant — the same machine-readable lens, applied to model and dataset releases. Per-lesson manifests are on the roadmap.
fast.ai and HuggingFace's free course are excellent complementary reads, and the Roadmap links to them where they fit.
Does the First Break AI cohort have an MCP server?
Yes. First Break AI publishes a free, public MCP server (Model Context Protocol) so your AI assistant can talk to the cohort directly — search lessons, answer roadmap questions, help with enrollment, and validate homework against the rubric.
The server uses Streamable HTTP (the current MCP transport). Install it in HuggingChat, Claude Desktop, Claude Code, Cursor, or OpenAI Codex by pasting one URL or config snippet. Step-by-step cards are in the Install the First Break AI MCP server section at the top of this page.
Tools exposed by the server include ask (cohort Q&A from lessons and FAQ), find (locate lessons and office-hours topics), do (enroll and Discord invite), validate (check submitted work), and next (suggest your next step). You can also use the Open Cohort Lens chat widget on this site without installing anything.
No API key is required for read-only cohort tools; the cohort teaches MCP by having you install the same server you learn from — the same “agent-readable supply chain” idea as Lesson 1’s Croissant walkthrough, applied to the cohort itself.
About AI-generated content: Office hours transcripts and cohort material are used to create AI-generated podcast lessons, summaries, and other learning resources published on this site and The Journey. All learner names are anonymized — no real names are used. By participating in office hours, you consent to your contributions being used in this way.