See what the mod videos showed
Videos showing Claude Code “mods” circulated. In one of them, Claude started using the computer use of the Codex app: you clicked, typed, and read the screen through another AI.
In another, several Claude sessions worked at the same time, each in a different role, with a dashboard showing who was leading and who had stalled. The nickname was "Threads".
Almost everything was done by reverse engineering: look at the running application, find out how it communicates internally, and connect a wire there, without an invitation from the manufacturer.
🆕 New here? Five words that come up all the time
- Claude Code — Claude running in your machine's terminal. It reads and creates files and runs commands in the folder where you open it.
- Codex — OpenAI's equivalent tool, also in the terminal. The command
codex execsends a request and gets the response without opening the screen. - Agent — an AI equipped with tools that acts on its own: creates files, runs commands, opens websites.
- Mod (plugin) — a package that adds things to Claude Code: a panel, a new command, a rule that runs before each action.
- Reverse engineering — figuring out how a program works internally by observing it, without official documentation.
How to read the diagram: The blue section is what the manufacturer provides (the app using the computer). The red dashed section is the improvised connection someone built to access it. That connection disappears when the app updates.
what the videos showed
the nickname "Threads"
the AI operates the screen
reverse engineering
Separate product discovery
Reverse engineering has value. It proves that something is possible, shows how the program communicates and shuts down the premature "it can’t be done."
The problem starts when the discovery becomes part of your daily system. Then you depend on something the manufacturer could change tomorrow without notice. The rule for this course fits on one line: reverse engineering is a discovery technique, not a product.
✓ As a discovery tool (works)
- ✓ Shows the idea works before you invest
- ✓ Shows what data the program exchanges and where it goes
- ✓ Prompts the right question: “Is there an official way to do this?”
- ✓ Runs on a test machine, with the date recorded
✗ As a product (doesn’t work)
- ✗ Breaks with every app update
- ✗ No one guarantees that the behavior stays the same
- ✗ May violate the provider’s terms of use
- ✗ You become the only one who knows how to fix it
How to read the diagram: reverse engineering appears only in the first frame. What determines the path is the green box in the middle. In most cases in 2026, the answer is "yes," and your system follows the blue arrow.
💡 Phrase to remember
When someone shows you an impressive trick, ask: "is this the way, or proof that a way exists?". Almost always, it's the latter.
proof that it's possible
what you use every day
the method maintained by the manufacturer
date label
Learn about today’s official approaches
What the videos reverse-engineered was necessary when they were recorded. Today, most of it already has through the official channel: a command or feature that the manufacturer documents and maintains.
The table shows the switch. The last column tells you which module in this course shows you how to do it with the kit.
| The video did | Official method today | In the course |
|---|---|---|
| Claude calling Codex from inside the app | codex exec, wrapped in the bridge runtime/pontes/codex-exec.sh | 2.1 |
| "Threads": multiple sessions in parallel | claude --bg to release and claude agents to list | 3.3 |
| Panel and buttons inside Claude Code | official Claude Code mods and plugins | 4.3 |
| Helper that approved permissions on its own | permission rules in the .claude/settings.json and autonomy levels | 4.1 |
| AI operating the screen | browser for agents, with a written policy | 2.4 |
What to look for in the table: not one row in the middle column depends on "opening the app internally". They’re all commands or settings published by the vendor.
🆕 New here? What are "official route" and "CLI"
Official method it’s the route the manufacturer provides for other programs to communicate with its system, and that it commits to maintaining. CLI (command-line interface) is a program you use by typing commands in the terminal, such as claude, codex or git.
Yesterday: only available internally
Without an official command, the approach was to observe the app and connect a wire. Testing it made sense.
Today: the vendor published the approach
Background sessions, plugins, screenless execution: it all became a documented feature.
Tomorrow: check again
Command names change. That’s why the kit has a diagnostic (doctor.mjs) and each recipe has a test that you run on your machine.
Claude asks Codex
parallel sessions
official mods
written permissions
Understand what the kit is
O INEMA Agent Runtime is a kit you copy into your project. It helps Claude Code and Codex discover, connect to, and use the tools you already have, with clear rules about what they can do on their own.
It's not a platform or a server. It's a folder runtime/ with four rules files, diagnostic and verification scripts, bridges, a guard, and seven tested recipes. It works through the subscription for Claude Code and/or Codex.
🆕 New here? A subscription isn't an API
A subscription is the monthly plan you already pay for to use Claude or ChatGPT. It has a quota: a usage limit per period. A API is another account, billed based on usage and designed for programmers. The kit runs only through the subscription: no paid API is required.
runtime/LEIA-ME.md the cycle and the ladder of paths runtime/CAPACIDADES.md tools map and which path to use runtime/POLITICA.md what the agent does on its own, asks before doing, or never does runtime/ROTEAMENTO.md which model to use for which task (saves quota) runtime/scripts/ doctor · observar · verificar runtime/pontes/ codex-exec.sh · mcp-modelo/ runtime/mods/ runtime-guarda · runtime-painel runtime/receitas/ R1 a R7, each with a Proof AGENTS.md · CLAUDE.md Codex and Claude read the same rules .claude/ applied policy, a team of 3 roles, guard enabled
✓ The kit is
- ✓ Files you read, edit, and version
- ✓ Same for Claude Code and Codex
- ✓ Recipes with proof you can run
- ✓ MIT License, open
✗ The kit isn’t
- ✗ A server running in the background
- ✗ A web dashboard to subscribe
- ✗ Depends on a paid API
- ✗ A reverse-engineered shortcut
💡 Why files instead of an application
A text file the agent reads on its own, you can understand without programming, and Git tracks every change to. When a rule changes, you edit one line. No provider needs to be online for your kit to work.
the folder you copy
README, CAPABILITIES, POLICY, ROUTING
each one with proof
no paid API
Go through the seven-step cycle
The file runtime/LEIA-ME.md organize everything into a cycle: discover → connect → route → execute → observe → verify → learn. Each step has a file and a question.
And every action, at every stage, goes through the POLITICA.md: it says what the agent does on its own and what it asks first.
How to read the diagram: The dashed ring shows the order of the steps, and "learn" (green) sends the cycle back to the beginning with better rules. The lines to the center show that no step escapes the policy.
| Step | Kit file | Question | Track |
|---|---|---|---|
| Discover | CAPACIDADES.md | What tools are available, and how does the agent access them? | T1 |
| Connect | receitas/, pontes/ | How do I turn this approach into something the agent calls? | T2 |
| Route | ROTEAMENTO.md | Which model solves this for the lowest cost? | T3 |
| Run | R2 recipe | One agent or a team? | T3 |
| Observe / Verify | "command → expected output" criterion | How do I prove it worked? | T3 |
| Learn | POLITICA.md (Learning section) | What should change? The agent proposes; you approve. | T4 |
What to look for in the table: the questions are the same ones you’d ask when hiring someone: what exists, how it gets there, who does it, how to check it, and what improves.
💡 What counts as “proof” in the kit
Every recipe ends with a Proof: a command and the text that must appear. Example from recipe R1: run the bridge with "Reply only PONG" and see PONG. If it doesn't appear, it didn't work, even if the agent says it did.
seven steps that repeat
by stage
command → output
applies to all
Meet Clara and Sônia, and see the course map
Two people accompany you through the course. Their data is in the example files that already come with the kit, in runtime/exemplos/. This way, every command you run returns exactly what the lesson shows.
Clara, 47 — clinic owner
The clinic schedule is a spreadsheet. Dr. Ana and Dr. Bruno see patients there. Clara wants the agent to tell her which time slots are available without anyone having to open the spreadsheet.
File: runtime/exemplos/agenda.csv
Sônia, 51 — accountant
Her client is a distributor whose ERP has no API: it only exports a sales CSV. Sônia wants the total per customer without putting a spreadsheet together by hand every month.
File: runtime/exemplos/erp-vendas.csv
🆕 New here? ERP and CSV
ERP is the company management system (sales, inventory, invoices). CSV is a plain-text spreadsheet: each line is a record, and the columns are separated by commas. Almost every legacy system can "export CSV," and that's enough for the agent to read it.
How to read the diagram: each step depends on the previous one. You only connect (T2) what you discovered (T1), only build a team (T3) with working bridges, and only give the agent more freedom (T4) with policies and a failure log.
Quick test (optional): what role does reverse engineering play in this course?
spreadsheet calendar
ERP without an API
one depends on the other
your bridge and your team
🎓 Module summary
Next module:
1.2 — Install and troubleshoot