I ────── The Free Codex Engine™

HY3 free window · tested 8 July 2026

The Free Codex Engine™.

OpenAI's coding agent, running on Tencent's free 295B model. Zero dollars. I wired it into my Agent OS and it built a full landing page first try.

A grand brass engine room with golden flywheels and emerald gauges reading zero, powered by a jade-and-gold Chinese dragon coiled through its pipes, breathing golden code glyphs

Codex is OpenAI's coding agent — it plans, writes files, runs commands. Normally it runs on OpenAI's paid models. But Codex has a hidden superpower: you can point it at ANY model with one config file. So I pointed it at HY3 — Tencent's new 295-billion-parameter model, which is free on OpenRouter right now — and dropped it into the Codex tab of my Agent OS as a one-click engine. Real receipts below, plus every command and all three gotchas I hit so yours works first time.

$0per token, free window

295Bparameters driving Codex

262Kcontext window

1st tryfull landing page built

II ────── How it works, in simple words

Codex is the hands. HY3 is the brain. The bill is zero.

Three pieces, one sentence each.

1. Codex — the hands. OpenAI's coding agent. You give it a task, it writes the files, runs the commands, and reports back. The agent itself is free to install — you normally pay for the model behind it.

2. HY3 — the brain. Tencent's new 295B Mixture-of-Experts model, built for exactly this: agentic coding and reliable tool-calling. On OpenRouter, the tencent/hy3:free version costs nothing — $0 in, $0 out — for a limited window.

3. The swap — one config file. Codex lets you define a custom "model provider". One small file tells it: talk to OpenRouter instead of OpenAI, use HY3. That's the entire trick. Your OpenAI login stays untouched.

The whole engine — same Codex hands, a free 295B brain swapped in behind them.

III ────── Exactly how it works, step by step

What happens when you hit send — no mystery.

Codex reads its profile file. Running codex exec --profile hy3 loads ~/.codex/hy3.config.toml. That file holds exactly two facts: the model is tencent/hy3:free, and the provider is OpenRouter at https://openrouter.ai/api/v1.

Your key rides along. The profile names an environment variable — OPENROUTER_API_KEY. Codex reads it and signs the request. No key ever gets pasted into the config itself.

Codex packs the task with its tools. Your prompt goes out with the tool list — write file, run command, read file — so HY3 can act, not just chat. Same agent machinery Codex always uses; only the address changed.

OpenRouter routes it to HY3 for $0. The :free suffix on the model name selects the free tier. OpenRouter forwards to Tencent's model and meters it at zero.

HY3 answers in tool calls. "Create landing.html with this content." Codex executes each call on your machine, sends the result back, and HY3 continues until the job is done. That loop is the whole agent.

In the Agent OS, a dropdown does all of this. My Codex tab now has an engine picker — 🐉 HY3 295B (free) is the default. The dashboard injects the same provider settings per request, so the OS never touches your Codex config either.

Your OpenAI setup stays untouched. Profiles are isolated. Plain codex still runs your normal models; --profile hy3 runs the free dragon. Delete one file to remove it.

One task's journey — profile picks the free brain, OpenRouter meters it at zero, files land on your disk.

So when someone asks "but what IS this?" — the honest one-line answer: it's one small config file that redirects Codex's brain from OpenAI's meter to Tencent's free 295B on OpenRouter.

IV ────── Straight from the source

The sources. Read it and run it yourself.

Official sources + everything I built ↓

the free modeltencent/hy3:free on OpenRouter → the mechanismCodex config + profiles docs → get a keyOpenRouter API keys → the hy3 free waveKilo Code's HY3 alert → my receipt · live buildThe landing page HY3 built → the os surfaceCodex tab in Agent OS →

V ────── My story · why this matters

I was you. Then I gave Codex a free brain.

Before

Codex was my favourite pair of hands — and every task it ran burned paid tokens.

Long agent loops meant watching the meter, not the work.

Free models either couldn't drive an agent or fell over on tool calls halfway through.

So the expensive brain did everything, including jobs a free one could handle.

Then HY3 went free on OpenRouter — and Codex only needed one file to switch brains.

After

Now Codex runs on a 295B model that costs nothing.

It one-shotted a complete landing page through my Agent OS on the first test.

The engine picker in my Codex tab defaults to the dragon — every task free by default.

And my paid models only wake up when a job actually earns them.

You can have this too. One file. Five minutes.

VI ────── The receipts

Real people. Real wins. Inside the Boardroom right now.

3,600+ Founders inside AIPB
400k YouTube subscribers
38 Countries · live members
163k X / Twitter followers
29k+ Udemy students

I'm not going to paste invented quotes here. The wins are real and written by the members themselves — agency owners, ecom founders, course creators, solo operators across 38 countries. Read them in their own words.

Read the 158-page wins doc →

Before you scroll on —

Commit to transitioning today. Not tomorrow.

This one is a five-minute job with a countdown on it — the free window won't wait.

Everything below is tested, this morning, on my machine.

So here's the deal.

If you're reading this — promise yourself one thing right now. You're going to finish this guide AND create the one config file before you sleep tonight. One file. Because the moment your coding agent runs on a free frontier-class brain, every experiment stops costing money — and you start running ten times more of them.

The people sitting still keep paying full price per idea. The people implementing today experiment for free.

Be one of those people.

Commit to the transition. Commit to taking action today. This changes everything about your workflow.

VII ────── I tested it — here's the proof

Real test: "build out a beautiful landing page."

Not a toy prompt — the exact request I typed into the Agent OS Codex tab with the HY3 engine selected. One shot: it planned, wrote landing.html — an 11KB, self-contained dark-theme page with a sticky nav, gradient hero, feature grid, stats band and CTA panel — and reported back. Zero edits from me.

The dark-themed landing page HY3 built through Codex — gradient hero headline, sticky navigation and call-to-action buttons

HY3's one-shot landing page, rendered straight from the file it wrote. Open the live version from the sources panel above.

And here's the tab itself — the engine badge reads the truth now: via OpenRouter · tencent/hy3:free · 295B · free.

The Agent OS Codex tab with the HY3 engine badge showing via OpenRouter tencent/hy3:free 295B free

The Codex tab in my Agent OS — 🐉 HY3 295B (free) is the default engine while the window is open.

"Is this actually the real HY3, or some proxy trick?"

Real. The request goes to OpenRouter's API with my key, model tencent/hy3:free, and OpenRouter returns the token usage on every call — the test build used 57,196 input and 7,864 output tokens, metered at $0. You can watch the same numbers stream back in your own terminal.

VIII ────── The exact setup

Wire it up — 5 minutes, 1 file.

You need: Codex installed (npm install -g @openai/codex) and a free OpenRouter account with an API key.

Step 1 — the profile file. Create ~/.codex/hy3.config.toml:

model = "tencent/hy3:free"
model_provider = "openrouter"

[model_providers.openrouter]
name = "OpenRouter"
base_url = "https://openrouter.ai/api/v1"
env_key = "OPENROUTER_API_KEY"
wire_api = "responses"

Step 2 — the key. Put export OPENROUTER_API_KEY=sk-or-... in your shell profile.

Step 3 — run it.

codex exec --profile hy3 "build me a landing page"

Three gotchas I hit so you don't:

Gotcha 1 — profiles are separate FILES now. Codex 0.143+ wants ~/.codex/hy3.config.toml. Putting [profiles.hy3] inside your main config.toml errors with a "legacy profile" message.

Gotcha 2 — wire_api must be "responses". New Codex dropped the old chat protocol. Set "chat" and it refuses to start. OpenRouter speaks responses fine.

Gotcha 3 — a broken install looks like a config error. My codex spawn failed with ENOENT on a missing native binary — nothing to do with the profile. npm install -g @openai/codex fixed it in ten seconds.

"I'm not technical. Can I really do this?"

One text file, one environment variable, one command. Copy the blocks above exactly. And inside the Boardroom, the Agent OS ships with this pre-wired — the Codex tab's engine dropdown, no terminal needed at all.

IX ────── The framework

The Free Codex Engine™.

Three parts. Each one removes a cost.

The Hands

You stop doing the typing — Codex plans, writes files and runs commands on your machine. The agent itself was always free; the meter was the model behind it.

ii.

The Brain Swap

You stop paying per thought — one profile file redirects Codex from OpenAI's meter to tencent/hy3:free on OpenRouter. 295B parameters, 262K context, $0 while the window's open.

iii.

The Cockpit

You stop juggling terminals — the Agent OS Codex tab carries an engine picker (🐉 HY3 by default), streams the build, and previews every file it writes in the Workspace tab.

One set of hands, three brains — the dropdown picks who thinks, and the free dragon is the default.

"Doesn't running Agent OS burn a fortune in tokens?"

No — that's the biggest myth about it, and this guide kills it twice. Agent OS runs the everyday 90% on free local models, catches free windows like HY3 for the heavy agent work — 295B at $0 — and saves the CLIs you already pay for for the few jobs that truly need them. Add the Caveman token-diet and even those paid replies shrink 60–75%. Full token-optimisation tutorials are inside the AI Profit Boardroom.

X ────── Old way vs new way

Old way vs new way.

Old way every task metered

Codex runs on paid models only — each loop costs real money
You ration experiments because the meter is always on
Free models can't hold an agent loop, so you never try
Testing a new model means new keys, new tools, new billing
Long 100K-token contexts are exactly the expensive runs
You read about free windows on X after they close

New way $0 by default

One profile file — Codex's brain becomes a free 295B
Engine dropdown in the OS: 🐉 HY3 free is the default
Real agentic builds, tested: full landing page, one shot
262K context at zero cost — feed it whole codebases
Your OpenAI setup untouched; switch back any time
Next free window = change one model slug, done

Real OpenRouter pricing — the free tier costs nothing, and even after the window HY3 stays ~60× cheaper than frontier output.

"Why not just keep using OmniRoute — it's already free?"

Keep both — that's why it's a dropdown. OmniRoute is a local gateway over 90+ small free providers: great for quick jobs, but the models are smaller and the reliable ones rotate. HY3 is a single 295B frontier-class brain with 262K context — a different weight class for agent loops and big codebases. While the window's open, the dragon is the better default; OmniRoute stays one click away.

"What happens when the free window closes?"

Three easy exits. Keep HY3 paid — it's cheap ($0.20/$0.80 per million, roughly 60× less than the big names). Switch the slug to the next free model — the profile pattern stays identical. Or just select OmniRoute in the dropdown and keep building free on the local gateway. Nothing about your setup breaks either way.

Get the whole operating system

Want this pre-wired, dropdown and all?

The Free Codex Engine is one dropdown inside the Agent Operating System — the dashboard I run my whole business on. Join the AI Profit Boardroom and you get everything:

Codex pre-wired — the engine picker, HY3 free default, workspace previews of everything it builds

Every free window — HY3 in Hermes AND Codex today; the next drop added the week it ships

Free local models — the everyday 90% of work at $0 on your own machine

The Caveman Command Engine — 60–75% off every paid reply

Every CLI you already pay for — Claude, Codex, Gemini, Kimi, GLM, Grok in one dashboard

Hermes Astros — the 24/7 YouTube competitor watcher that writes your titles

The Hermes Oracle — its sibling that watches X and drafts your posts

The Video Director + Agent Kanban — topic in, finished content out

The memory vault — an Obsidian brain your agents actually read

4,000+ founders + me — daily tutorials, weekly calls, someone online 24/7

You're not buying a tool. You're getting the whole operating system I run a seven-figure business on — as a zip, with coaching calls where we set it up together.

Get the Agent OS →

Inside the AI Profit Boardroom · skool.com/ai-profit-lab

Set up in an afternoon · used in 38 countries · new tools added every week

XI ────── Three beliefs to drop

What's holding you back.

Wrong: "Codex only works with OpenAI models."

Right: Codex has custom model providers built into its config — one file points it at any OpenAI-compatible API. Same hands, any brain. That's the most under-used feature in the whole tool.

Wrong: "A free model will fall apart on real agent work."

Right: HY3 was built FOR agent work — and it one-shotted a complete landing page through Codex on my first test, tool calls and all. Free right now doesn't mean weak; it means a launch window on a strong one.

Wrong: "Switching models means rebuilding my setup."

Right: Profiles are isolated files. Plain codex = your normal setup, --profile hy3 = the free dragon, and in the OS it's a dropdown. Nothing gets rebuilt, ever.

Don't take my word for it

158 pages of members who already broke through these exact beliefs. Their stories — real businesses, real wins — are documented here.

Read the 158-page testimonials doc →

XII ────── The SOP

Run it yourself — 8 steps.

Install Codex. npm install -g @openai/codex — also the fix if your existing install throws ENOENT.

Get an OpenRouter key. Free account at openrouter.ai → Keys → create. Export it as OPENROUTER_API_KEY.

Create the profile file. ~/.codex/hy3.config.toml — copy the block from the setup section, exactly. Separate file, not inside config.toml.

Smoke test. codex exec --profile hy3 --skip-git-repo-check "Reply with exactly: READY".

Give it a real build. A landing page, a script, a refactor — watch the token usage stream back metered at $0.

Use the giant context. 262K tokens — point it at a whole codebase and ask for the map.

Keep receipts. Note what it nails vs your paid model. Evidence now beats reviews later.

When the window closes, choose. Keep it paid (it's ~60× cheaper than the big names), swap the slug to the next free model, or flip the OS dropdown back to OmniRoute. One line either way.

XIII ────── Recap

What you gain.

You stopped paying per thought. Codex now runs on a free 295B brain — every loop, every retry, every experiment at $0.

You kept the hands you love. Same Codex, same tools, same workflow — only the address behind it changed.

You saw it work for real. "Build out a beautiful landing page" — one shot, 11KB file, receipts and token counts above.

You got the dropdown. In the Agent OS, 🐉 HY3 is one click — and it's the default while the window's open.

You dodged three gotchas. Profile-file format, wire_api "responses", and the ENOENT reinstall — solved here.

You kept your OpenAI setup. Profiles are isolated — plain codex is untouched, and removal is deleting one file.

You learned the reusable trick. Every future free window is the same file with a different slug. The engine outlives the window.

You experiment freely now. Ten ideas cost what one used to: nothing.

Your move

Free windows reward the already-built.

HY3 went free and I had it driving Codex inside my OS the same morning — because the machine room already existed: the engine picker, the profiles, the free local models, the Caveman token-diet, Astros watching YouTube, the Oracle watching X, the Video Director and Kanban teams shipping around the clock. That's what the Agent OS is: the room where every new tool plugs in the week it ships. You get it as a zip file, with coaching calls where we set it up together, step by step. Daily tutorials. A 30-day roadmap. 4,000+ founders across 38 countries, someone online whenever you get stuck.

Join the AI Profit Boardroom →

Inside the AI Profit Boardroom · skool.com/ai-profit-lab

Keep every file if you leave · 7-day refund · the only way you lose is by not trying

I'll see you in the next one.