You can now run a full SEO content system for free.
The harness is free, and the brain can be free too.
It shows you every single step it took.
So you can check the work in thirty seconds instead of thirty minutes.
Everything on this page is running on my machine right now.
150413GitHub stars
MITlicence
38plugins loaded
$0harness + brain
§2What actually shipped
Released 13 August. MIT licensed. Growing fast.
MIT means you can download it, change it, even build a product on top of it without asking anyone.
Live figures pulled from the GitHub API while this page was built — the repo went public on 13 August 2026.
But none of that is why you should care.
§3Where this is going
Three levels of running SEO on this thing.
Setup. Then the checking system no other harness has. Then scale.
§4What's actually new
The other harnesses are sealed boxes. This one is all plugins.
You could always swap the brain. You could never swap the body.
The sealed boxsomeone else decides
The interface is fixed
The rules are fixed
The sidebar is theirs
The core loop is theirs
You wait for their roadmap
All plugins38 loaded here
The sidebar is a plugin
The chat window is a plugin
The tools are plugins
The core loop is a plugin
Swap any piece, the rest keeps working
What you're watching: the 38 plugins my copy actually loads on boot — the sidebar, the chat, the trajectory panel, even the core loop, each one a separate swappable piece.
pulled live from localhost:3080§5The one line that matters
It reads agents.md and CLAUDE.md.
Those are instruction files — little text files that tell an agent about your business. If you build with Claude Code, you already have them.
What you're watching: a real session in my SEO folder. Step three, it opens CLAUDE.md on its own — nobody told it to.
It slots into what you already have.
§6Level one · the setup
My SEO system runs on skill files.
A skill file is a recipe the AI follows step by step, every time, without you repeating yourself. Mine is a 13-step quality control file.
What you're watching: the whole file scrolling past — 145 lines covering how articles get structured, what language to use, and what gets checked before anything is published.
step 0 · read before you writestep 1 · claims with sourcesstep 5 · structure firststep 6 · 3rd grade reading levelstep 9 · schema markupstep 12 · the grounding audit
§7The system underneath
Record the screen. Transcribe it. Write only from that.
Every claim in the article traces back to something that actually happened on screen.
§8The problem
Google doesn't penalise AI content. It penalises thin, generic content.
Grounded content — built from real work, real data, real experience — ranks. Generic AI slop doesn't. That's the whole difference.
§9The workspace
One folder. That's the whole setup.
It can read every file in that folder, create files, change files — and it can't touch anything outside it.
What you're watching: my real workspaces in the sidebar — the SEO folder sitting next to the others, with its sessions underneath it.
§10The brain
Free harness. Free brain. Full system, no cost.
Settings, models, add provider. Connect OpenCode — and one of its free models is DeepSeek V4 Flash.
What you're watching: Settings → Models on my machine, with the extra provider already added underneath the DeepSeek one.
What you're watching: the model picker opening. Under "OpenCode Zen" there it is — DeepSeek V4 Flash Free. One click and the whole harness runs on it.
§11Which brain does what
SEO has thinking work and volume work.
Pro handles the thinking for pennies. Flash handles the volume for nothing.
What you're watching: the free model doing a real volume job on my article — and catching that the schema wouldn't match the visible page.
Most people pay for tools that do a fraction of this.
§12What it looks like day to day
Point at the folder. Ask for the article. Walk away.
Come back later and the draft is sitting in your folder.
What you're watching: the real run from my machine. It read the skill file, then all seven transcripts in full, and only then started writing.
1,772
words in the finished draft
7 of 7
transcripts opened before a word was written
0
claims it couldn't trace to a transcript
THINKING IT?"So it just makes things up like every other AI writer."
It dropped a figure from its own draft because the same transcript contradicted itself, and said so in the audit.
That's the skill file doing its job — and you can see it happen in the log.
Every step the agent took, in order, clickable. The files it opened. Its actual reasoning. Every tool it called. Nothing hidden.
What you're watching: the trajectory panel from my real article session — the timing waterfall up top, then every step, switching between Duration, Turns and Calls.
§15Why this matters
Every agent makes mistakes. The question is whether you can see why.
With most tools a failed job gives you a wrong output and no explanation.
Black boxgood luck
Wrong output, no reason given
You re-run it and hope
You rewrite the prompt blindly
You never learn what broke
Same mistake next week
Trajectory viewone scroll
Oh — it looked in the wrong folder
Oh — it skipped transcript three
Oh — the file it needed was empty
You fix that one thing
You run it again
§16The 30-second check
Did it actually read the transcripts?
Open the panel and look at which files it opened before writing. That's the whole check.
What you're watching: clicking the tool rows in my real session. Every transcript, opened in full, before a single word got written.
That's the difference between AI content you can trust on a real website, and AI content you're gambling your rankings on.
§17One more layer
Export any session as a complete log.
Every step, every call, every result, in a standard file. Your reviewer checks the log instead of redoing the research.
What you're watching: hitting Session log on my article session. It really downloads — 1.38MB of every event in that run.
THINKING IT?"My team will never look at a log file."
They don't need to read it line by line. They open it, check the article was built from the sources, and approve.
You review the work the way a manager reviews a report.
§18Why this is the differentiator
No other harness gives you this much visibility.
One bad claim in a published article sits there hurting you for months. Seeing every step is worth more than any single feature.
§19The honest answer
The install is one command. After that it's a web page.
Chat window, sidebar, folder picker. It looks like software you already use.
What you're watching: the real install on this machine — one command, then it opens in the browser.
Wrong: "This is developer software. I'd need to learn to code first."
Right: The people running this in my community are agency owners, freelancers and ecommerce sellers. A lot of them had never opened a terminal before.
Wrong: "I need to understand how the model works to get anything out of it."
Right: If you can organise a filing cabinet and write a clear message to an assistant, you can run this.
Wrong: "The skill of the next few years is learning to code."
Right: It's writing clear briefs. What to look at, what to produce, what rules to follow. That's it.
Don't take my word for it
Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries. Real businesses, real numbers, in their own words.
One drafting, one doing keyword research, one fixing internal links — all reading the same folder, all following the same skill file.
§21Creator mode
Describe a tool you wish existed. It builds it in the session.
Whatever you check manually right now, you can describe — and it becomes a panel on your screen.
What you're watching: I asked for a panel showing my recent articles and whether each one is indexed yet. It loads the plugin-development skill, reads my drafts and my Search Console file, then goes looking for the API it needs to build the panel.
STRAIGHT UP"Did it finish and ship the panel?"
On my machine today, two creator-mode runs got as far as inspecting the API and then stalled before writing the plugin.
That's this being version 0.1 — the same caveat as the rest of the page. The article writing and the trajectory checking are the parts that are genuinely ready today.
You didn't write a line of code. You didn't wait for a developer.
§22The full loop
The GOLDIE Loop.
Generate, Optimise, Launch, Distribute, Index, Evolve — and here's how the harness covers each piece.
§23Evolve · the underrated one
The pages that fell are your fastest wins.
Drop your Search Console export in the folder and ask it to find what's sliding. Updating beats writing new.
What you're watching: a real run over my own Search Console export — 8,488 rows across 10 sites, pairing two periods to find what slipped.
§24The pipeline it plugs into
The same transcripts drive my whole SEO panel.
Research, transcripts, the skill, generate, deploy — the loop with a front door on it.
What you're watching: my live SEO pipeline — the transcript library, the skill it runs, the generate step, and the deploy history behind it.
§25Why it keeps working
Every agent reads and writes to one shared vault.
Obsidian is just a folder of plain text notes. That plainness is the superpower — plain text is the one thing every agent can read.
What you're watching: the first time DeepSeek Harness opened my vault. It came back knowing the business, the sites, the audience and the house style — and named the file each answer came from.
§26The feeling I know you have
A new tool isn't homework. It's a free upgrade.
Mastering individual tools was never the skill. Running one system that new tools plug into is.
Wrong: "There's a new AI tool every month. I can't keep up."
Right: DeepSeek Harness plugged into a year of my existing setup in an afternoon — because it was built to speak the languages the agent world already speaks.
§27One honest thing
This is version 0.1. A developer preview.
Things will break. A plugin that works today might break next month and get fixed the month after.
Don't do thisrisky
Wire your entire operation to it this week
Treat a 0.1 preview as finished software
Bet a client deadline on a preview feature
Delete the thing that already works
Start over when something breaks
Do this insteadcompounding
Run it on real SEO tasks — that part is ready
Keep it inside a system with shared memory
Keep your instruction files portable
When something breaks, swap one piece
Keep moving
That's the difference between people who compound with these tools and people who start over every few months.
§28Where this leaves you
The volume layer of SEO now costs nothing.
i.
You stopped paying for the boring half. Meta descriptions, title tags, alt text and schema across hundreds of pages run on a free model.
ii.
You stopped guessing whether it's grounded. Open the trajectory, see which files it read, approve in thirty seconds.
iii.
You stopped doing the loop by hand. Article to links to indexing to updates, running while you're doing something else.
The sites winning search over the next year are the ones publishing consistent, grounded, properly structured content at a pace humans can't match manually.
Want this running on your site without building it?
Book a free SEO strategy session with my team. You'll get a custom gameplan to rank your site and get more customers from Google and AI search engines.
The AI Profit Boardroom is where the actual Agent OS lives — the templates, the prompts, the daily rooms, the weekly walkthroughs. The full DeepSeek Harness SEO setup lives in there — the workspace, the skill files, the free models, and the Agent OS they plug into.
