On August 7 Anthropic announced that auto mode becomes the default in Claude Code. From August 14, sessions on Pro, Max and Team stop asking permission for every command. Anthropic's own number for what that changes: Claude works 9x longer between interruptions.
Anthropic's own announcement. From August 14 your agent stops asking permission.
I decided to make my Claude Code an SEO manager. It's excellent at finding keywords, researching, and drafting articles and interactive tools, if you give it the right access.
Google traffic compounds: one guide that ranks keeps sending people every month, for free, while you sleep.
That's the job I gave mine. Everything below is open source, so you can copy the whole thing completely for free.
The advantage
There's an advantage here nobody uses. Every SEO tool on the market learns about your product by crawling your homepage. Your agent doesn't have to. It already knows your product, because it probably built it. It has the codebase, the README, the real features.
It knows things your homepage doesn't even mention.
So instead of paying $99/month for a tool to badly reverse-engineer what your agent already knows, give the agent the job.
1. Hand it the job description
DispatchSEO (mine, and open source) is a fully self-hostable tool that lets your agent run your SEO: research keywords, check Google rankings, draft content, identify trends and AI citations.
What you need:
- Your site's source in a GitHub repo. Content ships as pull requests, so git-based sites only. Sorry WordPress.
- A machine with Docker and about 1 GB of RAM.
- Google Search Console access.
- Claude Code, on the subscription you already pay for.
Then:
-
docker compose up. Database, migrations, schedules and the headless builder are all in the box. - Run the setup wizard: connect Search Console and your repo.
- Paste one line into Claude Code. Your agent now has 61 tools over MCP: the queue, keywords, rankings, Search Console data, backlink prospects, trend scanning.
- Turn the schedule on.
The dashboard. Search Console is free, and your own impression data beats any seed list.
One honest note about "while you sleep": schedules only run while the machine is awake. A laptop is fine for a test drive. A $5 VPS or an old desktop that never sleeps is the real move.
On Codex or Cursor instead? Both work, overnight builder included. The backend doesn't care which agent turns up for the shift.
2. It researches from your own codebase
Every run, the agent re-reads the repo and derives keyword candidates from the product itself. It hunts commercial intent first: "X vs Y", "best X for Y", "X alternatives". Easy informational traffic converts at zero, so it isn't the target.
Survivors get validated for volume and difficulty, then the agent reads Google's actual page 1 to see whether the keyword is winnable. Everything lands in a queue with the reasoning attached, so you can see why it wants to write something before it writes it.
The queue, with the reasoning attached. Approve, reject, reorder, or set it to auto and walk away.
3. The night shift
Every morning it takes the oldest approved idea and opens a pull request. One guide a day maximum, however long the queue gets, because thirty approved ideas turning into thirty posts in a week is how sites get flagged.
A real PR my agent opened overnight. I reviewed it, merged it, live the same day.
It builds free tools too. Small utilities related to your product: a calculator, a checker, a generator. Nobody links to your blog post about keyword research. People link to a free tool that does keyword research. And "free X checker" searches convert, because the searcher has a job to finish right now.
Guides bring the traffic. Tools convert it, and they earn links while you sleep.
A free tool my agent built and shipped. People bookmark tools. Nobody bookmarks your 2,000-word guide.
The agent never touches the live site. It proposes, you merge. That's the whole safety model, and it's also why the output doesn't rot into slop. You're still the editor.
4. Three gates that stop it writing slop
The part I'd want to read the source of before trusting it:
- The SERP gate. It reads page 1 before writing a word. Whatever those results cover becomes the must-cover list. If the draft can't beat page 1, it refuses to build. An SEO tool that refuses to write.
- Sameness detection. Every guide gets checked against your own back catalogue: opening patterns, heading skeletons, stock phrases. Write 30 articles and they converge on one template, and that's the scaled-content fingerprint Google discounts. Surfer and Clearscope compare you to competitors, never to yourself.
- Information gain. Before it writes, it needs one thing no competitor page has. A command it actually ran, a real measurement, a screenshot from your dashboard.
Feed it what only you know: product knowledge, screenshots of your app, the words your customers actually use. That's the entire difference between content and slop.
What it costs
- Search Console only — $0. Rankings from GSC, keyword ideas from Autocomplete and your own impression data.
- Plus a free SerpApi key — still $0. Live page-1 checks, 250 free searches a month.
- Plus DataForSEO (optional, most accurate) — cents per call. Search volume, keyword difficulty, domain rating.
Free mode finds keywords you can win. Paid mode also tells you which ones are worth winning. The agent itself runs on the Claude subscription you already pay for, and one PR a night is a small slice of it.
Does it work? Here are my charts
My site is very young. Two months old, no domain rating, no bought links: 84 clicks and 4.3K impressions in the last 60 days.
My site's Search Console, last 60 days. 84 clicks, 4.3K impressions, and climbing.
Anyone who's done SEO knows how to read that. A two-month-old domain with no DR is usually still sitting at zero, and the standard advice is to expect nothing for the first six months. This one didn't wait.
Why should you listen to me?
I build DispatchSEO, and all of it runs in public. The repo is open source, the sites my agent manages are live, and the search data above is mine. Last week I posted this same chart one week younger. Next month I'll post it again, whichever way it went.
You don't have to trust me. You can read the commits.
github.com/NeoZi12/dispatchseo
Everything in this article runs from this repo. Read the commits.
If you'd rather not run a machine, there's a hosted version at dispatchseo.com. Self-hosting stays feature complete either way. The cloud sells convenience, not capability.
What job should your agent take over next? I'm collecting ideas for the next article. Tell me below 👇
Top comments (0)