SEO With Claude Projects: One Connector, Every Thread

Anthropic redesigned Claude Projects on September 17, 2026. A project is now one conversation where Claude coordinates a stream of work for you, starting a thread for each task. Each thread is a Claude Code cloud session that keeps running after you close the laptop. The launch post is about shipping code. It says nothing about SEO.

One line in the docs decides what it means for SEO, and it is not in the launch post. Threads run in the cloud, so they do not have the MCP servers installed on your machine. The SEO server you added to Claude Code last month is invisible to every thread in a project. The fix is a connector on your claude.ai account, set once. After that, every thread has live keyword and SERP data with no setup at all, and one Google consent screen adds your Search Console.

This post is that setup, step by step. It also covers the two parts of the redesign that matter more for SEO than the parallel threads everyone is writing about. Those are project memory and routines. And it names the defaults that will drain a Pro plan if you leave them alone.

What changed in Claude Projects

Two different things are called Projects right now, and most of the advice online is about the wrong one.

The older Projects live in claude.ai chat and Cowork. They are folders. You group conversations, attach reference files as project knowledge, add custom instructions, and every chat in the folder starts with that context. No threads, no coordinator. Those projects keep working exactly as they do today. Anthropic’s own words in the announcement are that existing projects on Pro and Max plans keep working as they do today, and will be upgraded as the rollout expands to chat and Cowork.

The redesigned Projects live at claude.ai/code, in the Code tab of the desktop app, and in the mobile app. Not in the terminal CLI. The announcement puts it in one sentence. Projects have threads that do the work and a coordinator that directs them. The Claude Code documentation fills in the rest. You paste a task into the conversation. Claude scopes it, starts a thread or hands it to the thread already working in that area, reviews what comes back, and assembles the result. Each thread is a cloud session with its own sandbox, and every thread starts with the project’s instructions, memory and files.

Three parts are new, and all three matter for SEO.

• Threads: cloud sessions that run in parallel and keep going after you close the laptop. You can check on them from your phone.
• Memory: notes Claude keeps about the project, requirements, decisions and pitfalls, stored as files that every new thread reads on start.
• Library: a tab on the project’s Overview pane that collects the files you added and the files threads produced.

Who gets it. Public beta on Pro and Max, rolling out gradually, starting with accounts that have already used cloud sessions and have no existing projects in chat or Cowork. Team and Enterprise are not included yet. If Projects is not in your sidebar at claude.ai/code, the rollout has not reached your account, and there is a waitlist. The announcement also says running threads on your own machine, alongside local tools and behind your network, is coming very soon. Until it does, the cloud is the only place a thread runs, and that is the whole reason the next section exists.

I pulled the live top ten for this exact query the day the redesign shipped. About 132 results. A Reddit thread first, then a GitHub skill, a Substack post, three YouTube videos and one Search Engine Land guide about Claude Code. One YouTube video names Projects in its title. Not one title on that page mentions threads or a coordinator.

Why threads cannot see your SEO server

The documentation is blunt about it. Threads are cloud sessions, so they do not have the skills, MCP servers, plugins and tools installed only on your machine. A separate line says threads do not pick up anything from the Claude Code setup on your own machine.

So the SEO server you added with claude mcp add is not there. The npx server running on your laptop is not there. The docs even list the kind of work that still needs a local session, a local database, a device emulator, an API behind your VPN. A keyword server listening on localhost belongs on that list.

Where threads do get MCP from is the connectors on your claude.ai account. You connect a server once, at claude.ai/customize/connectors or through the Manage connectors link in Project settings under Environment, and every thread in every project can use it with no per-project setup. In a project with a single repository, threads also load whatever that repository’s .mcp.json declares, which is a second route for a hosted server if your site lives in a repo.

Three consequences fall out of that.

The server has to be hosted. Public HTTPS, reachable from Anthropic’s cloud. Any hosted MCP server qualifies, ours or anyone else’s. A server that only exists on your machine does not, whatever it is.

You connect once, on the account, not per project. A second project for a second site needs nothing new. Turning a connector off inside a running thread removes it from that thread and from threads started later, until you turn it back on.

The project conversation has no connectors. This is the one that catches people. Ask for keyword volumes in the project conversation and you get an answer with no data behind it. Send the same request as a task and the thread that picks it up has every tool. The rule is simple. If it needs data, it is a task.

That is the same rule the Grok Bot cloud agent taught a few weeks ago, and SEO with Grok Bot walks through the same problem on that runtime. The difference here is that Anthropic gives you a connectors page, so there is no chat command to remember.

The setup: an SEO project with no repository

Most people will point a new project at their website’s repository, because that is what the launch demo does. For SEO, start without one. The docs say a project does not need a repository at all. Its threads still research, write documents, and deliver files to the Library tab. That is the shape of most SEO work.

Step 1. Create the project. New project, name it after the site, skip the repository. If your site is in a GitHub repository and you want threads to fix title tags and open pull requests, add it. The pull request becomes your review step, which is a good place for one. If your site lives in a CMS, leave the repository off and keep the project research-only.

Step 2. Write the project instructions. Project settings, then Memory, then Project instructions. Up to 16,000 characters, sent to every new thread and to the coordinator. This is the brief every thread starts from, so it carries your site facts and your rules. Something like this, with your own domains in it.

Project instructions

Site: example.com. Astro blog, US audience, English only.
Goal: grow organic clicks to /blog/ and /pricing/.
Competitors: rival-one.com, rival-two.com, rival-three.com.

Use the ContextBolt SEO connector for every keyword, SERP, audit,
backlink and Search Console number. Never estimate a search volume.
If a tool or connector you need is missing, say exactly what is
missing in your first message and stop. Do not substitute or guess.

Deliver every report as a markdown file in the Library, named
<job>-<date>.md. Never publish, edit a live page, open a pull
request or merge one unless I ask for it in the task.

The stop-and-say-so line is adapted from Anthropic’s own example brief, and it matters more in SEO than anywhere else. A model asked for search volume with no data source will produce a number that looks right. That instruction turns a missing connector into a message to you, instead of a keyword plan built on invented figures.

Step 3. Connect ContextBolt SEO. Open claude.ai/customize/connectors, or click Manage connectors in Project settings under Environment, which lands in the same place. Add a custom connector. Give it a name and the shared URL.

https://seo.contextbolt.app/mcp

Claude fills in the options under the URL and marks them Detected. Leave them alone. One is worth naming, because it is labeled Recommended and it does not work here. Claude offers “Use Anthropic’s hosted client metadata”. Our server does not support it, so keep whichever option Claude selected and press Add. Then sign in with your ContextBolt SEO account email, enter the code from the email, and approve the connection. If anything in that sign-in goes sideways, the connect your agent page has the recovery steps, and the Claude Desktop guide walks through the same dialog.

Search Console comes last, and it comes from inside a thread rather than the connectors page. Send a task asking the thread to connect Search Console, approve read-only access on Google’s consent screen, and every routine after that runs on your real clicks and positions.

Send the first task as a thread

Do not start with a keyword plan. Start with a task that proves the tools arrived, and make it deliver something to the Library so you can see the round trip.

Copy this prompt

New task. First, list the tools available through the ContextBolt SEO
connector. Then audit the ten most important pages on example.com.
Save the result to the Library as site-audit-2026-09-17.md, with the
issues that most affect indexing or clicks, each sized by the number
of pages it hits. Do not fix anything.

Claude starts a thread. Naming the page count matters, because a site audit costs a credit per page and runs 25 pages if you do not, so this one costs ten. Open its card in the conversation to read the transcript as it goes. If the tool list comes back empty, open the thread and pick Connectors from the plus menu beside its message box. The connector should be listed and switched on. If it is not there, the sign-in did not finish, and the fix is the connectors page, not the project.

Once the audit is in the Library, send the keyword pass, and make it stop before it drafts anything.

Copy this prompt

New task. Expand "ai bookmark manager" into 30 keyword candidates for
example.com, with US volume, difficulty and intent from ContextBolt
SEO. Pull the live top ten for the five strongest candidates. Shortlist
three targets with one line each on why, save the full table to the
Library as keywords-ai-bookmark-manager.md, and stop before drafting.

The shortlist is the decision, and it stays with you. The thread’s job is to make it a decision with evidence behind it. The workflow inside that thread is the same one keyword research with Claude walks through in a local session. Nothing about the research changes because it moved to the cloud. What changes is that you can send three of these and go to lunch.

Memory is the part worth the beta

Parallel threads are the headline of the redesign, and for SEO they are the wrong thing to be excited about. SEO work is serial. You cannot pick a target before the research comes back. You cannot write a brief before the target is picked. You cannot judge the page for weeks after it goes live. Ten threads at once solves a problem SEO does not have.

Memory solves the one it does have. Every session you have ever run on your site started from zero. The reason you dropped a keyword in June, the two pages that cannibalize each other, the query that sits at position 12 because the title is wrong and not because the page is weak. All of that lived in your head or in a document nobody opened.

Project memory is that document, and every thread reads it. The docs describe it as notes Claude keeps about the project, requirements, decisions and pitfalls, stored as files, with an index file called MEMORY.md that every thread reads when it starts. It is separate from the CLAUDE.md in a repository and separate from the memory Claude Code keeps on your own machine. Project settings under Memory lists the files, and you can read, edit or delete any of them.

You feed it by telling it. In the conversation or in any thread.

Copy this prompt

Remember: we are not targeting "best seo tools" again. The top ten is
all trade press and our page sat at position 41 for four months. Any
thread that proposes it should say so and skip it.

Do that after every decision and the project gets smarter in a way a single session never can. It also protects you from the way projects manage context. You do not manage context windows in a project. Threads compact automatically, and the conversation works from recent messages, recent threads and project memory rather than its full history. Anything that must never be dropped goes in memory or in the instructions, because the transcript itself will not hold it.

Routines for the weekly check

The Overview pane has three tabs beyond the thread list. Library, Pull requests, and Routines. Routines are scheduled work the project runs on its own, and the weekly Search Console check is the one to set up first, because it costs nothing. Search Console calls on ContextBolt SEO are free. Keyword and SERP calls cost a credit each, and a site audit costs a credit per page, 25 pages unless you name a smaller number.

Copy this prompt

Every Monday at 9:00 AM, start a thread that compares the last 7 days
in Search Console with the 7 days before. Report only the pages and
queries that lost clicks and had at least 100 impressions, with the
position change for each. Save it to the Library as
gsc-week-<date>.md. Change nothing on the site.

What you do with that list is the work in striking distance keywords with Claude. Split the click jobs from the ranking jobs, then send each one as a task. The routine finds them. You decide which ones get a thread.

What a project costs you in plan limits

A project draws on the same plan limits as your other Claude Code sessions, and it uses them faster. On a Pro plan in particular, the docs say to expect to hit your limit sooner on days you run one. Four things are worth knowing before the first task.

• Thread model and effort. A new project runs every thread on Opus at high effort, which draws on your plan fastest. Project settings under General has a thread model and a thread effort setting, and a separate pair for the coordinator. A keyword expansion or a Search Console comparison does not need the largest model at the highest effort.
• How many threads at once. There is no fixed number. Claude starts as many as the work calls for, and the enforced cap is 200 new threads a day across your projects. Ask it to propose threads before starting them and to run a few at a time, and it saves that preference to memory.
• What happens at the limit. A thread you started that hits your plan’s limit shows Service is busy, keeps retrying, and continues on its own when the limit resets, so a job you thought had failed at 6pm picks up again in the next usage window. Click Stop in the thread if you would rather it did not spend the next one. A thread a routine started does not wait. Its turn stops with a limit error and you send it a message after the reset, so a Monday routine that hits the limit needs a nudge from you.
• The idle thread. A follow-up sent to a thread that has sat idle for more than an hour re-reads that thread’s whole conversation before doing anything. For new work, start a new thread.

ContextBolt SEO credits are a separate meter and a smaller one. The $35 plan carries 1,000 credits a month, the 7-day trial carries 100, and Search Console calls are free. The keyword pass above is one credit for the expansion and one for each of the five SERPs, six in total if the thread does it the direct way. The ten-page audit is ten. Name a page count on every audit, and the plan limit is the meter to watch.

Threads research, you publish

Threads run in auto mode when the thread’s model supports it, which means most tool calls run without asking you. That is what makes a thread useful for research and what makes it dangerous for a live title tag.

The pitch for handing SEO to an agent is that it runs without you. Ours is the opposite, because search feedback takes months. A page changed on Monday tells you nothing until October, so a thread that publishes on its own leaves you blind for a quarter with no idea which of its changes did the damage. That is why the last line of the project instructions above says never publish unless asked, and why it is doing more work than any other line in the file.

Two ways to keep the line honest. If your site is in a repository, let threads open pull requests and merge them yourself, and put that in the instructions in so many words, because a thread in auto mode can merge if nothing tells it not to. Anthropic’s own example brief carries the line, do not merge, force-push or change CI configuration without asking me in the thread. The thread’s card shows a button for the pull request’s next step, Fix CI, Address comments, Merge it, and each one is a message you send. The button is a convenience. The instruction is the gate. If your site is in a CMS, do not put the CMS credentials in the project’s environment at all. A thread cannot publish through a login it does not have. SEO automation with AI agents has the longer list of jobs worth handing over and the three you never do.

Where ContextBolt SEO fits

A project needs one data connector, not five. ContextBolt SEO covers the five jobs a thread will be asked to do. Research, which is what to target. Audit and fix, which is what is broken. Authority, which is backlinks. AI visibility, which is whether ChatGPT and the other engines name you. And your real data, which is Search Console. You do not pick a tool. The thread asks in plain English and the right one fires.

It is hosted, so it is the shape a cloud thread can reach, with nothing to install and nothing on your machine to keep running. The numbers are estimates from a large keyword index, decision-useful and directionally right, and every result saves to your SEO Board on its own, so the thread’s transcript is never the only copy of the research. ContextBolt is ours, so weigh the bias, and then run the audit task above on your own site during the trial. It is free for 7 days, then $35 a month.