SEO Inside Muse Code: Audit, Fix, Re-Check in One Session

Muse Code is Meta’s terminal coding agent. One command installs it, muse starts it inside a project, and it plans, edits, and checks changes across a large repo on Muse Spark 1.3, Meta’s flagship model with a 1M-token context window. It is the Meta answer to Claude Code and Codex, and it speaks MCP.

That last part is what makes this post possible. MCP, the Model Context Protocol, is the standard Anthropic released in 2024 so agents could call outside tools through one plug. Muse Code reads its MCP servers from a settings file at startup. Point one of them at a live SEO data source and the keyword question that lands while you are editing a page gets answered in the same window.

This guide covers the settings block, the login, the one setting that will bite you, and the loop only a coding agent can close. We run it on our own site, so this is the workflow as it actually runs.

Why pull SEO data into Muse Code

The case is not that Muse Code is now an SEO tool. It is that the cost of asking an SEO question should be near zero, and in the usual setup it is not.

You are in a session, Muse Code has just drafted a docs page, and you want to know whether anyone searches for the phrase it put in the H1. The old answer was to break flow, open a browser, log in to a dashboard, run one lookup, and come back. Most people guess instead, and guessing is how a page ends up optimized for a phrase nobody types.

Muse Code changes the second half of that too. A dashboard shows you a missing meta description. Muse Code can open the file, write one, and run the site’s checks, because editing the repo is its whole job. Meta’s product page pitches it for the most complex coding workstreams, which is a marketing way of saying it will happily touch forty pages in one go. That is a gift for an audit and a risk for a live site, and the rest of this post is about keeping both.

If you want the longer argument for why asking the agent beats opening another SEO suite, the SEO MCP server explainer makes it. The short version is that an answer you can act on in the same window is worth more than a prettier chart two tabs away.

How do you connect SEO data to Muse Code?

Muse Code has no add command for MCP servers. You edit the settings file, then sign in. Meta’s extending and automating docs are the source for the field names below.

With an active ContextBolt SEO trial or subscription, open ~/.config/muse/settings.json and add the server. If the file is new, schema_version has to be there, or every muse command fails at startup with a malformed settings file.

{
  "schema_version": 1,
  "mcp_servers": {
    "contextbolt-seo": {
      "transport": "streamable_http",
      "url": "https://seo.contextbolt.app/mcp",
      "enabled": true,
      "mode": "optional"
    }
  }
}

Two fields deserve a sentence each.

transport: streamable_http is for a remote server reached by URL. stdio is for a local process you start with a command. ContextBolt SEO is hosted, so it is the first one, and there is no framing field on an HTTP server. Adding one fails validation.

mode: Muse Code defaults to required, and a required server that fails to start aborts the whole run. That is the right default for a database your task cannot proceed without. It is the wrong default for a research tool. Set it to optional, so a blip on the network gets you a warning and a working session instead of no session.

Then sign in. The shared URL has no secret in it; your account is connected by OAuth.

muse mcp login contextbolt-seo

The browser opens the ContextBolt sign-in page. Check the address is on seo.contextbolt.app, enter your SEO account email, enter the code, and choose Allow connection. muse mcp logout contextbolt-seo disconnects it later.

Start muse in any project and type /mcp. It shows the connected servers and the tools each one gives the session. You want keyword research, keyword difficulty, SERP overview, domain overview, the audit tools, and the backlink tools. If the server is missing, the login did not finish. The connection guide covers reauthentication and the personal-URL compatibility option, which is a URL with a token in it, for a machine that cannot open a browser.

If you have not installed Muse Code yet, the install is one command on macOS and Linux, curl -fsSL https://dev.meta.ai/install.sh | bash, with a PowerShell version at install.ps1 for Windows, then muse login. The shorter, step-by-step version of the whole setup is on the Muse Code integration page.

The audit, fix, re-check loop

This is the workflow that only makes sense inside a coding agent, and it is the reason to do SEO here rather than in a chat window.

Say “audit our site” and ContextBolt SEO finds the sitemap, picks the most important pages itself, homepage and section and product pages first, and grades their mechanical health, title and meta lengths, one H1 per page, thin content, broken links, canonical tags, then hands back a 0 to 100 score with the fixes worst first. The audit is metered per page, 25 pages by default and 50 at most, one credit each, so the first run is a bounded spend rather than a whole-site crawl. For one section, ask for a page audit on each URL instead.

Then the part a dashboard cannot do. Muse Code opens the flagged files, makes the changes, and runs the site’s own checks. Ask it to re-audit and it re-checks the same pages and leads with the delta. Health 78 to 91, four pages still flagged, here they are.

Copy this prompt

Use ContextBolt SEO to audit our site at its default 25 pages. Group the
findings by cause, not by page. For each cause, tell me how many pages
it touches and what the fix is.

Then fix only the title and meta description issues in this repo. Do
not touch body copy. Run the project's checks. Stop before committing
and show me the diff.

The stops are deliberate. Titles and descriptions are safe to batch. Body copy is not, and neither is a commit you have not read. A coding agent that can touch forty pages will, if you let it, and search feedback takes long enough that you will not know it was wrong until the quarter is over.

Once the diff is in and shipped, the re-check can be a one-liner from the shell. muse exec runs a single prompt without opening a session.

muse exec "Re-audit our site with ContextBolt SEO and lead with what changed."

That is the loop. Find it, fix it in the repo, prove it moved. The site audit with Claude post walks the same loop with the triage rules in more detail, and they transfer unchanged.

What else you can ask once it is connected

The audit is the headline. The everyday use is smaller questions, asked in the seconds you would otherwise spend guessing.

• Keyword reality checks: “Is anyone searching for the phrase in this H1, and what would they type instead?” Muse Code runs keyword research and returns related terms with volume, difficulty, and intent, before the page is written.
• Difficulty plus SERP in one ask: “How hard is ‘rust web framework’ to rank for, and who is in the top 10?” It chains a difficulty score and the live page-one results, so you learn whether you can win and what kind of page already wins.
• Domain and competitor mining: “What does vercel.com rank for, and who else competes for those terms?” A content map plus the rivals you would not have named, run from the shell you deploy from.
• Your own numbers: “Which docs pages lost clicks in the last 28 days compared with the 28 before?” That reads your connected Search Console, and Search Console calls never spend a credit.
• Backlink gaps: “Who links to my top competitor but not to me?” A spam-scored prospect list. Backlink questions cost three credits each because the data behind them is heavier.
• Chained research: “Find ten low-difficulty keywords for the docs, check which ones a competitor already ranks for, and list the gaps.” One prompt, and the whole chain lands on your SEO Dashboard as it runs.

The thread through all of these is plain language. You are not learning a tool’s menu. You ask the way you would ask a colleague, the agent picks the tool, and then the next one.

Where ContextBolt SEO fits

ContextBolt SEO is ours, so weigh the bias. It is a hosted MCP server that wraps a large keyword index and hands you one URL. You add the URL to settings.json, sign in once, and the SEO tools are in the terminal. Research tools cost one credit each. Backlink tools cost three. The plan is $35 a month for 1,000 credits, with a 7-day free trial, and there is no account to provision or deposit to fund. The Search Console tools and the automatic memory never spend a credit.

Every lookup also lands on your SEO Dashboard, one card per keyword or domain, dated, so next month’s question starts from what you already know. Ask about the same keyword a week later and the answer leads with what changed, a volume or difficulty delta, with no extra prompt and no extra credit. Most agents lose their research when the session ends. Here it accretes into something you own, in a dashboard you can open from any machine.

Volume and difficulty are estimates, decision-useful and directionally right, the same class of data inside the dashboards you already trust. For deciding what to build and what to write, that is more than enough. Use Search Console, which is your real data, for how it is actually going.

Is it worth it?

Four things to know going in, and then the answer.

MCP tools are not sandboxed. Meta’s docs say it plainly. Muse Code sandboxes shell commands. An MCP server runs outside that sandbox, as a network connection, so connect servers you trust. ContextBolt SEO’s data tools are read-only lookups against your own credit balance, and nothing in it writes to your site or your accounts.

Read the pricing label. Meta’s model page lists a cheaper contributor model with the label “Used to improve our products” and the standard one as “Not used to improve our products.” If you are paying per token on the contributor price, do not route a client’s Search Console data through it. On the monthly plans, Everyday at $5, High at $15, Power at $50, check which model you are on before an audit that reads real traffic.

Scheduled rank tracking is a different job. Checks here run when you ask, and the Search Console connection takes a daily snapshot of your own clicks and positions. If you think in charts, the terminal will feel sparse at first. The SEO Dashboard is where the charts live.

The agent can be confidently wrong about strategy even when the numbers are right. Treat its SEO advice the way you already treat its code. Useful, fast, and in need of a human read before it ships.

What you gain is the thing that matters when you build solo. The answer arrives where you already are, and the fix happens in the same session, at a price that does not make you wince. The keyword you were nervous about might be wide open. Start the 7-day trial, add the block above, run muse mcp login contextbolt-seo, and find out without leaving the session.