- Intermediate
- 12 minutes
- No SEO knowledge needed
SEO with AI agents: connect Claude Code to Crawl Foundry
Connect Claude Code to a Crawl Foundry workspace over MCP. See which data its tools can reach, how the SEO skill guides tasks, how purchase limits work, and why publishing still needs your approval.
Let's startIn a nutshell
MCP connects tools. A skill describes a workflow.
After you connect Claude Code and authorize a Crawl Foundry workspace, MCP lets it call the tools exposed to that connection. Available data depends on the selected workspace and its connected sources: Search Console figures appear only when Search Console is connected and data is available. The SEO skill provides task instructions; it does not grant data access or replace the product's permission checks.
Why it matters: Try a read-only task first: ask for saved crawl findings on /pricing and a title suggestion, and ask the agent to name the selected workspace, tool and data date. A crawl finding does not prove what Google indexed. Review any suggestion against the live page before applying it.
First things first
What to look out for first
1.Trace every number
Do thisIf a tool returns a number, state its source, scope and date. If no value is available, say so instead of estimating.
Not thisGuessing search volume or positions from model memory.
2.Check the cost before fresh data
Do thisCheck saved data first. Before a paid research or content action, check its quoted cost and the workspace budget: an active budget may allow the charge without a separate prompt.
Not thisAssuming every paid action pauses for approval. If you want to block automatic purchases, set the budget to “No spending”.
3.Keep publishing approval with a person
Do thisReview titles, descriptions and links. Publishing requires explicit approval, even when research spending is within budget.
Not thisTreating a spending limit as permission to publish.
Step by step
How to set it up
1Step 1 of 4
Connect Claude Code to Crawl Foundry
Install Claude Code. In Crawl Foundry, open Agent, MCP & API and copy the connection command shown there. Run it in a terminal, then type /mcp in Claude Code and complete OAuth in your browser. Confirm the selected workspace and requested access. The command's --scope user setting makes the server available across your Claude Code projects; choose a different scope if that is too broad.
Done when: /mcp lists crawl_foundry as connected, and you have confirmed its workspace and requested access.
Simplified sketch with sample data 2Step 2 of 4
Install the skill
Download SKILL.md and put it in ~/.claude/skills/crawl-foundry-seo/. Claude Code uses the skill's name and description to decide when it fits, then loads its instructions when relevant. Use /crawl-foundry-seo to invoke it directly. To keep it in one repository, use that project's .claude/skills/ folder instead.
Done when: When you type a slash in Claude Code, /crawl-foundry-seo shows up in the list.
Simplified sketch with sample data 3Step 3 of 4
Start safely
Start with a read-only task. Ask Claude Code to select the right workspace, name it, and summarize stored SEO data without buying or changing anything. Then try: “For /pricing, show saved crawl evidence, name the data date and suggest a clearer title. Don't edit or publish.”
Done when: The agent identifies the workspace, names the tool and date for each figure, and says when data is unavailable.
Simplified sketch with sample data 4Step 4 of 4
Set the budget
The agent budget sets caps for paid research and content actions, per action, workspace and organization. An active budget can allow purchases within those limits without a separate approval each time. Set the workspace to “No spending” to block automatic purchases; saved-data reads remain available. Publishing always requires explicit approval, regardless of budget. Other account changes use separate permissions. Organization administrators control the budget.
Done when: The "Agent budget" tab shows the setting you want for your workspace.
Simplified sketch with sample data
Ready to copy
Every command for the setup
Three commands, in this order. Crawl Foundry also shows the connection command under "Agent, MCP & API".
1. Install Claude Code
Terminal · macOS, Linux, WSLcurl -fsSL https://claude.ai/install.sh | bashTerminal · Windows PowerShellirm https://claude.ai/install.ps1 | iexThen start Claude Code in your project folder with the command claude and sign in on first launch. Claude Code needs a Claude subscription or an Anthropic Console account.
2. Connect Crawl Foundry
Terminal · any systemclaude mcp add --transport http --scope user crawl_foundry "https://app.crawlfoundry.com/api/mcp"The command adds Crawl Foundry for all your projects. Then type /mcp in Claude Code and sign in. The commands for Codex, Cursor and VS Code are on the SEO MCP server page.
3. Install the skill
Terminal · macOS, Linux, WSLmkdir -p ~/.claude/skills/crawl-foundry-seo && curl -fsSL https://crawlfoundry.com/downloads/crawl-foundry-seo/SKILL.md -o ~/.claude/skills/crawl-foundry-seo/SKILL.mdTerminal · Windows PowerShellNew-Item -ItemType Directory -Force "$HOME\.claude\skills\crawl-foundry-seo" | Out-Null; irm https://crawlfoundry.com/downloads/crawl-foundry-seo/SKILL.md -OutFile "$HOME\.claude\skills\crawl-foundry-seo\SKILL.md"The command creates the folder and downloads SKILL.md into it. You can also save the file with the download button and put it there yourself.
Common mistakes
Mistakes that happen a lot at the start
Telling the agent nothing more than "do my SEO".
Better: Give it a concrete task with a URL and a goal, such as checking one page or suggesting titles for a category. The skill gives it the procedure.
Assuming every purchase waits for your approval.
Better: An active budget can permit automatic purchases within its limits. If you want to approve each purchase yourself, set the workspace to “No spending”.
Accepting suggestions without checking them.
Better: An agent with good data can still be wrong, especially with metadata. Read every suggestion before it goes live, and measure the result later.
Mini glossary
Terms you need here
- AI agent
- A program built on a language model that does more than answer: it uses tools to read files, run commands and query data. Claude Code is such an agent.
- MCP
- Model Context Protocol, an open standard that lets AI applications connect to external tools and data. Anthropic introduced it in November 2024.
- Skill
- A folder with a SKILL.md file: name, description and instructions for one kind of task. Since December 2025 the format has been an open standard that other tools read too.
- Agent budget
- Crawl Foundry's limit for paid agent actions. It can allow automatic spending within its configured caps; it does not authorize publishing.
Done
Connect your first agent to Crawl Foundry.
After you connect Claude Code and select a workspace, it can read data exposed through the available MCP tools for free. Paid research and content actions follow the configured budget and approval policy; publishing always requires your explicit approval.