Ask your AI to read a competitor’s site, page by page.
Any page read with a real browser and returned as clean markdown, a site crawled page by page up to a hundred pages, and search results with their content already fetched — answered inside ChatGPT, Claude, Codex or Cursor. No Firecrawl plan or API key of your own: Soku makes each call and quotes it first.
Claude CodeCodexCursorVS CodeChatGPTClaude (web & desktop)Your own agent
Soku Connect
Sign in, grab the setup prompt from your console, paste it to your agent, ask something. That’s the whole setup.
Soku research stack
traffic, SEO, app-store and ads intelligence
Firecrawl
Scrape or crawl any site into clean text
J
Jina Reader
Any web page as clean text, cheaply
ScreenshotOne
Screenshots of any page, full-length or mobile
200+
APIs
40+
apps
30+
skills
Ask once, let the agent choose the right source, and see the price before a paid data call runs.
No vendor keyQuoted per callOne research stack
About five minutes
Set it up once, in the app you already use.
There is nothing to install for Firecrawl specifically. One connection reaches every source in the catalogue, and this is one of them.
Claude Code
One sign-in covers every project, and headless runs (`claude -p`, CI, the Agent SDK) work without a browser — `claude mcp add` only writes config, so a non-interactive run can never finish the MCP authorization. Both surfaces can act on your accounts once you approve each change.
1
Create your Soku workspace
The connection is public and identical for everyone — your Soku sign-in is what points it at your data. Connect one account in that workspace too: that is what your AI reads. The first time your AI uses the connection, Soku asks you to approve the access and names the workspace it will read through — so the workspace has to exist by then.
Already have one? Nothing to do here — sign in when the connection asks for it.
2
Copy the Soku MCP connection
https://api.soku.ai/mcp
3
Add it to Claude Code
claude mcp add --transport http soku https://api.soku.ai/mcp
Run the command in any project. It writes the configuration and prints an `Added …` line.
Authorize separately — adding a server does not open a browser.
Run `claude mcp list` to confirm the server reads `✔ Connected`.
4
Approve the access, then ask it something
Inside Claude Code, run `/mcp` and follow the browser flow — or, from your shell, `claude mcp login soku`.
The Soku consent page names the host, the workspace the grant will read through, and what the connection will be able to do.
Approving creates the grant listed above on this page. Cancelling creates nothing.
On a machine with no browser (SSH, headless Linux), `claude mcp login soku --no-browser` prints the URL and takes the pasted callback.
Paste this first
Using Soku, search for what capabilities I can reach.
A list of capabilities with names, what each reads and what each costs. An empty list means the search matched nothing — try broader words. An error naming authorization means the host never completed the browser step.
Checked 2026-08-25 against Claude Code documentation. Menu names change between releases — look for connectors, apps, tools or integrations.
Ask in plain English. These are the tools behind the answer.
Each one is a tool your AI is offered by name once you connect. You never have to type the name — it is here so you can tell what actually ran.
One page, rendered and cleaned
firecrawl_scrape
Fetches a URL with a real browser and returns its content as markdown, HTML or structured fields — so it works on pages a plain fetch returns empty.
“Scrape this competitor’s pricing page into markdown.”
“Pull the plan names and prices out of this page.”
A whole site, page by page
firecrawl_crawl
Follows links from a starting URL across a site and collects up to a hundred pages of content. It runs as a job; your AI reads the pages back when the crawl is done.
“Crawl this documentation site.”
“Get the pages under this competitor’s blog, up to a hundred of them.”
Search results with their content
firecrawl_search
Runs a web search and scrapes the results in the same request, so the answer comes back as content rather than links.
“Search for MCP marketing tools and read their pages.”
“Find reviews of this product and pull what they say.”
And 1 more, same connection
Named rather than worked through — your AI discovers all of them through the connection itself, so you never have to pick one.
Read the pages a crawl has collectedfirecrawl_crawl_status
Pricing
Choose the Soku plan for your research and agent workflow.
Start free, upgrade when your team needs more credits, more brands, or deeper workflow execution.