An agent with a shell can already run curl. So why bother with a tool server?
Because curl returns HTML, and HTML is a bad input for a language model. It is long, noisy and mostly layout. The agent spends its context reading navigation menus, and the answer you wanted is somewhere on line 900.
What the MCP server changes
The AgentsRecon MCP server gives the agent tools that return useful shapes:
- A page as clean Markdown, not HTML.
- A company’s logo, colors and fonts as data.
- A whole site through crawl, run as a job.
- An industry code from classification.
The agent picks the tool that fits the question. “What does this company do?” becomes one call, not a scraping project.
Wiring it up
The server speaks Streamable HTTP at https://mcp.agentsrecon.com/mcp and takes your API key as a Bearer token.
In Claude Code:
claude mcp add --transport http agentsrecon https://mcp.agentsrecon.com/mcp \
--header "Authorization: Bearer agr_YOUR_KEY"
Claude Desktop and Cursor need a few lines of JSON config. The MCP page has all three, ready to paste. The server arrives with the public launch, and the instructions are final.
A few habits that help
- Be specific in the prompt. “Look at example.com and give me its brand colors” beats “research this site”.
- Mind the quota. Every page the agent reads counts. A crawl of 500 pages is 500 pages. See pricing.
- Keep the key out of chat. Put it in config, never in a prompt or a repo.
The Free plan includes the MCP server, rate-limited, so you can try it before paying anything.
What the agent actually does with it
It helps to picture the loop. You ask Claude Code something like “look at this company’s site and tell me what they sell and what their brand looks like”. Without a tool, the agent writes a curl command, gets back a page of markup, and tries to read it. With the server connected, the agent sees a short list of tools with descriptions, picks the scrape tool for the text and the brand tool for the look, and gets two compact answers. It spends its context on your question instead of on the site’s navigation.
That difference matters more as the task grows. One page is a small problem. “Compare the pricing pages of these five competitors” is five pages, each of which would be a long dump of HTML. As Markdown, all five fit in one working session, and the agent can quote the parts that matter.
Where it works well, and where it does not
It works well for questions that need the content or the look of a public site: summarizing a product, pulling a palette, listing what a docs site covers, reading a PDF that a page links to. It works less well for anything that needs a login, a form submission or a click path through an app. Those are browser automation problems, not reading problems, and this tool does not pretend to solve them.
It also costs quota. Every page the agent reads counts against your monthly plan, so give it a clear job and a clear limit. “Read these three URLs” is a fine instruction. “Research this whole industry” is how a free month disappears in an afternoon. If you let an agent start a crawl, tell it the page limit you are comfortable with in the same prompt.
Setting it up safely
Treat the API key like any other secret. Put it in the client configuration, not in a prompt, a chat transcript or a repository. If a key ends up somewhere it should not, rotate it from the dashboard and update the config. Use a separate key for each machine or project, so one leak does not mean changing everything.
Test with something you know the answer to. Ask the agent to read a page you wrote and summarize it, then compare. If the summary matches, you have confirmed the connection, the key and the output format in one step, and you can start trusting it on pages you have not read yourself.
Why I built it as a server, not a library
I could have shipped only an SDK and let people write their own tool wrappers. Plenty will, and the TypeScript and Python SDKs exist for exactly that. But a wrapper is a small project every team repeats: describe the tools, handle the key, shape the output. The MCP server is that project done once. If your agent client speaks the protocol, you paste a few lines and it works.
If you try it and the tool descriptions confuse your agent, tell me. The wording of those descriptions is part of the product, and I would rather fix it than have you work around it.