AI agents & MCP
Give your agent the open web, safely.
A native MCP server, per-key budgets, citations on every fetch, and the same eight endpoints your humans use. Your agents don't need a custom scraper; they need a contract that says "this is the web, here's your budget."
8
tools, native MCP
60s
install to first call
100%
cited responses
$0
beyond your budget cap
What you actually need
The hard parts, already solved.
Native MCP, not a wrapper
Real Model Context Protocol. Tools, resources, prompts. Works with Claude Desktop, Cursor, VS Code, Windsurf, plus the OpenAI Agents SDK and Google ADK in Python.
Per-key budget caps
Cap how many credits an agent can burn per session (an <code>X-Ujeebu-Budget</code> header on the hosted endpoint, or <code>UJEEBU_BUDGET</code> when self-hosting with Docker). Hard stop at the limit. No nightmare bills from a runaway eval loop.
Citations built in
Every tool result includes source URL, fetch timestamp, credit cost. Your agent can ground its claims; you can audit every step.
12 tools, typed
<code>scrape</code> (html/screenshot/pdf/json/rules), <code>extract</code>, <code>card</code>, <code>serp</code>, <code>markdown</code>, <code>chatgpt</code>, <code>gemini</code>, <code>account</code>. Full JSON Schema on each, so the agent picks the right tool the first time.
Streaming responses
For long fetches, results stream into the agent's context as they arrive, so the model can reason on partial results before the full call returns.
Agent-friendly auth
Issue scoped sub-keys per agent or per session. Revoke instantly. Audit logs show every tool call with full request + response trace.
The pipeline
4 steps. One pipeline.
01
Install
Drop the config into your MCP host or use the one-click installers in the <a href="/mcp" class="uj-accent-text">setup wizard</a>. 60 seconds to first tool call from Claude Desktop, Cursor, VS Code, Windsurf, OpenAI Agents, or Google ADK.
02
Cap
Issue a sub-key with a credit budget. The agent can't exceed it; failed attempts cost nothing; you sleep at night.
03
Run
The agent picks tools by JSON Schema, executes them, and reasons over typed results. Streaming keeps long fetches responsive.
04
Audit
Every call logged with tool, args, response, citation, cost. Reproduce any agent step exactly. Critical for evals + customer-facing AI.
full pipeline · 30 lines
// claude_desktop_config.json
{
"mcpServers": {
"ujeebu": {
"url": "https://mcpdev.ujeebu.com/mcp",
"headers": {
"Authorization": "Bearer uj_live_...",
"X-Ujeebu-Budget": "5000"
}
}
}
}
// In Claude:
> Compare React Aria, Radix, Headless UI,
> and Reach UI on accessibility primitives.
🛠 ujeebu.serp({ q:"react aria vs radix" })
🛠 ujeebu.extract(url:"github.com/...") ×4
🛠 ujeebu.markdown(url:"docs/aria") ×4
// Agent reasons over the typed results,
// cites every claim, stops at 5000 credits.Endpoints in play
Use what fits. Skip what doesn’t.
MCP server
The protocol surface. One endpoint, twelve tools, every MCP host.
Scrape
Render any URL with JS execution. Screenshot and PDF as parameters.
Extract
Auto-detects page type - article, product, listing - and returns clean structured data.
SERP
Web search as a typed tool. Geo-targetable. CAPTCHA-free.
Markdown
When the agent needs to read a long doc, not just snippet it. Chunked, RAG-ready.
ChatGPT (URL)
Pass a URL plus a prompt; ChatGPT reads the page and replies. Skip the round-trip.
Gemini (URL)
Same as ChatGPT-by-URL, routed through Gemini. Pick the model that fits your task.
Common questions
Giving an AI agent web access, answered.
How do I give an AI agent access to the live web?
Connect the Ujeebu MCP server. The fastest way is our hosted endpoint - add <code>{ "url": "https://mcpdev.ujeebu.com/mcp", "headers": { "Authorization": "Bearer ..." } }</code> to your MCP host config (Claude Desktop, Cursor, VS Code, Windsurf); nothing to install. Prefer to self-host? Run <code>docker run -i --rm -e UJEEBU_KEY ujeebu/ujeebu-mcp</code> and point the config at that instead. Either way the agent gets twelve typed tools (<code>scrape</code> in five variants, plus <code>extract</code>, <code>card</code>, <code>serp</code>, <code>markdown</code>, <code>chatgpt</code>, <code>gemini</code>, <code>account</code>) and can fetch, search, and read the web through a single API key. The <a href="/mcp">setup wizard</a> generates one-click install links for Cursor and VS Code, copy-paste configs for the rest, and Python snippets for the OpenAI Agents SDK and Google ADK.
What is an MCP server for web scraping?
MCP (Model Context Protocol) is the open standard for connecting AI agents to external tools. Ujeebu ships a native MCP server, not a wrapper, that exposes its scraping, search, and extraction endpoints as typed tools with full JSON Schema, so the agent picks the right tool the first time. It works with any MCP-compatible client.
How do I stop an agent from running up a huge scraping bill?
Cap the per-session spend in the MCP config: on the hosted endpoint add an <code>X-Ujeebu-Budget</code> header, or when self-hosting with Docker set <code>UJEEBU_BUDGET</code> in the env. It is a hard stop. When the budget is hit, further calls fail rather than charge. Failed requests are never billed, so a runaway eval loop cannot generate a surprise invoice. The <code>account</code> tool also lets an agent self-check its remaining quota before kicking off a large batch.
Can the agent cite its sources?
Yes. Every tool result includes the source URL, the fetch timestamp, and the credit cost. The agent can ground each claim in a real source, and you get an audit trail of exactly which pages were fetched, when, and at what cost.
Does Ujeebu work with Claude, Cursor, VS Code, OpenAI Agents, and Google ADK?
Yes. The MCP server works with Claude Desktop, Cursor, VS Code, Windsurf, and any other MCP-compatible host. For Python agents, the OpenAI Agents SDK (<code>MCPServerStdio</code>) and Google ADK (<code>MCPToolset</code>) both consume the same MCP server with a short snippet. See <a href="/docs/mcp">the setup guide</a> for the exact code per host. If you are not using MCP at all, the same capabilities are available as plain REST endpoints at <code>https://api.ujeebu.com</code> using the <code>ApiKey</code> header.
How much does it cost to give an agent web access?
Start free with 5,000 credits and no credit card on a 14-day trial. Pricing is usage-based and a single credit pool is shared across every endpoint, so you do not pre-commit budget to one tool. Failed requests are not billed. See <a href="/pricing">/pricing</a> for current plans; top-ups are available on every plan including the trial.
Can an agent extract structured data without me writing selectors?
Yes. The <code>extract</code> tool auto-detects the page type (article, product, listing) and returns structured JSON with no selectors required. When you want to name specific fields, the <code>scrape</code> tool takes <code>extract_rules</code> and returns exactly those. For long documents, the <code>markdown</code> tool returns clean, chunked, RAG-ready Markdown the model can reason over directly.
How does Ujeebu handle anti-bot protection and JavaScript-heavy pages?
It is handled server-side. Ujeebu renders pages with a real headless Chrome browser, executes JavaScript, and routes through rotating, premium, residential, or mobile proxies with geo-targeting, built on years of anti-bot infrastructure. The agent never has to manage a browser or proxy pool itself.