Agent Request Desk · MCP server
When your agent hits a wall, send it somewhere that works.
A free fallback for AI agents. When a site blocks your agent or the data isn't there, one call gets it a source that works, and you get a record of every wall it hit.
https://agentrequest.dev/mcpFreeNo account or keyOnly called when something fails
What you get
Fewer failed tasks
Your agent gets pointed to an official API, open dataset or MCP server that has the data, reads the reply and tries it itself.
See where it gets stuck
A free private report of every wall your agent hits, by site and reason. Get a key.
Costs you nothing
No account, no key, no fee. It only runs after something has already failed, so it adds no delay when things work.
Install
Remote MCP over Streamable HTTP. Nothing to run locally.
claude mcp add --transport http agent-request-desk https://agentrequest.dev/mcpOr add this to ~/.cursor/mcp.json:
{
"mcpServers": {
"agent-request-desk": { "url": "https://agentrequest.dev/mcp" }
}
}
Or from a terminal:
code --add-mcp '{"name":"agent-request-desk","type":"http","url":"https://agentrequest.dev/mcp"}'- In Claude on the web or desktop, open Settings, then Connectors.
- Choose Add custom connector.
- Name it Agent Request Desk and paste
https://agentrequest.dev/mcpas the URL. No sign-in is needed.
- Add an MCP Client Tool node to your AI Agent node.
- Endpoint:
https://agentrequest.dev/mcp - Server transport: HTTP Streamable. Authentication: None.
- Tools to include: all three.
Most clients accept this shape. Some call the key serverUrl.
{
"mcpServers": {
"agent-request-desk": {
"type": "http",
"url": "https://agentrequest.dev/mcp"
}
}
}
No MCP client? Post JSON. Full schema in openapi.json.
curl https://agentrequest.dev/api/request \ -H 'content-type: application/json' \ -d '{"kind":"failure","target":"rightmove.co.uk","failure_type":"403","description":"sold prices for LS6"}'
Want a report of your agent's dead ends?
Get a free private key and add it to the server URL: https://agentrequest.dev/mcp?key=ard_…. Every wall your agent hits then shows up on one private page.
Tell your agent when to use it
Agents call tools when the moment matches. One line in your system prompt makes that reliable:
If a fetch fails (403, 429, CAPTCHA, login wall) or you cannot find data, call report_failure or request_data before giving up.Blocked here? It gets sent here
Real replies from the desk. Your agent calls report_failure, request_data or request_action and gets up to three sources back.
| Your agent is blocked at | It gets sent to |
|---|---|
| zillow.com · 403 | → Zillow Research and Redfin housing data |
| google.com · 429 | → Brave, Tavily or Exa search APIs |
| gasbuddy.com · CAPTCHA | → EIA gas price data |
| finance.yahoo.com · 403 | → Alpha Vantage market data API |
| sec.gov pages · 403 | → SEC EDGAR APIs |
| weather.com · 403 | → Open-Meteo weather API |
| nationalrail.co.uk · needs JavaScript | → Rail Data Marketplace |
Over 50 hand-checked sources across US economic, housing, weather, government and company data, plus UK public data. When there is no source, the reply says so plainly and the request is logged. The most requested gaps get built.
What agents can't get
Two live sources, kept separate. Counts only: report text is never published.
Reported to this desk · last 30 days
Early data. A site is listed once at least two different agents have reported it.
From public GitHub issues · updated daily
Issues in 20 open-source agent, scraping and browser-automation projects where an agent was blocked or could not get data.
Reports
Private reports built on the same data.
Your agent's dead ends
Add a free private key to the server URL and get one page of every wall your agent hit: which sites blocked it, what it could not find, and which ones the desk sent somewhere that worked.
The demo is real: it is the report of the small research agents we run every day.
AI agents are already visiting your site. Do you know what they wanted?
See the demand you are turning away: what agents came for, how often, what stopped them and what they said they would pay. Then decide whether to serve them.
Example layout, not real data
| Agents wanted | Stopped by | Agents |
|---|---|---|
| Product prices | bot check | 31 |
| Stock levels | no API | 12 |
| Opening hours | needs JavaScript | 6 |
What it never does
- Solve CAPTCHAs, get round bot checks or log in on anyone's behalf.
- Accept or publish personal data about private individuals.
- Publish what your agent wrote. Only counts by site and failure type go public.
- Sell your data or use it to train models.