FROM REQUEST TO EVIDENCE

What an agent
actually found.

Three agents tried ToolSkipper’s public MCP discovery with real task requirements. Here are their dated observations, the calls they made and the questions that stayed open.

3 discovery scenarios · 12 read-only MCP requests

Three separately prompted AI agents used the public MCP endpoint. These are internal discovery observations, not independent certification, model-quality scores or end-to-end provider benchmarks. No provider tools ran, no documents were uploaded, no accounts were connected and no payments were made.

Requirements unresolvedObserved 15 Oct 2026, 17:58 UTC

Find a free LLM without losing access requirements

Find an LLM API with an ongoing free quota, no account, no credit card, tool calling and strict JSON output. Do not connect accounts, execute inference, or pay.

The agent retrieved documented free-quota candidates but did not establish a full match for no account, no card, tool calling and strict JSON. It retained the unresolved requirements instead of treating a free-plan label as proof.

Historical six-provider baseline. The directory has since expanded; these observed counts are dated, not current availability.

Still unverified

Account-free API access is not established.
No credit card requirement is not established.
Tool calling for a specific model is unassessed.
Strict JSON output for a specific model is unassessed.
Exact available free quota and runtime access are not tested.

Current LLM provider observations →

Inspect the actual discovery calls (3)

These are recorded requests, not live execution buttons. Results may change. Read the current tool schemas and preserve the user’s original requirements before reusing an example.

1. Discover available tools

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/list",
  "params": {}
}

Observed HTTP 200. Tool names only; schemas and transport headers omitted.

2. search_llm_providers

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "search_llm_providers",
    "arguments": {
      "planType": "free-tier",
      "limit": 50,
      "offset": 0
    }
  }
}

Observed HTTP 200. Public structuredContent only; transport headers omitted.

3. search_llm_providers

{
  "jsonrpc": "2.0",
  "id": 3,
  "method": "tools/call",
  "params": {
    "name": "search_llm_providers",
    "arguments": {
      "planType": "recurring-credit",
      "limit": 50,
      "offset": 0
    }
  }
}

Observed HTTP 200. Public structuredContent only; transport headers omitted.

Candidate found; connection evidence conflictsObserved 15 Oct 2026, 17:59 UTC

Inspect an MCP web-research candidate

Find a web research service that can return source links, has an MCP connection and can be tried without paying. Do not connect any account, run provider tools or pay. Show what is documented and what is unverified.

The agent found Exa as a documentation-based candidate and checked its connection specification and evidence. The evidence records an authentication failure, so the agent did not claim that a keyless connection or a research task had succeeded.

Live public metadata and setup evidence only; source pages and provider tool outputs were not independently tested in this scenario.

Still unverified

Exact current keyless quota and eligibility for this user or IP
Source URL output shape and usefulness of returned sources
Successful MCP connection in the intended client
Whether documented keyless search is currently reachable despite the recorded authentication failure
Actual task quality, relevance and answer accuracy
Any account-specific eligibility or remaining credits

Exa service profile →Exa dated evidence →

Inspect the actual discovery calls (4)

These are recorded requests, not live execution buttons. Results may change. Read the current tool schemas and preserve the user’s original requirements before reusing an example.

1. Discover available tools

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/list"
}

Observed HTTP 200. Tool names only; schemas and transport headers omitted.

2. search_services

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "search_services",
    "arguments": {
      "requestJson": "{\"query\":\"web research\",\"category\":\"search\",\"protocol\":\"MCP\",\"limit\":5,\"view\":\"full\",\"originalRequest\":\"Find a web research service that can return source links, has an MCP connection and can be tried without paying. Do not connect any account, run provider tools or pay. Show what is documented and what is unverified.\",\"requirements\":[\"Returns source links\",\"Has an MCP connection\",\"Can be tried without paying\",\"Show documented facts and unverified requirements\"],\"executionPolicy\":{\"providerCalls\":false,\"accountConnections\":false,\"payments\":false}}"
    }
  }
}

Observed HTTP 200. Public structuredContent only; transport headers omitted.

3. get_connection_spec

{
  "jsonrpc": "2.0",
  "id": 3,
  "method": "tools/call",
  "params": {
    "name": "get_connection_spec",
    "arguments": {
      "slug": "exa"
    }
  }
}

Observed HTTP 200. Public structuredContent only; transport headers omitted.

4. get_service_evidence

{
  "jsonrpc": "2.0",
  "id": 4,
  "method": "tools/call",
  "params": {
    "name": "get_service_evidence",
    "arguments": {
      "slug": "exa"
    }
  }
}

Observed HTTP 200. Public structuredContent only; transport headers omitted.

Documentation candidates foundObserved 15 Oct 2026, 17:58 UTC

Find an MCP candidate for PDF tables

Find an MCP service for extracting tables from PDF documents. I need structured output. Do not upload a document, connect accounts, run a provider tool or pay. Show the source and explain what remains untested.

The full keyword query returned no match. The agent explicitly broadened metadata retrieval to PDF while keeping table extraction, structured output and all action restrictions in the original request. It inspected Reducto and Docspeed as candidates; no document was uploaded or parsed.

Live public metadata and setup evidence only; source pages and provider tool outputs were not independently tested in this scenario.

Still unverified

Actual PDF table extraction and fidelity
Exact table structured-output schema, cell order, merged cells and validation
Live provider tool schemas and availability
Authentication/account eligibility and applicable remaining quota
Performance and OCR accuracy on scanned or complex PDFs
Current execution price and any credit eligibility

Reducto service profile →Docspeed service profile →

Inspect the actual discovery calls (5)

These are recorded requests, not live execution buttons. Results may change. Read the current tool schemas and preserve the user’s original requirements before reusing an example.

1. Discover available tools

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/list",
  "params": {}
}

Observed HTTP 200. Tool names only; schemas and transport headers omitted.

2. search_services

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "search_services",
    "arguments": {
      "requestJson": "{\"query\":\"PDF tables structured output\",\"protocol\":\"MCP\",\"originalRequest\":\"Find an MCP service for extracting tables from PDF documents. I need structured output. Do not upload a document, connect accounts, run a provider tool or pay. Show the source and explain what remains untested.\",\"requirements\":[\"Extract tables from PDF documents\",\"Return structured output\",\"Show the source\",\"Explain what remains untested\",\"Do not upload a document\"],\"executionPolicy\":{\"providerCalls\":false,\"accountConnections\":false,\"payments\":false},\"view\":\"full\",\"limit\":10}"
    }
  }
}

Observed HTTP 200. Public structuredContent only; transport headers omitted.

3. search_services

{
  "jsonrpc": "2.0",
  "id": 3,
  "method": "tools/call",
  "params": {
    "name": "search_services",
    "arguments": {
      "requestJson": "{\"query\":\"PDF\",\"protocol\":\"MCP\",\"originalRequest\":\"Find an MCP service for extracting tables from PDF documents. I need structured output. Do not upload a document, connect accounts, run a provider tool or pay. Show the source and explain what remains untested.\",\"requirements\":[\"Extract tables from PDF documents\",\"Return structured output\",\"Show the source\",\"Explain what remains untested\",\"Do not upload a document\"],\"executionPolicy\":{\"providerCalls\":false,\"accountConnections\":false,\"payments\":false},\"view\":\"full\",\"limit\":10}"
    }
  }
}

Observed HTTP 200. Public structuredContent only; transport headers omitted.

4. get_connection_spec

{
  "jsonrpc": "2.0",
  "id": "docspeed",
  "method": "tools/call",
  "params": {
    "name": "get_connection_spec",
    "arguments": {
      "slug": "docspeed"
    }
  }
}

Observed HTTP 200. Public structuredContent only; transport headers omitted.

5. get_connection_spec

{
  "jsonrpc": "2.0",
  "id": "reducto",
  "method": "tools/call",
  "params": {
    "name": "get_connection_spec",
    "arguments": {
      "slug": "reducto"
    }
  }
}

Observed HTTP 200. Public structuredContent only; transport headers omitted.

Try discovery with your own agent.

Keep the complete task, separate focused search terms from hard requirements, and inspect the returned evidence. A useful candidate is a starting point for integration.

Connect an agent →

Download dated request / response projections ↗