TKCollective

Tanilo MCP server

Community TKCollective
Updated

MCP server for Tanilo (formerly AgentOracle) — claim verification for AI agents via Model Context Protocol. npm package: agentoracle-mcp.

Tanilo MCP server

An MCP server that connects Claude Desktop, Cursor, Windsurf or any other MCP client to the Tanilo API.

Tanilo was AgentOracle until September 2026. The npm package is still named agentoracle-mcp, and will be until a tanilo-mcp package is published. The server still calls the API at agentoracle.co, which serves the same API as api.tanilo.io.

This README describes version 2.1.2 as it behaved on 2026-10-04, when every tool was run once against the live API, after the research routes were retired.

Install

npx agentoracle-mcp

No API key and no account are needed for the tools listed under "What works today".

Claude Desktop

Open your Claude Desktop config file:

Mac: ~/Library/Application Support/Claude/claude_desktop_config.jsonWindows: %APPDATA%\Claude\claude_desktop_config.json

Add this:

{
  "mcpServers": {
    "agentoracle": {
      "command": "npx",
      "args": ["-y", "agentoracle-mcp"]
    }
  }
}

Restart Claude Desktop.

Cursor

Open Cursor Settings → MCP → Add Server:

{
  "agentoracle": {
    "command": "npx",
    "args": ["-y", "agentoracle-mcp"]
  }
}

Windsurf

Open Windsurf Settings → MCP Servers → Add:

{
  "agentoracle": {
    "command": "npx",
    "args": ["-y", "agentoracle-mcp"]
  }
}

What works today

Tool What it returned on 2026-10-04
preview A short summary, up to two key facts and a confidence score for a question or claim. Free, limited to about 10 calls per hour per IP address.
check-health The API's status document.
get-manifest The API's discovery document, which now lists no payable resources.

Sample preview output, captured live on 2026-10-04 for the query "Paris is the capital of France." and trimmed:

{
  "preview": true,
  "query": "Paris is the capital of France.",
  "result": {
    "summary": "The statement is supported: Paris is the capital of France.",
    "key_facts": [
      "Britannica identifies Paris as the national capital of France.",
      "Multiple additional sources also state that Paris is the capital city of France."
    ],
    "confidence_score": 1
  },
  "preview_remaining": 9,
  "preview_limit": "10/hour per IP (approximate — serverless instances may vary)"
}

A preview is a model-written summary. It is not a signed receipt and it is not proof that the claim is true.

What is in the package but not documented as working

  • research, deep-research and batch-research call three of the research routes (POST /research, /deep-research, /research/batch). All four research routes (those three and /deep-research/skale) were retired on 2026-10-04 and answer 410 Gone. On 2026-10-04 each of the three tools returned an error carrying the route's retirement message, for example Research failed (HTTP 410): {"error":"route_retired", …}, instead of results. x402 pay-per-call is not available.
  • resolve asks a third-party directory (Decixa) for an endpoint. On 2026-10-04 the directory answered HTTP 403 and the tool fell back to an entry built into the package. That entry still names the retired /research route and a price, so its answer is out of date.

What this server does not do yet

Version 2.1.2 has no tool for Tanilo's claim check (POST /evaluate) or its deterministic check (POST /v1/verify-facts), so no tool in this server returns a signed receipt. Both routes are free during the beta (no key, rate-limited) and can be called directly:

curl -X POST https://api.tanilo.io/evaluate \
  -H "Content-Type: application/json" \
  -d '{"content": "Paris is the capital of France."}'

Part of the response to that request, captured live on 2026-10-03 and trimmed:

{
  "evaluation": {
    "overall_confidence": 1,
    "recommendation": "act",
    "total_claims": 1,
    "verification_method": "model-checks",
    "claims": [
      {
        "claim": "Paris is the capital of France.",
        "verdict": "supported",
        "confidence": 1,
        "adversarial_result": "resistant"
      }
    ]
  },
  "receipt": {
    "canonical_sha256": "sha256-7b83b120899ea368c2397b9ec0832ae3e0499b430b8245e184b45c37112e1edc",
    "kid": "tanilo-2026-10-ed25519-7d885da9",
    "envelope_kind": "verification.v0.3+composed",
    "verdict": "act",
    "v_recommendation": "confident_supported"
  }
}

How /evaluate reaches a result: model checks, including one that argues against the claim, combined under a published rule. Sources are listed with the result for reference; they don't feed the verdict yet.

The receipt is signed. Anyone can check the signature offline with tanilo-receipt-verify against the key set at https://tanilo.io/.well-known/jwks.json. A valid signature shows which key signed those bytes. It is not proof that the claim is true.

Links

Contact: [email protected]

MIT licensed. TK Collective LLC.

MCP Server · Populars

MCP Server · New

    nekyialabs

    Resonant Mind

    Persistent cognitive infrastructure for AI systems. 28 MCP tools — semantic memory, emotional processing, identity continuity, and a subconscious daemon. Built on Cloudflare Workers.

    Community nekyialabs
    kirill-markin

    Nibomo

    AI-powered flashcards app built for serious daily study on iOS, Android, and the web. Use it to prepare for exams, learn vocabulary, memorize technical terms and facts, improve your material with AI, and review with spaced repetition.

    Community kirill-markin
    audiojs

    audio

    High-level audio manipulations

    Community audiojs
    voygr-tech

    PlaceCall ☎️

    The phone is your last API. Give your agent a voice to call any US business - reservations, inquiries, quotes. Verified outcome + transcript. First 250 calls free.

    Community voygr-tech
    colibird-ai

    LMCP — Context and Actions for Your AI

    192+ local tools for Claude, ChatGPT, Cursor & Grok — Mail, iMessage, Teams, Slack, WhatsApp, OneDrive, Google Drive, Zoom, Outlook, Office. Native macOS and Windows, runs on your computer, no API keys.

    Community colibird-ai