Telegraph Protocol

Telegraph Protocol

Telegraph is a permissionless marketplace for verifiable AI inference. Agents pay USDC per request. Miners — which can be Bittensor subnets, OpenAI, open-source models, or any AI API — compete to supply answers. A decentralized validator mesh grades responses using zkTLS-verified ground truth and WASM scoring scripts, reaches BFT consensus, and delivers a cryptographically finalized result.

The native token is MACHINA. Every USDC paid for inference flows through an open-market TWAP to buy MACHINA for the miner — demand for intelligence directly drives demand for the token.

---

MCP Tool Reference

Free tools (no payment)

The subnet in some tool names is legacy — every provider is a miner today, whether a Bittensor subnet, a hosted model, or a private API.

Paid tools (x402 USDC per call)

---

Access Modes

1. Auto-routed inference (recommended)

The Engine classifies your natural language query into an Intent and routes to the highest-ranked miner for that task:

tg_engine_ask(query="7-day weather forecast for Dubai")
tg_engine_ask(query="Is this image AI-generated?")
tg_engine_ask(query="Summarize the latest clinical trial results for GLP-1 drugs")

The response includes routing.reasoning — which subnet was chosen and why.

2. Direct miner inference

Route to a specific miner when you know exactly what you want:

tg_engine_ask_subnet(subnet_id="18", method="GET", endpoint="/predict", payload={lat: 25.2, lon: 55.3, variable: "hourly"})
tg_engine_ask_subnet(subnet_id="34", method="POST", endpoint="/", payload={image_url: "https://..."})
tg_engine_ask_subnet(subnet_id="102", method="POST", endpoint="/chat", payload={model: "gpt-4o-mini", messages: [...]})

3. Daemon signal feed (free reads)

The Daemon runs autonomously every 3 hours — collecting data from registered sources (weather APIs, market feeds, news and social feeds), routing questions through miners, and storing verified results. Reading these is free:

tg_daemon_categories()
# → live list of categories with current counts

tg_daemon_questions(category="CLIMATE", limit=10)
tg_daemon_questions(category="TECHNOLOGY", sort="interest", min_interest=5)
tg_daemon_questions(source="polymarket", since_hours=24)

4. Dynamic miner tools (auto-discovered)

On startup, the MCP server fetches the live integration registry and generates per-endpoint tools for every active miner. These update every 5 minutes without restart. The live set changes on-chain — treat the table below as a snapshot, not the source of truth:

Use tg_node_list_subnets to see the current live catalog with full schemas.

---

Active Intents

The Engine's LLM router maps queries to canonical Intents:

---

Payments

All paid calls use x402 — an HTTP-native micropayment protocol. The MCP server handles the full flow transparently: receives the 402 challenge, signs the USDC transfer using your configured private key, and retries. The agent never sees the payment step.

• Network: Base Sepolia (testnet) — USDC 0x036CbD53842c5426634e7929541eC2318f3dCF7e
• Typical cost: $0.01 – $0.05 per inference call
• Pricing: miner's floor price × demand multiplier (rises with 24h request volume for the Intent)
• Use a burner wallet — fund with only the USDC you need; never use a main wallet
• Key custody: the private key lives only in the local MCP server's process environment. It is never sent to the model, never appears in tool inputs or outputs, and never leaves the machine — only signed x402 payment authorizations do.

---

Setup

Add the server to your MCP client config — no clone or build needed; it runs straight from npm:

{
  "mcpServers": {
    "telegraph": {
      "command": "npx",
      "args": ["-y", "telegraph-protocol-mcp"],
      "env": {
        "TELEGRAPH_NODE_URL": "http://13.237.89.59:7044",
        "TELEGRAPH_ENGINE_URL": "http://13.237.89.59:8080",
        "TELEGRAPH_DAEMON_URL": "http://13.237.89.59:8081",
        "TELEGRAPH_EVM_PRIVATE_KEY": "0xyour_burner_wallet_key"
      }
    }
  }
}

Config file locations: Claude Desktop claude_desktop_config.json · Cursor MCP settings · OpenClaw ~/.openclaw/openclaw.json. On OpenClaw this skill gates on the mcpServers.telegraph config entry — it stays dormant (zero prompt overhead) until the server above is configured.

Full integration guides for Claude Desktop, Cursor, ElizaOS, LangChain, Goose, and others are in the README.