IBM Granite via x402

@three-ws/ibm-x402-mcp

Pay-per-use IBM Granite AI over MCP — chat, code, embeddings, analysis, and forecasting, billed in USDC on Solana. No IBM account required.

Install · Quick start · Tools · Payment flow · Requirements · three.ws

A Model Context Protocol server that exposes IBM Granite foundation models as pay-per-use tools via the x402 payment protocol. End users pay USDC on Solana per call — no IBM Cloud account of their own. The server operator supplies IBM credentials (WATSONX_*) and a receiving Solana wallet (MCP_SVM_PAYMENT_ADDRESS); callers supply only USDC. One free tool (ibm_granite_getting_started) explains prices and the flow before any payment.

Built by three.ws. Community-built and not affiliated with IBM.

How it works

1. An MCP client (Claude Desktop, Claude Code, Cursor, or an agent) connects to this server.
2. The client calls a tool — e.g. ibm_granite_chat.
3. Without an x402 payment payload, the server returns a 402 PaymentRequired envelope quoting the USDC price and the Solana receiving address.
4. The client signs a Solana USDC transfer and retries with the payment in _meta["x402/payment"].
5. The server verifies and settles the payment via the facilitator, calls IBM watsonx.ai, and returns the result with a settlement receipt in _meta["x402/payment-response"].

x402-capable MCP clients handle this loop automatically.

Install

npm install @three-ws/ibm-x402-mcp

Run it directly with npx (no install needed):

MCP_SVM_PAYMENT_ADDRESS=<your-solana-wallet> \

WATSONX_API_KEY= \ WATSONX_PROJECT_ID= \ npx @three-ws/ibm-x402-mcp

Or install globally for the ibm-x402-mcp binary on your PATH:

npm install -g @three-ws/ibm-x402-mcp

Quick start

Whoever runs the process is the operator: the server fails fast at startup without a payment address and IBM credentials (WATSONX_API_KEY, WATSONX_PROJECT_ID). Callers of the running server need only USDC — no IBM account. With Claude Code, one command:

claude mcp add ibm-granite-x402 \

-e MCP_SVM_PAYMENT_ADDRESS=your-solana-wallet \ -e WATSONX_API_KEY=your-ibm-api-key \ -e WATSONX_PROJECT_ID=your-watsonx-project-id \ -- npx -y @three-ws/ibm-x402-mcp

Or wire the server into your MCP client config (claude_desktop_config.json, Cursor's mcp.json):

{

"mcpServers": { "ibm-x402": { "command": "npx", "args": ["-y", "@three-ws/ibm-x402-mcp"], "env": { "MCP_SVM_PAYMENT_ADDRESS": "your-solana-wallet-address", "WATSONX_API_KEY": "your-ibm-cloud-api-key", "WATSONX_PROJECT_ID": "your-watsonx-project-id" } } } }

Inspect the tool surface with the MCP Inspector:

npx -y @modelcontextprotocol/inspector npx @three-ws/ibm-x402-mcp

Tools

Every tool is a read-only model-inference call — nothing on your machine or in any account is modified — and declares MCP tool annotations (readOnlyHint, openWorldHint, idempotentHint) so clients can reason about side effects before paying.

Input parameters

`ibm_granite_getting_started` — section (optional: overview (default, everything) / pricing / payment / tools / setup).

`ibm_granite_chat` — messages (required: 1–50 { role, content } pairs; roles system/user/assistant, content ≤ 32,000 chars), model (default ibm/granite-3-8b-instruct), max_new_tokens (1–4096, default 1024), temperature (0–2, default 0.7).

`ibm_granite_code` — task (required: generate/review/refactor/explain/test/document), prompt (required, ≤ 16,000 chars: the description for generate, the code for everything else), language, context (≤ 4,000 chars).

`ibm_granite_embed` — inputs (required: 1–64 texts, ≤8000 chars each), model.

`ibm_granite_analyze` — document (required), analysis_type (general/contract/financial/technical/medical/sentiment, default general), language.

`ibm_granite_forecast` — timestamps (required: 64–1024 ISO-8601, uniform cadence, oldest first), values (required: 64–1024 numbers, same length), freq (required: pandas cadence, e.g. 1h, 1D), prediction_length (1–96), label.

Example calls

// ibm_granite_chat

{ "messages": [ { "role": "system", "content": "You are an expert data engineer." }, { "role": "user", "content": "Design a lakehouse schema for IoT sensor telemetry." } ], "max_new_tokens": 1024, "temperature": 0.7 }

// ibm_granite_code { "task": "review", "prompt": "def calculate_roi(revenue, cost): return revenue / cost", "language": "Python" }

// ibm_granite_embed { "inputs": ["enterprise data governance", "cloud-native AI pipeline", "real-time analytics"] }

// ibm_granite_analyze { "document": "This Software License Agreement is entered into between...", "analysis_type": "contract" }

// ibm_granite_forecast (timestamps/values must be 64–1024 points; abbreviated here) { "timestamps": ["2025-01-01T00:00:00Z", "...", "2025-03-05T00:00:00Z"], "values": [12500, "...", 13200], "freq": "1D", "prediction_length": 14, "label": "daily_revenue_usd" }

Payment flow

This server uses the x402 protocol for micropayments:

1. Client calls a tool without payment → 402 PaymentRequired with the USDC amount and Solana address.
2. Client builds and signs a Solana USDC transfer transaction.
3. Client retries with the signed tx in _meta["x402/payment"].
4. Server verifies and settles via the configured facilitator (default PayAI).
5. Server calls IBM watsonx.ai and returns the result with _meta["x402/payment-response"] (settlement receipt).

MCP Client (Claude Desktop / Cursor / agent)

│ tools/call (with x402 payment in _meta) ▼ ibm-x402-mcp (stdio MCP server) │ verify + settle USDC on Solana ├──► x402 facilitator (default https://facilitator.payai.network) │ │ inference call with IAM Bearer token └──► IBM watsonx.ai (us-south.ml.cloud.ibm.com) └── IBM Granite 3 8B Instruct / Embedding / TTM

Requirements

• Node.js >= 20.
• A Solana wallet address to receive USDC (MCP_SVM_PAYMENT_ADDRESS).
• IBM Cloud credentials: an API key (create one) and a watsonx.ai project id (Project → Manage → General → Project ID), or a deployment space id.

Environment variables

Regional hosts: us-south, eu-de, eu-gb, jp-tok, au-syd, ca-tor — e.g. https://eu-de.ml.cloud.ibm.com.

Errors

Every failure state is explicit — there is no silent fallback:

| Unexpected server fault | Tool result { "ok": false, "error": "internal_error", "message": "…" }. | Retry