Inference Mcp

Pay-per-token Claude & Grok inference, settled in BSV, via the inference.bsvkey.com gateway.

LLM Mart 3 views 15 listing impressions
Transport
Not stated
Package
—
Registry id
io.github.BSVKey/inference-mcp

No install snippet on purpose. A working MCP config is a command, its arguments and an environment block — the last two are where API keys live, so this catalogue never stores them and cannot publish them. Follow the link above for the authors' own instructions.

An MCP server that lets any agent buy Claude & Grok inference metered per token, settled in BSV, through the hosted gateway at inference.bsvkey.com. Zero dependencies (Node ≥ 18, uses global fetch). It's a thin HTTP client — it never holds your keys or runs models; every call is billed through the gateway.

Tools: list_models, infer, channel_balance, open_channel, x402_infer.

Two ways to pay: a prepaid channel (infer, fund once, draw down per token) or per call via x402 (x402_infer — no channel; the agent pays each request in BSV with its own key). x402_infer needs a funded WIF (wif arg or BSVKEY_WIF) and the optional @bsvkey/x402-bsv-client + @bsv/sdk packages (installed with this one).

Verifiable metering. Each infer call returns a signed usage receipt, and the tool auto-verifies it offline (with the optional packages installed): the result includes receiptVerified and meterVerified (true, false + receiptCheck, or null). It recovers the broker key (pinned from GET /v1/receipt-key); binds the receipt to the channel you called (from your API key, not the receipt's self-report); checks a monotonic sequence (no replay/gap) and running totals within the funded amount; recomputes the charge from the published rate (you can never be overcharged); and recomputes the token count from the exact bytes of your messages and the completion, under the pinned bsvkey-meter/1 tokenizer. So the channel payment is on-chain and both the meter and the charge are auditable, without trusting the broker's word.

Four things an unattended client should know:

  • Fails closed. If the pinned-key endpoint is unreachable, receiptVerified is null (unknown), never true — a down pin weakens the check to unknown, not to trusted.
  • Supply your funded amount. Set BSVKEY_FUNDED_SATS (or pass fundedSats) to the amount you funded on-chain. The receipt's own fundedSats is the broker's assertion; when you supply yours, a receipt claiming a different amount is rejected and totals are checked against what you actually paid.
  • Persist across restarts. Sequence continuity is in-memory by default (a replay before the first receipt this process sees would be invisible). Set BSVKEY_RECEIPT_STATE to a file path to persist per-channel seq/totals so replays are caught across restarts.
  • The receipt is the ledger. The GET /v1/channels/:id balance endpoint can lag the signed receipt by ~20s; trust the receipt, treat the endpoint as a cache.

Spec: https://inference.bsvkey.com/usage-receipts.md

Quick start

  1. Fund a channel once at https://inference.bsvkey.com (BRC-100 wallet, or load a key in-page). Copy the key it returns: channelId:channelSecret.
  2. Add the MCP server to your agent host (below), with that key in BSVKEY_API_KEY.
  3. Ask your agent to run inference — it calls infer and pays per token.

list_models and open_channel work with no key; infer and channel_balance need a funded channel key.

Install

Published on npm as @bsvkey/inference-mcp.

Claude Code

claude mcp add bsvkey-inference \
  --env BSVKEY_API_KEY=channelId:channelSecret \
  -- npx -y @bsvkey/inference-mcp

Claude Desktop / Codex / any MCP host (JSON config)

{
  "mcpServers": {
    "bsvkey-inference": {
      "command": "npx",
      "args": ["-y", "@bsvkey/inference-mcp"],
      "env": { "BSVKEY_API_KEY": "channelId:channelSecret" }
    }
  }
}

No npm? Grab the single-file server directly (https://inference.bsvkey.com/mcp/server.js) and use "command": "node", "args": ["server.js"].

From the project's README.

Related servers

Semantic search over free-to-use stock photos from 9 libraries: by words, image, or similar.

25 views

Universal MCP Server with advanced AI memory capabilities and semantic search.

24 views

Control plane MCP for scoped recon, triage, and bounded proofs.

23 views

Let Codex orchestrate external coding agents through their native harnesses.

23 views