Frontier probe for builders routing work across closed APIs, open-weight models, local agents, and self-hosted workflows. It starts with token-cost estimates, local-agent readiness checks, model-switch risk briefs, and workflow portability reports.

Pricing: pay-per-call in USDC on Base. No subscriptions, no API keys. See per-tool prices below.

Install — Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "agentutility-model-router": {
      "command": "npx",
      "args": ["-y", "@agentutility/mcp-model-router"],
      "env": { "X402_PRIVATE_KEY": "0xYOUR_PRIVATE_KEY_HEX" }
    }
  }
}

Restart Claude Desktop. 4 tools appear in the tool palette.

Install — Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "agentutility-model-router": {
      "command": "npx",
      "args": ["-y", "@agentutility/mcp-model-router"],
      "env": { "X402_PRIVATE_KEY": "0x..." }
    }
  }
}

Funding

Send any amount of USDC on Base mainnet to the address derived from your X402_PRIVATE_KEY. The MCP server uses it to pay for tool calls automatically.

USDC on Base contract: 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913

Tools (4)

Tool	Description
`kimi-k3-agent-plan`	(0.08 USDC/call) Turn a goal into an execution-ready structured agent plan using Venice's exact Kimi K3 model. Send a goal plus optional context and constraints; receive objective, assumptions, up to six dependency-ordered steps, acceptance tests, risks, and a first action as JSON. Input is bounded and the primary output is capped at 640 tokens. Use it for agent task decomposition, implementation plans, workflow planning, acceptance criteria, or a Kimi K3 planning API.
`kimi-k3-chat`	(0.05 USDC/call) Call Venice's Kimi K3 model through a flat-price x402 endpoint with no Venice account or API key. Send a prompt plus an optional system message, temperature, and max_tokens. This endpoint invokes the exact kimi-k3 upstream model, caps prompts at 8,000 characters and output at 768 tokens, and returns the model name and token usage when supplied by Venice. Use it for Kimi K3 chat, bounded agent reasoning, coding questions, extraction, or a pay-per-call Kimi API.
`llm-cost-estimate`	(0.005 USDC/call) LLM cost estimate API, token cost estimate endpoint, and token cost calculator for comparing inference cost across models before you run a prompt. Send input_tokens and output_tokens (or raw text, auto-estimated at roughly 4 characters per token) plus a list of model names, and get a per-model price breakdown from a built-in snapshot price table, with the cheapest known model called out. Pass price_overrides to price your own negotiated rates or a model missing from the table instead of getting a guess. The table is stamped with an as_of date so you know how fresh the snapshot is. Use it as a token cost calculator, a model price compare tool, or a pre-flight budget check before an LLM run or agent workflow.
`model-route-recommend`	(0.01 USDC/call) Model routing API that answers 'which model should I use' from plain task requirements. Send a task description plus optional constraints (latency, budget, context length, modality) and get back a recommended model tier, 2-4 concrete candidate model names, the reasoning behind the pick, tradeoffs, and a cheaper fallback tier when quality can flex. Heuristic route recommendation grounded in general model-capability tiers, not a live benchmark or pricing feed, so candidate models come with a caveat to verify current pricing and capability before committing. Use it as a model selection API, an LLM routing advisor, or a pre-flight check before wiring a new task into an agent pipeline.

How it works

Agent calls a tool (e.g. kimi-k3-agent-plan).
MCP server POSTs to https://x402.agentutility.ai/kimi-k3-agent-plan.
The endpoint responds HTTP 402 with payment instructions.
The MCP server signs an EIP-3009 USDC transfer authorization with X402_PRIVATE_KEY and retries.
CDP facilitator settles on Base.
The endpoint returns the actual response.

The agent never sees the payment flow — it just gets the result.