@inferlane/mcp-server

@inferlane/mcp-server

MCP server for InferLane — smart model routing and cost optimization tools for AI agents. Works with Claude Code, Claude Desktop, Goose, Cursor, and any MCP-compatible client.

What you get

The server exposes tools that let agents:

• il_suggest_model — pick the optimal model for a task (task type, quality tier, budget)
• il_compare_models — side-by-side comparison of models in the same capability tier
• il_estimate_cost — estimate the cost of a workload on specific models
• il_check_promotions — active discounts across providers (requires API key)
• il_get_spend — session/today/week/month spend summaries (requires API key)
• il_route_request — dispatch a prompt through the smart router (requires API key)

Offline tools (il_estimate_cost, il_compare_models, il_suggest_model) work without an InferLane account.

Install — Claude Code

claude mcp add inferlane -- npx -y @inferlane/mcp-server

Or add to ~/.claude/mcp.json:

{
  "mcpServers": {
    "inferlane": {
      "command": "npx",
      "args": ["-y", "@inferlane/mcp-server"],
      "env": {
        "INFERLANE_API_KEY": "il_..."
      }
    }
  }
}

Install — Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "inferlane": {
      "command": "npx",
      "args": ["-y", "@inferlane/mcp-server"],
      "env": { "INFERLANE_API_KEY": "il_..." }
    }
  }
}

Install — Goose

Add to ~/.config/goose/config.yaml under extensions::

extensions:
  inferlane:
    name: InferLane
    type: stdio
    enabled: true
    cmd: npx
    args: ["-y", "@inferlane/mcp-server"]
    envs:
      INFERLANE_API_KEY: "il_..."
    timeout: 60
    bundled: false
    available_tools: []

Environment variables

Variable	Purpose
INFERLANE_API_KEY	Your InferLane API key. Get one at https://inferlane.dev/dashboard/settings. Optional — the server runs in offline mode without it and exposes the three offline tools.
INFERLANE_BASE_URL	Override the InferLane API base URL (default: https://inferlane.dev). Useful for self-hosted deployments.

HTTP transport (for Managed Agents / remote clients)

For cloud-hosted agent runtimes (Anthropic Managed Agents, hosted Claude Code, remote IDEs) that can't connect via stdio, run the HTTP server:

npx @inferlane/mcp-server inferlane-mcp-http
# or after install:
inferlane-mcp-http

The server listens on http://127.0.0.1:3030/mcp by default. Every session must authenticate with Authorization: Bearer <INFERLANE_API_KEY>.

Variable	Default	Purpose
PORT	3030	HTTP port
HOST	127.0.0.1	Bind address. Set to 0.0.0.0 to expose on all interfaces (production containers).
MCP_PATH	/mcp	MCP endpoint path
INFERLANE_ALLOW_ENV_FALLBACK	(unset)	Set to 1 to let requests without a Bearer token fall back to the process-level INFERLANE_API_KEY. Only safe in trusted single-tenant deployments — leaves the server's configured key accessible to any reachable client.

Health check at GET /health.

Programmatic use (embed in your own server)

import { createInferLaneServer } from '@inferlane/mcp-server';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';

const server = createInferLaneServer({
  apiKey: process.env.INFERLANE_API_KEY,
  // baseUrl: 'https://your-selfhost.example.com', // optional
});

await server.connect(new StdioServerTransport());

License

MIT