MCP Server

Our MCP server lets coding agents search the docs and read your account. It’s a remote HTTP server at:

https://mcp.flexinference.com/mcp

Connecting

{
  "mcpServers": {
    "flexinference": {
      "type": "http",
      "url": "https://mcp.flexinference.com/mcp"
    }
  }
}

claude mcp add --transport http flexinference https://mcp.flexinference.com/mcp

The first account call opens a browser for dashboard login and shows the requested scopes, which you can approve or narrow. Docs tools need no login.

search_docs searches the docs and the full error catalog, and returns the best matches with links.
get_error_reference looks up what an error code means and how to fix it. Leave code out to list every code.
run_flex_demo sends one prompt twice on the same model, once through the Flex Race and once as a default request, and returns both answers with the billed comparison. One run per IP address per day.
compare_usage_export prices a usage CSV export (OpenAI platform, Anthropic console, or OpenRouter) at each model’s measured rate and returns the per-model breakdown. Neither tool stores what you send.

Each one takes one scope, and each one only reads.

get_usage_summary (usage:read) reads your usage and savings by model.
list_api_keys (keys:read) lists your keys by name, last 4, and created date. Never the full key.
get_byok_status (byok:status) reads which providers have a key on file, as a yes or no. Never the key.
get_billing_status (billing:read) reads your plan and balance. Never the full card number.

No tool here creates or takes a key. Keys live in the dashboard.