MCP Computeflux

What mcp-computeflux does

The mcp-computeflux MCP server connects an MCP host to ComputeFlux, an OpenAI-compatible multi-model inference gateway deployed on the Polkadot testnet. It presents the backend as two agent-callable tools rather than requiring the host to implement the HTTP API directly.

The server fits the Conversational AI category. It is useful when an agent needs to discover available models and then send chat messages to a selected model. ComputeFlux is the default destination, but the same interface can target another service that implements the OpenAI-compatible API.

How it works

The mcp-computeflux MCP server communicates with its host using newline-delimited JSON-RPC 2.0 over standard input and output. Protocol responses are written to stdout, while diagnostic logs are sent to stderr. The implementation runs with Node.js 18 or newer and has no runtime npm dependency for the server itself.

Two tools are exposed:

• computeflux_models retrieves the model IDs available from the configured endpoint.
• computeflux_chat sends a non-streaming chat completion using a model ID and OpenAI-format messages. It also accepts optional max_tokens and temperature values.

A chat response includes the selected model, finish reason, latency, usage information, and assistant text. Requests require a bearer API key supplied through the environment.

Setup and configuration

Set COMPUTEFLUX_API_KEY before starting the process. By default, requests go to https://api.computeflux.ai/v1. Set COMPUTEFLUX_BASE_URL when using another OpenAI-compatible host.

The server can be installed through the official MCP Registry under io.github.computeflux2026isgod/mcp-computeflux, where supported clients can retrieve it by name. A direct setup can instead clone the repository and launch server.mjs with Node:

export COMPUTEFLUX_API_KEY="your_key"
node /path/to/mcp-computeflux/server.mjs

For a local end-to-end check, the repository includes an optional client/ harness. That harness installs the official MCP SDK, starts the server over stdio, lists its tools, and tests a chat completion. The npm installation applies to the test harness, not the server runtime.

Tools and capabilities

The mcp-computeflux MCP server supports model listing and single-request chat completion. Chat calls are non-streaming, so the complete result is returned after inference finishes. The endpoint abstraction allows the same MCP configuration to be used with ComputeFlux or another compatible API by changing the base URL.

Limitations and notes

An API key is required to make requests. The server does not provide a streaming chat tool, and the documented chat operation accepts the OpenAI messages structure rather than a separate prompt-only interface. ComputeFlux is described as running on the Polkadot testnet, and its default endpoint is used unless configuration overrides it.