mcp-computeflux

io.github.computeflux2026isgodv1.0.0Updated Oct 4, 2026

OpenAI-compatible multi-model inference via ComputeFlux (Polkadot testnet, TEE-verifiable). stdio.

VerifiedSTDIODesktop onlyDeveloper ToolsAI & ML

Overview

AI-generated overview

Lets an assistant list and call OpenAI-compatible chat models through the ComputeFlux gateway as plain MCP tools.

What it does
Exposes two tools over stdio: one lists the model ids served by the configured endpoint, and the other runs a non-streaming chat completion using a model id and OpenAI-style messages, with optional max_tokens and temperature. The chat tool returns the model, finish reason, latency, token usage, and the assistant text. It is backend-agnostic, so the base URL can point at any OpenAI-compatible API, with ComputeFlux as the default.
When to use it
Useful when an MCP host such as Cursor, Claude Desktop, or another MCP client should be able to query available models and run chat completions against an OpenAI-compatible endpoint without a separate SDK integration. It fits simple, non-streaming model calls rather than streaming or tool-calling workflows.
Requirements
Runs locally as a stdio process on Node.js 18 or newer; the server itself needs no npm install. Requires the COMPUTEFLUX_API_KEY environment variable, and optionally COMPUTEFLUX_BASE_URL to override the default endpoint. Network access to the endpoint is needed. Desktop only.
Before you install
The server sends prompts and messages to the configured endpoint, so conversation content leaves the machine. It requires the COMPUTEFLUX_API_KEY bearer credential, which should be supplied through the environment rather than committed to configuration files. The default endpoint is a testnet gateway, so treat results and availability accordingly. Calls may incur usage costs depending on the provider.

Installation

In SourceWeft

  1. Open mcp-computeflux in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

mcp-computeflux

Zero-dependency MCP (Model Context Protocol) server. Give any MCP host — Cursor, Claude Desktop, Windsurf, any MCP client — one OpenAI-compatible endpoint, and call models on ComputeFlux as plain tools: list what's available, then run a chat completion.

Your MCP agents hit flash-tier models over one OpenAI-compatible endpoint.

  • ~175 lines of Node 18+, no npm install to run the server (only the optional client/ e2e harness pulls SDK deps).
  • Backend-agnostic: point it at any OpenAI-compatible API — ComputeFlux is the default, not the only target.
  • ComputeFlux is an OpenAI-compatible, TEE-verifiable multi-model inference gateway deployed on the Polkadot testnet.

Tools

ToolWhat it does
computeflux_modelsLists the model ids served by the endpoint.
computeflux_chatRuns a non-streaming chat completion (model + OpenAI messages; optional max_tokens, temperature). Returns model, finish reason, latency, usage, and the assistant text.

Configuration (env)

VariableRequiredDefaultNotes
COMPUTEFLUX_API_KEYyes—Bearer key for the endpoint.
COMPUTEFLUX_BASE_URLnohttps://api.computeflux.ai/v1Any OpenAI-compatible base URL.
bash
export COMPUTEFLUX_API_KEY="your_c***_key"# export COMPUTEFLUX_BASE_URL="https://your-openai-compat-host/v1"   # optional

Install

Published on the official MCP Registry (.mcpb bundle, no npm step):

io.github.computeflux2026isgod/mcp-computeflux — listed on the official MCP Registry (search by this name).

Any MCP client / agent that supports registry install (mcp add, Cursor, Claude Desktop, Claude Code) can pull it by name. Direct git-clone quickstart:

bash
git clone https://github.com/computeflux2026isgod/mcp-computefluxexport COMPUTEFLUX_API_KEY="your_...key"node /path/to/mcp-computeflux/server.mjs   # stdio JSON-RPC

mcpServers snippet:

json
{  "mcpServers": {    "computeflux": {      "command": "node",      "args": ["/path/to/mcp-computeflux/server.mjs"],      "env": { "COMPUTEFLUX_API_KEY": "your_c***_key" }    }  }}

Local e2e (optional)

client/ contains a tiny harness using the official MCP SDK — the exact path Cursor / Claude Desktop use. It spawns the server over stdio, lists tools, and runs one chat completion with a unique-token round-trip fidelity check.

bash
cd client && npm install && npm run e2e   # expects COMPUTEFLUX_API_KEY in the .env it points to

Transport

Newline-delimited JSON-RPC 2.0 over stdio. Only protocol messages are written to stdout; logs go to stderr.

License

MIT

Source: README.md at commit c4a0bcd

Tools

0
Tool metadata has not been indexed yet.

Version history

1
  1. v1.0.0LatestOct 4, 2026