Private · ProductionPythonFastMCPMCPSandboxesGitHub
NiPlex-MCP
One MCP connection that gives an AI agent 73 tools. A production Model Context Protocol server. Any MCP-capable AI client connects once and gets repository work, cloud sandboxes, remote file operations, web and YouTube search, Gemini analysis and persistent chat threads. The source is private; this page describes what it does.
Tools73
AuthFail-secure API key
DispatchAllow-listed
SourcePrivate
What it does
73 tools in seven groups
- GitHub, 38 tools: repositories, files, branches, commits, issues, pull requests, CI status and code search.
- ThreadForge chat, 12 tools: persistent multi-model threads and shared artifacts, proxied from ThreadForge-MCP.
- Remote files, 8 tools: list, read, write, rename, search and delete files on a remote host over SFTP.
- Web and video, 5 tools: web search with fallbacks, page reading, and YouTube search, details and stats.
- Sandboxes, 4 tools: run code in two cloud sandboxes through one interface.
- Gemini, 2 tools: ask Gemini, or analyse a YouTube video straight from its URL.
- Utilities, 4 tools: keyword help across all tools, installed versions, recent errors and a helper.
Security posture
- Fails closed: if no API key is configured, the whole server answers 503 instead of running open.
- Every call goes through an explicit allow-list. Unknown tool names never reach implementation code.
- There is no email-sending tool and no auto-merge. Nothing is destructive by default.
- Each domain is its own manager class, so one tool group cannot leak into another.
How a request flows
ConnectAn MCP client such as Claude connects to the server endpoint with an API key.
CheckRequests without a valid key are rejected before any tool runs.
DispatchThe tool name is matched against the allow-list and handed to its manager.
ActThe manager calls GitHub, a sandbox, a remote host, a search backend or Gemini.
ReturnThe result goes back to the agent as a normal MCP tool response.
Boundaries
- The repository is private. It is not open source and is not available to install.
- Security details are reported by the owner. They are not an independent audit.
- Staged tools for other providers exist in the codebase but are not exposed.