flux

Your AI agents have access to 10,000+ MCP servers. They should only see five.

[](https://github.com/fluxverseai/flux/actions/workflows/ci.yml) [](https://pypi.org/project/fluxverse-cli/) [](LICENSE)

---

The MCP ecosystem has exploded — 10,000+ servers, 60,000+ skills, and counting. But there's no way to manage them. API keys sit in plaintext config files. Every project gets the same 50 MCPs dumped into its context, whether it needs them or not. Your research agent can accidentally call your trading API. Setting up a new project means copy-pasting .mcp.json and praying nothing leaks.

Flux is the control plane that sits between your MCP ecosystem and your AI agents. One registry. Per-project scoping. Secrets in your OS keychain — never in files. Sandboxed execution. Full lifecycle management for every tool your agents touch.

> Think package.json for AI tooling — declare what each project needs, and Flux handles the rest.

Without Flux, managing MCPs at scale looks like this:

| Pain | What happens | |------|-------------| | Configuration sprawl | 20 projects × 50 MCPs = 20 hand-maintained .mcp.json files. Add a new MCP? Edit every project. | | Credential leakage | 48% of MCP servers recommend storing API keys in plaintext. Keys end up committed to git, leaked in logs. | | No scoping | Every agent sees every tool. A coding assistant gets your home automation MCPs. A wiki bot gets filesystem write access. More tools = more noise = worse outputs. | | No reproducibility | You can't git clone a project and get its MCP setup. There's no npm install for AI tooling. | | Skills are unmanaged | 60,000+ Claude Code skills exist as files you manually copy between machines. No versioning. No scoping. No package