Flux Cloud

Flux Cloud MCP server

An MCP server that lets any AI agent use Flux Cloud, the decentralized cloud: quote an app in US dollars, deploy it, pay for it, watch it come up, read its logs, update it, and cancel it. Works with Claude Code, Claude Desktop, Cursor, Windsurf, OpenCode and every other MCP host.

There is no account. Two keys are the whole identity:

The server generates a pair for you (flux_generate_keys); fund the payment address with FLUX and deploy.

Pricing policy

Every price this server shows is the Flux Cloud USD price, the same one home.runonflux.io charges, converted to FLUX at the live market rate with the 5 percent pay-in-FLUX discount. Quotes come from the network itself (/apps/calculatefiatandfluxprice), so they include the $0.99 minimum, hardware and term discounts, and update credits.

The blockchain's own acceptance threshold is several times lower. It is never shown as a price; it is only checked as a guard before a payment is broadcast.

Install

npm install -g @runonflux/flux-cloud-mcp     # or: yarn global add @runonflux/flux-cloud-mcp

Or run it without installing: npx @runonflux/flux-cloud-mcp.

Claude Code

claude mcp add flux-cloud -s user \

-e FLUX_ID_PRIVATE_KEY= -e FLUX_PAYMENT_PRIVATE_KEY= \ -- npx -y @runonflux/flux-cloud-mcp

Claude Desktop, Cursor, Windsurf (JSON config)

{

"mcpServers": { "flux-cloud": { "command": "npx", "args": ["-y", "@runonflux/flux-cloud-mcp"], "env": { "FLUX_ID_PRIVATE_KEY": " ", "FLUX_PAYMENT_PRIVATE_KEY": " " } } } }

Keys are optional. Without them every read-only tool works, and flux_generate_keys will create a pair to put into the config.

OpenCode

{

"mcp": { "flux-cloud": { "type": "local", "command": ["npx", "-y", "@runonflux/flux-cloud-mcp"], "environment": { "FLUX_ID_PRIVATE_KEY": " ", "FLUX_PAYMENT_PRIVATE_KEY": " " } } } }

Tools

Resources flux://guide/overview, flux://guide/spec-format, flux://guide/pricing and flux://guide/gotchas give the agent the domain knowledge; the deploy_on_flux prompt walks it through a deployment.

A typical session

> deploy nginx on flux, 3 instances, one month

flux_get_identity -> Flux ID 1Ab..., payer t1Cd..., 996.98 FLUX ($48.74) flux_build_spec -> v8 spec, port 39978 -> 80, r:/data, valid flux_quote_app -> $0.99 for 1 month, pay 19.24 FLUX (5% FLUX discount) user agrees flux_deploy_app confirm=true -> message hash 8d2f..., paid 19.24 FLUX, txid 0c41... flux_wait_for_app -> accepted at height 2941560, 3/3 instances running https://myapp.app.runonflux.io

How a deployment works

1. The spec is verified through api.runonflux.io exactly as it will be at registration (image reachable, architecture, ports, name free). If the balancer refuses, a few healthy nodes are probed and tried in turn.
2. The network quotes the USD price; the server checks the payer's balance.
3. With confirm=true the Flux ID signs the spec, the same endpoint broadcasts it to the network and returns a 64-character message hash.
4. The payment address sends the quoted FLUX to the network deployment address with the hash in an OP_RETURN output. Right before signing, the amount is re-checked against a freshly fetched price table, because an underpaid message is dropped silently and the FLUX is not refunded.
5. Nodes pair the confirmed payment with the message and publish the app. Instances then spawn and pull the image.

Private (enterprise) apps publish an empty component list, so flux_get_app, flux_get_app_logs and flux_get_app_stats fetch the decrypted specification from an ArcaneOS node with the owner key when it is configured or passed. Only component names, images, ports and sizes are returned; environment variables, commands and registry credentials never leave the server.

Updates use the same tool: an existing app of the same owner gets a fluxappupdate message and the network credits the unused part of the old term. Private (enterprise) apps are supported through the enterprise argument: components are encrypted for the network and only ArcaneOS nodes can run them.

Hosted server: mcp.runonflux.com

The same server also runs on Flux Cloud itself, over Streamable HTTP, for hosts that cannot start a local process (claude.ai connectors, ChatGPT, browser and mobile agents). Add it as a remote MCP server:

https://mcp.runonflux.com/mcp   (also https://mcp.runonflux.io/mcp)

The hosted server holds no keys. Read-only tools need none. Tools that sign or pay take fluxIdPrivateKey and paymentPrivateKey as call arguments, used in memory for that call and never logged or stored. The server's instructions tell agents to create a dedicated pair with flux_generate_keys and to have the user fund only what a deployment needs, rather than asking for the keys of a wallet used elsewhere. Anything typed into a chat passes through the AI vendor and the host's history, so treat such a pair as disposable and keep its balance small. For larger budgets, run the local server, where keys never leave your machine.

Run your own copy with Docker:

docker run -p 3000:3000 runonflux/flux-cloud-mcp

It is stateless, so any number of instances can sit behind one load balancer. GET /healthz reports liveness. The Flux app specification that runs the public instance is in deploy/cloudmcp.json.

Configuration

Security notes

• Private keys stay in the server process. No tool ever returns a configured key; flux_generate_keys returns only the keys it just created.
• Nothing is spent unless a tool is called with confirm=true. Hosts that ask before tool calls will show the USD amount in the arguments' context.
• FluxOS sessions are self-issued signed phrases valid for a few hours; no password or token is stored anywhere.
• Use a dedicated payment address holding only what you intend to spend.

Development

yarn install

yarn type-check && yarn lint && yarn format:check && yarn build && yarn test yarn test:live # differential test of the USD estimator against the network

The USD estimator, the specification formatter and the consensus guard mirror the FluxOS code (appSpecHelpers.js, appUtilities.js, messageVerifier.js) and the live test checks the estimator against the network on every run.

License

MIT