A Cloud Agent for the servers you own
Investigate it. Change it. Verify it.
Every workspace has one durable Cloud Agent that can work across the servers you authorize. It reads, edits, builds, uploads, runs commands, and keeps useful memory. The dashboard, Cloud Agent, API, and MCP all cross the same permission gate and land in the same activity trail. Server credentials never reach the model.
What the Cloud Agent did is exactly what you see. Its tool calls and your clicks cross the same gate and land in the same record.
How it works
Connect what you own. Hand over the tools. Keep the record.
01 · connect
Connect what you own
Point Serverwright at a game panel or anything that speaks SFTP. The host itself decides which tools exist: a panel gets console, commands, and power. A bare box gets files and search. Nothing is faked, and nothing sits greyed out.
02 · operate
Hand your agent the tools
The Cloud Agent receives exactly the tools the current operator and host allow. Ask investigates without changing resources. Edit can act without per-tool approval interruptions. External agents can use the same public MCP surface.
Connect this agent to Serverwright, which provides tools to operate my remote servers. 1. Add a remote MCP server using this environment's native MCP setup: - Name: serverwright - Transport: Streamable HTTP - URL: https://mcp.serverwright.app/mcp 2. Authentication is OAuth. Start the connection's sign-in flow. If it opens a browser, I will approve it there. If it prints an authorization URL, show it to me so I can open it. Do not ask me to paste a token or key. 3. If this environment does not support MCP OAuth, stop and tell me. I can create a Serverwright API key for you to send as an `Authorization: Bearer <key>` header to the same URL. 4. Once connected, call the `servers` tool and report which servers and tools I can reach.
03 · on the record
Keep the record
Every action is written down with its actor, what it touched, and how it ended, and rows only append. Access is granted per operator and revoked in one move, across MCP, API, CLI, and the dashboard at once.
Where it points
Built for the boxes you actually have
Minecraft and game servers
Let it read the crash log at 2am. Plugin conflicts, config drift, restarts: your agent debugs the server the way it debugs code, inside the panel's own permissions.
ExploreRaspberry Pi and homelab
The headless box finally has hands. Services get inspected, configs get fixed, and processes get bounced from the chat you already have open.
ExploreNAS and home storage
Careful hands for the box that holds everything: reads before writes, scoped shares, and a full record of every touch.
ExploreAny remote machine
A VPS, a dedicated box, a server you administer for someone else: the same verbs and the same trail on all of them.
ExploreCustody
The key stays on our side of the line
Giving a model shell access means giving it the password. Serverwright is built so that trade never has to be made.
Credentials are held, not shared
The panel key or SFTP credential is stored encrypted in Serverwright's custody. Tools act with it. No model, transcript, or API response can read it back.
Scoped, and revocable in one move
Capabilities are derived from the host and narrowed per operator. Revoking access takes one click and cuts every path at once.
An append-only trail
Actor, action, target, outcome, on every call. Passive reads coalesce instead of piling up, so real actions never get buried.
The complete Cloud Agent starts free.
Free includes Fast, workspace memory, artifacts, every server tool, and the full trail. Paid plans add High, more included Agent usage, and seats for named operators. Your history and knowledge remain yours if the plan changes.