MCP servers
An MCP server (Model Context Protocol) extends your AI tools with external capabilities. cnfgr stores a single definition and writes it into each tool that supports MCP.
Anatomy of an MCP server
| Field | Meaning |
|---|---|
id | Stable identifier for the server. |
name | Display name. |
transport | How the tool talks to the server (e.g. stdio or HTTP). |
command / url | The command to launch (stdio) or the endpoint URL (HTTP). |
args | Arguments passed to the command. |
env | Environment variables — secret values held as placeholder references. |
headers | Request headers — secret values held as placeholder references. |
trust | Trust setting for the server. |
timeout | Request timeout. |
Secrets stay local
MCP servers often need an API key or token. cnfgr never syncs those values: in
env and headers, secrets are stored as placeholder references
(${secret:<id>}) and re-hydrated from each device’s local secret store when the
config is written. See Secrets & privacy.
Editing an MCP server
Open the MCP tab and select or create a server. Enter the transport, command/URL, args, env, and headers. Secret values you enter are kept on the device and replaced with placeholders everywhere they’d otherwise travel.
Fidelity across tools
Tools vary in their MCP support. cnfgr projects the definition into each tool’s format; see the tool support matrix.