mcp-installer

Manage MCP servers

Use the bundled {SKILL_DIR}/mcp_apply.py for daemon and configuration operations. Do not construct raw daemon requests or edit mcp.json by hand while the helper can perform the operation. Run the helper with --help only when the exact subcommand or argument is not already known.

State machine

1. RESOLVE

Identify the exact server name, transport (http or stdio), endpoint or command, required arguments, environment variable names, target agent, and whether all or selected tools should be wired. Inspect current state first:

python3 "{SKILL_DIR}/mcp_apply.py" status [name]

Do not invent package names, endpoints, headers, OAuth mode, commands, tool names, or agent targets. Ask only for a missing value that cannot be derived from the provided server documentation or current configuration.

2. SAFETY

Use ${ENV_VAR} references for secrets. Never print, request in chat, embed in command history, or persist a raw API token when an environment reference or OAuth flow is available. Direct confidential OAuth credentials, when required by the helper, must be stored through its supported credential flow rather than written into mcp.json.

Resolve relative filesystem arguments to absolute paths because stdio runners do not inherit the user's interactive working directory. Inspect existing agent references before removing or renaming a server.

3. MUTATE

Choose exactly one operation and run it once:

add NAME --http URL [--header KEY=VALUE] [--oauth ...]
add NAME --stdio COMMAND --args ... [--env KEY=VALUE]
update NAME <transport options>
restart NAME
connect-oauth NAME
remove NAME
apply

Invoke it as python3 "{SKILL_DIR}/mcp_apply.py" <operation> .... Treat exit codes as follows:

• 0: configuration applied, or safely written while the daemon was offline;
• 1: validation/API error—correct the reported input before retrying;
• 2: runner error or authentication required—inspect status and fix that condition;
• 3: readiness wait timed out—report the current state instead of looping.