Skip To Content
Get Early AccessSign In

Extensions

Your tool already exists.
Let Iris reach it.

You have a local MCP server. It needs a command, arguments and an environment setting. Adding a name alone cannot make that server usable.

mcp.jsonIllustrative Situation
Commandnpx
ArgumentsPackage + server mode
EnvironmentExplicit key/value settings
ExecutionStill subject to policy
You Ask Iris

“Connect this server on my paired computer. Show me what is still missing.”

Installed. Authenticated. Permitted. Different things.

A useful setup flow separates what starts the server, what authenticates it and what the agent is allowed to do. A skill can explain a task; it cannot grant the missing authority.

Try The Situation

Interactive simulation. Fixed sample data. No real actions.

Back In Your Own Work

What Changes
When You Use This?

Before you try it for real.

The example explains the behavior. Your actual tools, account and evidence determine what is possible.

What You Need

A compatible server, its required runtime and credentials, and a paired computer for local stdio. Never put secret values in a shared skill or Matrix source.

What Is Enforced, Recorded Or Advised

The platform rechecks enabled resources and execution scope. Authentication does not override tool policy, and adding a package does not approve its actions.

Where It Stops

MCP servers differ in protocol, authentication and output. Local servers need a connected host and their own dependencies. A third-party server can return inaccurate or malicious content; review its access.

The Workflow, Step By Step
  1. Choose The Connection

    Add a supported remote MCP endpoint or a local stdio configuration through a paired computer. Supply required arguments, environment variables and supported authentication.

  2. Inspect What Is Available

    Discover the server's tools and load the schemas needed for the task. Skills supply reusable instructions; they do not grant credentials or new authority.

  3. Use And Review

    Enable the connection for the intended agent and task. Inspect tool results, errors and any required approval before depending on the outcome.

Bring Your Own Version Of This Problem.

There is work
behind that question.