JARVIS UltraManualv1.31

Concepts

The tools lock

The tools lock decides what each agent may do with your connected apps and MCP servers. It checks every call an agent makes to them, before the call runs.

What it checks

When you connect an app or add an MCP server, you choose which agents may use it and with what access (Connect apps, Add an MCP server). The dashboard writes those choices into each agent's profile, and sets the lock up in front of the agent's tools. Before each call to an app or a server, the lock asks two questions: may this agent use this app at all, and is this call a read or a change?

Access A read A change
Read only Runs Refused
Changes need your approval Runs Stopped; the exact call goes to your Decision queue as a Tool call card
Full access Runs Runs

A call to an app the agent was not given is refused.

Reads and changes

The lock reads a call's name. A call is a read when its name has a reading word (get, list, search, read, find, fetch and the like) and no change word (create, update, delete, send and the like). So "list messages" is a read, and "delete task list" is a change. When in doubt, the lock treats a call as a change: a wrong guess only ever sends a read to you for approval, never lets a change through. For an MCP server, a tool the server itself marks as read-only also counts as a read.

Approving a stopped call

A Tool call card shows the app, the tool and the exact call. When you approve it, the agent is told to run that call now, and the lock lets that agent's changes through for 15 minutes. Then the lock closes again. Each change is its own card, so agents are told to put as much as they can into one call.

If the same call is stopped again within 10 minutes, no second card is made.

When it is on

The lock is set up in every agent's profile once there is something to guard: a Composio key saved, an app connected or an MCP server added. It stays set up after that. A server you add to an agent's config.yaml by hand, outside the dashboard, is not guarded.

What it is not

The lock is a guard rail for an agent's mistakes, not a wall against a determined program: it sees the calls agents make through the dashboard's app launcher and MCP servers, and the agents' instructions tell them to use those. Give Full access only where an unwanted change would be easy to undo, and start new apps on Changes need your approval.