Choose the tools to expose to AI — pick on screen, approve in-house

Why separate proposal from approval?

A WireCanal MCP canal starts with all tools denied by default. The allowlist that decides which tools the AI may see lives only in wirecanal.json, the local file next to the Agent running inside your network — the cloud side does not hold it. So even if something happened to the dashboard, which of your internal tools are visible to AI would not change one bit.

That said, hand-writing the allowlist every time is tedious. So the dashboard provides a "Tool permissions" editor. Here is how it works:

• check_circleWhat you can edit and save on screen is only a proposal — "this is how I would like to allow things". Nothing changes the moment you save
• check_circleTo apply it, you run one approval command on the machine where the Agent runs. It shows you the diff before writing to the local wirecanal.json
• check_circleThe screen shows "Applied / Pending changes", so any gap between the proposal and the actual configuration is visible at a glance
• check_circleBecause you can state flatly that "the cloud settings screen can never change what is exposed internally", it is a structure that passes corporate approval and security reviews

Step 1: Pick the tools on screen

Open the canal's "MCP connection" tab and you will find the "Tool permissions (which tools the AI may see)" card.

Click "Edit proposal" and the tool names the connected Agent has detected from your in-house MCP appear as checkboxes. Just check the ones you want to expose. You can also add names by hand, or choose options such as "deny tools with destructive names (delete/drop, etc.) even if they are on the allowlist".

Step 2: Saving does not apply anything yet (this is the point)

Click "Save proposal" and the card changes to "Pending changes". At this point, nothing has changed in what the AI can see. In fact, if the AI calls a tool that is in the proposal, it is still denied.

Step 3: Approve it in-house

On the machine where the Agent runs, execute the approval command. It fetches the proposal from the cloud, shows the diff against the current configuration, and then writes it to the local wirecanal.json (without -yes, it asks for y/N confirmation).

$ wirecanal apply-policy -config wirecanal.json
wirecanal apply-policy: diff between the proposal for canal 72jo and the current configuration:
  add to allow: get_sales_summary
  add to allow: search_products
  deny_destructive: false -> true
Apply these changes to tools in wirecanal.json? [y/N]: y
wirecanal apply-policy: applied. Effective from the next Agent start

After approving, restart the Agent. Only then do the selected tools become visible to the AI.

Step 4: Confirm "Applied"

When the restarted Agent connects, the card changes to "Applied (matches the Agent's effective configuration)". Only the allowed tools respond to the AI; any other tool is refused before it ever reaches your in-house MCP (you can also choose the option to hide them from the tool listing).

Ongoing operations

• check_circleTo add more tools: the same 2 steps — check them on screen and save, then approve in-house. Even in a hurry, the order never changes (that is where the assurance comes from)
• check_circleTo withdraw a proposal: "Withdraw proposal" removes only the proposal. It does not affect the configuration already applied on the Agent side
• check_circleIf the Agent is old: automatic tool-name detection and the applied-state display require Agent v0.17 or later. With older Agents, you can still edit proposals by hand and approve them

If you prefer to configure this programmatically, the same can be done through the public API (tool permissions endpoints). For using it with your whole organization, see Let your organization's members use it (organization IdP integration); for the overall design philosophy, see Security.