Tools and permissions
An agent uses tools supplied by its native Codex or Claude Code configuration. CatalEx’s supported tools controls select discovered MCP servers, terminal access and approval mode for that agent. Availability depends on the installed provider and release.
Inspect before changing
catalex agents tools --agent AGENT_ID --json
catalex agents tools --agent AGENT_ID --discover --json
Discovery checks tools under existing trust. It does not sign in, connect new servers or grant trust. Read the desired policy, discovery inventory and actual native/session evidence separately.
| Control | Choices | Why change it? |
|---|---|---|
| MCP servers | Complete selection of discovered optional server IDs | Expose only the tools the role needs |
| Terminal access | inherit, off, on | Keep the native policy, or request a supported terminal restriction/change |
| Approval mode | inherit, manual, automatic | Keep native behavior or request the provider’s supported approval mapping |
| Required communication | Retained by CatalEx | Let the agent receive and answer conversations |
Native denials, mandatory approvals, package hooks and platform restrictions remain binding. “Automatic” does not authorize bypassing them. Folder isolation is a separate setting; see folders and tools.
Save a reviewed selection
Use the exact saved revision and inventory revision returned by inspection. CHANGE_ID is a new stable operation ID for this particular change. Review the full optional-server list: omitted optional servers are not retained automatically.
catalex agents tools --agent AGENT_ID --operation-id CHANGE_ID --revision REVISION --inventory-revision INVENTORY_ID --mcp-server SERVER_ID --terminal-access off --approval-mode inherit
Replace the uppercase placeholders with the inspected values. Use repeated/comma-separated --mcp-server IDs for several servers, or --mcp-server '' to disable all optional servers. The command saves the complete requested policy; adding --apply submits it for application.
A save fences affected execution until the policy can be confirmed at a safe start/resume boundary. Check effective native evidence before claiming it is active. Offline, pending, failed or unsupported evidence is not success.
After a lost reply, retry the same operation ID and identical flags. After a confirmed failure or pending result, refresh and review state before deciding on a new operation. A stale inventory or policy revision requires a new review, not a forced write. Test both a permitted action and a deliberately refused one after application.