Claude Agent SDK

Create one Claude Agent SDK client for the run you intend, and hand it a stdio MCP configuration that names the Kaleidoscope tools explicitly. One client, one process, for the whole run.

Connect it

kaleidoscope config --profile default --json

Put that descriptor in the SDK’s own stdio MCP configuration. One client, opened around the run and closed at the end — not a fresh engine process per tool call.

What it writes

Nothing on your disk. The descriptor sets an empty environment for the child process, so account tokens, vault locations and provider keys never reach it. Your model-provider credentials stay in your own application.

The one thing that is different here

The tool allowlist is two names, and you write them out:

mcp__kaleidoscope__search
mcp__kaleidoscope__remember

Anything else is not a Kaleidoscope tool. Operator commands are never offered to a model, so there is no third name to add.

Disconnect

Close the client. Remove the descriptor from your SDK configuration; nothing else was changed.