user_algorithms/

1. Validate — the source is AST-checked against a conservative denylist:

no network, no subprocess, no filesystem escape hatches. Accepting arbitrary generated code is the risk here, so this runs before anything touches disk (validation).

1. Store — accepted source is written to

~/.tuiml/user_algorithms///algorithm.py, versioned, so an edit never destroys the version that worked (storage).

1. Register — the file is imported and registered under both a versioned

alias (MyAlg_v1_0_0) and a bare latest alias (MyAlg), which is what makes tuiml_train, tuiml_benchmark and tuiml_describe work on it with no special-casing (registration).

1. Reload — load_all() rescans the directory at import and

re-registers everything, so an agent's work survives an MCP server restart.

• research_log — experiment history per

algorithm version, so an agent can see what it already tried.

• sources — reading, searching and

editing algorithm source, user-authored and built-in.

The denylist is a guard against an agent's mistakes, not a sandbox. Source that reaches here still runs in your interpreter with your permissions.