LM Studio's Bionic agent adds sandboxed cryptography operations, safer shell-command reviews and file-deletion prompts; changelog names no insurance terms

LM Studio shipped two dated updates to Bionic, its local AI agent, in the first half of August: version 1.0.6 on Aug. 10 and version 1.0.5 on Aug. 7, according to the company's changelog.

The Aug. 10 release lets the agent "use sandboxed (Pyodide) operations for encryption, signatures, and other cryptography tasks," running that code in an isolated environment rather than directly on the host machine. The same version brought "more reliable automatic shell-command reviews during long sessions," a check LM Studio runs before executing terminal commands the agent proposes.

Three days earlier, version 1.0.5 added "permission prompts for destructive file deletion operations," requiring explicit user confirmation before Bionic deletes a file, plus "automatic context sizing for local MLX models based on available memory" and "faster processing for large files, with support for files up to 100 MiB."

Bionic runs against locally hosted open-weight models. LM Studio says it lets users "run Bionic with the latest frontier open models such as GLM 5.2, Kimi K3, and DeepSeek V4 Pro," and that it is "available in initial preview for macOS and Windows," according to the company's product page as captured in AI Dispatch's index. The company describes its change tracking in terms of file edits, not agent decisions: "every change is automatically saved, so you can work with your agent freely."

The Aug. 7 and Aug. 10 changelog entries name no insurance carrier, coverage limit or indemnity term, and neither describes how the sandboxed cryptography operations are audited after the fact.