GitHub Repo
AutoGPT 0.11.3 Adds a Remote Local PC Execution Target, Binding Agent Access to a Machine and Folder
On November 15, AutoGPT released platform beta 0.11.3, adding a Local PC execution target that connects chat tasks to a paired computer. The feature is controlled by a gradual rollout flag; local-model routing and workflow recording are still unavailable.
In autogpt-platform-beta-v0.11.3, released on November 15, AutoGPT added a remote Local PC execution target. Eligible accounts can bind a new chat to a paired computer and a specified folder; the cloud remains the default execution environment. The release is available, but that does not mean every account can connect to a local machine right away. Release announcement
AutoPilot could already create agents through conversation, call platform blocks, and use supported code execution tools. This change brings the execution location into the chat state: when an agent works with files, runs commands, or uses some computer-control tools, those operations can be handled through the local companion app. The engineering significance is that the same conversational tool interface can now connect to a user's working environment. AutoPilot documentation, implementation details
Access control is also more granular. The companion app initiates an outbound control connection, while folder selection creates a signed authorization bound to the machine, chat, version, and filesystem identity. The data channel must match the saved execution target. If authorization expires, is revoked, or the identity does not match, the system rejects the operation and does not automatically rerun it in the cloud. This makes it easier to determine which environment a failure occurred in. Pull request
Deployment teams still need to distinguish “local execution” from “data stays local.” The maintainers explicitly state that the platform deployment can see intermediate prompts and results. Local-model support remains preparatory work and is not yet a supported chat mode. Workflow recording also remains disabled, so this release does not mean desktop operations can be recorded to automatically generate usable skills. Feature limitations
The next things to watch are the rollout scope, compatibility across companion-app versions, and how authorization and sessions recover after a disconnection. For teams that want agents to work directly on local projects, this release provides a concrete integration path; before adopting it, they should still verify directory authorization, revocation, and reconnection behavior.