Agent capabilities, multi-repository delivery, and local agent setup

This week gives teams one place to manage the skills, MCP servers, and plugins available to agents. Workflow delivery is also more reliable across several repositories, while Desktop makes local agent setup easier to reuse and diagnose.

Control what agents can use

The new Capability Catalog brings Agent Skills, MCP servers, and plugins into workspace settings. Workspace admins can inspect a source, approve or reject a verified version, choose who can attach it, and manage its credentials without placing secrets in agent configuration.

Capabilities can be made available, required, or disabled for an agent and narrowed again for a workflow or individual step. A required capability blocks launch when its version, credential, repository evidence, or runtime support is unavailable, with a clear path to reconnect or ask an admin for help.

Each session now records what it received. The Environment view separates workspace-managed capabilities, local capabilities observed by the worker, and anything excluded by policy, making it easier to understand the tools and context behind a run.

Complete work across several repositories

Workflows that change more than one repository now keep a separate branch and delivery result for each repository. Every changed repository can produce its own pull request or merge request, and recovery actions stay tied to the repository that needs attention.

When starting work, the branch picker can suggest starting points from parent, ancestor, related, and recent work. You can also search the full branch list or enter a branch manually.

Reuse a local agent setup

Desktop can duplicate a local agent from its details page. The new agent carries over its runtime, repository folder, concurrency, pull request behavior, instructions, hooks, MCP preference, and visual settings, then starts from its own details page.

Desktop startup no longer waits on a slow login shell before showing a window. A loading screen appears while the shell path is resolved, and a visible warning explains when Desktop has to fall back to a minimal path.

Improved

• Continue, Retry, Skip, and Cancel use the same availability rules across workflow steps and run menus
• Unavailable workflow recovery actions explain why they cannot run instead of disappearing
• Backlog and Roadmap empty states appear correctly in workspaces that only contain completed onboarding work
• Agent session reports show the runtime, model, repository, permissions, plugins, skills, and MCP servers used for the current attempt

Fixes

• Multi-repository workflows no longer finish with commits and a pull request for only the first repository
• one auth login exits cleanly after authentication instead of hanging
• Code steps can recover from inherited generated files in a workflow worktree
• Stale local workers stop retrying claims when their runtime revision is no longer accepted
• New capability access is no longer hidden by an older cached billing decision