One extension library
Extension profiles give way to one list of MCP servers and skills, each limited to the repositories and agents you choose. Enabled items apply automatically to matching task runs, pull request follow-ups and automation runs; the disable-mcps, disable-skills and disable-extensions labels switch them off for one issue or pull request. An HTTPS server can sign in with OAuth from Settings.
Agent extensions
Local MCP servers from npm packages
An administrator adds an npm package at an exact version, with its arguments, variables and write-only secrets, and ProPR pins its integrity value. The worker installs the package once into its own cache, outside every task workspace, and each run mounts it read-only, checks the files against the pinned value and starts it for Claude Code and Codex inside the task container. Secret variables are readable there, so saving a server that has them asks for an acknowledgement. Python packages are not supported yet.
How agent extensions work
A Claude goal asks you a question
Claude goal sessions get an ask_operator tool: one question, optionally with suggested answers, shown under Needs you. Your next goal input is the answer. Pause and cancel work while it waits, and after 24 hours by default Claude is told nobody answered. Ordinary task runs never get the tool.
How Goals work
MCP plans hold 50 issues
create_plan, update_plan and publish_plan accept up to 50 plan issues, and implement_plan starts up to 50 per call, the same size Planner Studio always allowed.
MCP from chat
Reviews in the PR discussion
get_pull_request_discussion pages a pull request's submitted GitHub reviews and its inline review comments, with file, line and reply thread, next to the conversation it already read.