The latest release adds a native, project-based chat interface, stronger multi-agent coordination and clearer controls for auditing local AI work on the Mac.
Mac MCP 2.3.1 marks a notable shift from a background automation tool towards a more visible control centre for local AI work on the Mac. The open-source project, which lets agents interact with files, browsers, shell commands and native apps under user approval, says the latest release adds a native chat interface and broadens support for coordinating multiple coding agents across a single project.
According to the project’s GitHub repository, the new Mac MCP Chat window is designed around project-based conversations rather than isolated command-line sessions. Users can open a chat from the menu or with a keyboard shortcut, choose a project folder and then work with supported providers including Claude, Codex, OpenCode, GitHub Copilot, Antigravity, Kilo Code, Kimi Code, Goose and Cline. Each provider keeps its own model settings and operating style, while the conversation history is preserved so work can resume after a restart.
The interface is built to make agent activity easier to audit. File changes appear as diffs, terminal output is shown in a readable form and individual steps are collapsed into compact cards that can be expanded when more detail is needed. Users can reply while a task is still running, queue the next instruction, attach files or images, and approve sensitive actions directly inside the chat. The release also adds better visibility into background work, with sidebar state, notifications and clearer status markers intended to reduce the need to hop between terminals and separate tools.
The update also tightens the way teams of agents coordinate. The project says agents now check configuration before creating anything, bundle multiple errors into a single report, preserve task dependencies more cleanly and pass completed work further down the chain with less context loss. A new history option for team messages gives users a fuller record of what happened during a run. The same release expands persistent memory use, allowing delegated agents to surface relevant skills and stored preferences before they begin searching, a change that may reduce repetitive setup but also raises the importance of privacy controls and data retention settings.
Security remains central to the release. The project notes that earlier 2.3.0 updates changed delegated Codex agents so they start with applications disabled by default, after concerns that broader account connectors could create an unintended route to command execution outside the intended scope. It also added safeguards around browser uploads, including confirmation prompts and checks for secrets in file contents. For 2.3.1, installation options include an Apple Silicon package, a terminal script and an npm installer, with support varying by macOS version and chip architecture. The Apple Silicon package is not signed with an Apple Developer ID and is not notarised, so macOS may warn before first launch; the project also publishes a SHA-256 checksum for users who want to verify the download before installing it.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
