Cursor CLI checks MCP resource saves and connects local worker servers Cursor shipped a CLI update on November 17, 2026 that routes MCP resource downloads through the same permission checks and approvals as ordinary file edits, refusing paths starting with ~, using a URL scheme, or containing .., and blocks other processes running under the same user account from reading the CLI's memory, environment, or open files on Linux. The release also lets self-hosted workers run URL MCP servers from ~/.cursor/mcp.json whose address is localhost, 127.0.0.1, or [::1], and adds a privateTmp sandbox.json option that maps /tmp inside sandboxed Linux commands to the workspace's own scratch folder. Cursor said admin prompt hooks from the team dashboard now run with their prompt and model in the CLI instead of as an empty command, and that agent worker --help now documents that workers exit with code 0 when released, recommending supervision via systemd, a Windows scheduled task, or a Kubernetes Deployment. CLI Cursor CLI changelog The latest features, improvements, and fixes shipping to Cursor CLI. Run agent --version to check your installed version, and agent update to upgrade in place. Agent nov-17-2026-agent - Admin Prompt hooks from the team dashboard now run as prompts. Prompt hooks a team admin sets up in the dashboard https://cursor.com/docs/hooks cloud-distribution-enterprise-only now run in the CLI with their prompt and model, as they do in the desktop app. Before, the CLI ran each one as an empty command, so the prompt never ran. Dashboard command hooks with no script are now skipped. - Sandboxed commands can use /tmp on Linux. Set "privateTmp": true in sandbox.json https://cursor.com/docs/reference/sandbox , with readBoundary set to "workspace" , and /tmp inside sandboxed commands becomes the workspace's own scratch folder, the same one $TMPDIR points to. Tools that write to a literal /tmp path stop failing, files stay there for the next command, and other workspaces can't see them. macOS ignores the setting. - Skip login scripts in the agent's shell. Set CURSOR LOGIN SHELL=0 and the agent starts its shell without running login scripts such as /etc/profile , ~/.bash profile , or ~/.zprofile , so they can't reset your PATH . Your rc file, such as ~/.bashrc or ~/.zshrc , still runs. Leave it unset to keep the current behavior. Security nov-17-2026-security - Other programs running as you can no longer read the CLI's memory on Linux. The CLI now blocks other processes under your user account, including commands the agent runs, from reading its memory, environment, or open files, or attaching a debugger to it. This keeps the credentials the CLI holds out of their reach. To attach a debugger or tracer, set CURSOR AGENT ALLOW PROCESS INSPECTION=1 before starting agent . - MCP resource downloads follow your file-write rules. When the agent saves an MCP https://cursor.com/docs/mcp resource to disk, the write goes through the same permission checks and approvals as any other file edit, and it must land inside your workspace. Paths that start with ~ , use a URL scheme, or contain .. are refused, and Ask and Plan modes no longer save resources. Self-hosted workers nov-17-2026-self-hosted-workers - Self-hosted workers can use MCP servers running locally on your machine. agent worker now runs URL servers from ~/.cursor/mcp.json whose address is localhost , 127.0.0.1 , or ::1 , such as a desktop app's local MCP endpoint, so agents on the worker can call their tools. Cursor's backend can't reach those addresses. Other URL servers still run on the backend, as described in MCP servers https://cursor.com/docs/cloud-agent/self-hosted/my-machines mcp-servers . - agent worker --help explains when self-hosted workers exit. A new Lifecycle section says a worker exits with code 0 whenever Cursor releases it: its agent was archived, killed, or expired, its claim went stale, or it sat idle past --idle-release-timeout . It recommends running each pool worker under a supervisor, such as a systemd unit, a Windows scheduled task, or a Kubernetes Deployment, and restarting it with agent worker --pool