Index
Features
Long-running agents
SpaceXAIStays with multi-step work: research, analysis, codebase edits, and turning a product idea into a working first version.
Self-testing and verification
SpaceXAIOn longer trajectories, checks its own work before moving on.
Own cloud computer
SpaceXAIEach Bot runs on a persistent cloud VM with a browser, filesystem, and terminal. All Bots on an account share that user-scoped computer (files, browser sessions, logins), with a separate screen per Bot.
Always-on runtime
SpaceXAIKeep working 24/7 in the cloud after the client is closed.
Routines
SpaceXAIA routine tells one Bot when to run a workflow — on a schedule or, where supported, after an event. Background routines can run while the laptop is closed. A Bot can own up to 50 routines.
Approvals
SpaceXAIBots only pull you in for judgment calls. They follow up on dropped threads and nudge stalled handoffs.
Plan mode
SpaceXAIThe agent explores and drafts a plan before edits. Only the session plan file may be edited until you approve; bash is not gated. Auto and always-approve do not skip plan review. Enter with /plan or Shift+Tab; the agent can enter plan mode on its own. Reopen with /view-plan.
Auto and always-approve
SpaceXAIAsk (default) prompts for anything not already allowed. Auto uses a classifier to auto-approve safe tools. Always-approve skips permission prompts; deny rules and PreToolUse hooks still apply. Shift+Tab cycles Normal → Plan → Auto → Always-approve. Allow/deny rules in [permission] or --allow/--deny; deny wins.
Agent Client Protocol
SpaceXAIUse Grok Build through the Agent Client Protocol (ACP) in other apps.
Sessions
SpaceXAIEvery conversation is saved under ~/.grok/sessions/. Resume with /resume, grok --resume, or grok -c. /fork branches a peer session; /rewind restores files and truncates the conversation. grok sessions list/search/delete and grok export manage them.
Agent Dashboard
SpaceXAIFullscreen overview of every session (needs input, working, idle, completed, failed). Open with Ctrl+\, /dashboard, or grok dashboard. Reply or answer permission prompts inline; Ctrl+W starts new agents in a worktree.
/goal and /loop
CursorAssign a long-lived objective (e.g. fix flaky tests and make CI green). Pair with /loop for recurring check-ins.
Custom modes
CursorPin skills as always-on modes so agents stay focused. Select from the / menu.
Steering
CursorSend a follow-up without interrupting the current turn. Press return twice or Send now.
Builds
CursorReady-to-use copies of the dev environment. Agents boot ~3× faster to first token; included with Cloud Agents.
Conversation hooks
CursorObserve and control prompts, responses, thinking, subagents, compaction, and turn completion.
Start from scratch
CursorCloud Agents can start without a connected GitHub or other SCM. Cursor creates an Origin repo in the background; save with Create repo when ready.
Live preview
CursorPort-forwards the cloud agent environment to the browser for preview and design mode.
Vercel publish
CursorConnect a Vercel account and publish the agent build to a live URL.
Self-hosted machines
CursorRun cloud agents on machines in your network. My Machines connects a laptop or VM; team pools scale workers on demand with hibernate and restore. Integrates with AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, and E2B. Self-hosted workers support computer use on Linux and Mac.
Projects
CursorLong-running work with a coordinator agent that plans and delegates to implementing agents. Shared context files sync across cloud and local machines; runs on a cloud computer when the laptop is closed.
Rollouts
CursorChange monitor on every pull request: writes a monitoring plan, tracks deploys per environment, and reports verified healthy, regression detected, or inconclusive. Connects Origin or GitHub, CD deploy events, and telemetry (Datadog and others).
Security Review
CursorPosts one review comment per pull request reporting exploitable bugs (injection, auth bypass, secrets in source, SSRF, unsafe deserialization, vulnerable dependencies). Team rules enforce codebase-specific policies. Draft PRs are skipped.
Agent SDK
CursorAgents as code: instructions.md, tools/, skills/, MCP connections, and channels. `agent-sdk dev` runs locally; `agent-sdk deploy` hosts on Cursor. Evals and `agent-sdk github replay` test PR-driven bots.
Live apps
CursorCloud agents scaffold web apps, provision Postgres and Cursor SSO auth, run security review, and publish to a chosen domain for team or public sharing.
Pull requests
CursorTimeline, commits, checks, diffs, comments, and merges in Origin, two-way synced with GitHub.
Agents in every repo
CursorQuery code, make changes, update PRs, and push branches from hosted repos.
27 of 89