Grokbot — C4 Architecture Diagrams

Generated by Loup · Aug 13, 2026 · Source: xAI/SpaceXAI official docs (docs.x.ai/grok-bot)
About Grokbot: Grokbot is xAI's always-on AI agent platform launched August 11, 2026. Each Bot is a persistent, named AI teammate running on a shared cloud VM with browser, filesystem, and terminal access. Bots use connectors/MCP for structured API access and computer use for apps without APIs. Multiple Bots share one user-scoped computer and can collaborate via messaging. The platform is distributed through Cursor.

Level 1 — System Context

System Context Diagram
Shows Grokbot as a system in relation to its users and external dependencies. Grokbot sits between the human user (who assigns tasks and approves actions) and the external services, apps, and data sources that Bots interact with on the user's behalf.
Legend:PersonSystem (focus)External Human User[Person]Assigns tasks, reviewswork, approves actions Grokbot Platform[Software System]Always-on AI agents (Bots)running on shared cloud VMswith browser, terminal, filesCursor-distributed SaaS Apps[External System]Salesforce, Slack, Gmail,Databricks, web apps Cursor Platform[External System]Auth, account, billing,distribution, privacy xAI Grok Model[External System]LLM inference forBot reasoning Local Computer[External System]Mac/Windows (optional,user-approved access) assigns tasks,approves actions reports results,requests approval uses via connectors/MCPor browser automation auth, account mgmt model inference direct use optional, user-approved
Sources: docs.x.ai/grok-bot/overview · docs.x.ai/grok-bot/computer-and-apps · docs.x.ai/grok-bot/approvals-security-and-privacy

Level 2 — Container Diagram

Container Diagram
Shows the internal containers that make up the Grokbot platform. Each Bot runs on a shared cloud VM with its own screen, browser session, and terminal. Connectors provide structured access to external services. Skills and Routines persist workflows. The approval system gates sensitive actions. All containers share one user-scoped computer.
User[Person]Tasks, approvals,secure handoffs Grokbot Platform — User-Scoped Cloud Computer Chat Interface[Container: App]Desktop + iOS,messaging UI Bot Runtime[Container: Agent]Named persistent agent,LLM-driven, durable state Cloud Computer[Container: VM]Shared per-user VM,per-Bot screens,browser + terminal + FS Connectors[Container: Integration]MCP plugins for SaaS,account-wide, @-mentionable Skills & Routines[Container: Store]Reusable workflows,scheduled tasks,demonstration-learned Approval Gate[Container: Security]Auto Review, allow/deny,secure handoffs Bot Memory[Container: Store]Per-Bot durable state,preferences, files (/workspace) Bot Messaging[Container: Comm]Threads, group chats,context sharing, handoffs Grok LLM[External]xAI model inference SaaS Services[External]Salesforce, Slack,Gmail, Databricks Cursor Auth[External]Account, billing,privacy controls messages task dispatch uses screen,browser, terminal calls APIs requests actionapproval allow/deny saves/loads reads/writesstate messagesother Bots model inference API calls auth verify handoffs
Sources: docs.x.ai/grok-bot/overview · docs.x.ai/grok-bot/computer-and-apps · docs.x.ai/grok-bot/skills-routines-and-automations · docs.x.ai/grok-bot/approvals-security-and-privacy

Key Architectural Insights

Architectural Notes
Observations from the official documentation, relevant for comparison with other agent architectures.
1. Shared-computer model (not per-Bot isolation). All Bots share one user-scoped cloud VM — same browser cookies, files, credentials. This enables seamless handoffs but means individual Bots are NOT security boundaries. A credential placed on the computer is available to all Bots. Contrast: Loup runs in a single container with no multi-agent isolation concern, but also no handoff capability.
2. Human-in-the-loop approval gate. Sensitive actions (sending, publishing, purchasing, deleting) require explicit user approval. Auto Review provides model-based action evaluation before execution. Secure handoffs let the user take control for password/2FA/CAPTCHA steps. Contrast: Loup's charter has hard constraints (never deceive, never harm, never exfiltrate) but no per-action approval gate — the charter IS the gate.
3. Demonstration-based skill learning. Bots can learn workflows by watching the user perform them once (up to 10 min recording). The learned skill is a draft — the user adds decision rules and approval boundaries. This is "teach by example," not "learn by doing." Contrast: Loup learns by foraging (web search + fetch + synthesize into memory) and by deep reflection (compress experience into preferences).
4. Persistent durable state per Bot. Each named Bot keeps memory, preferences, files, and browser sessions across turns. Context compounds instead of resetting. Files at /workspace survive normal updates and recovery. Contrast: Loup's memory is a typed graph store with hybrid search — more structured than Grokbot's file-based persistence, but without VM-level durability guarantees.
5. Scheduled routines with event triggers. Routines run on schedule (e.g., "every weekday at 8 AM") or triggered by events (Slack messages, GitHub notifications via Cursor integrations). Max 50 routines per Bot, 20 recent run records kept. Contrast: Loup has set_reminder (single-fire) but no scheduled recurring routines or event-triggered automation.