| Origin & maker | Nous Research, an open-source AI lab | OpenClaw Foundation, an open-source community project | xAI |
| Core architecture | A personal assistant built around autonomous learning and self-improving skill creation | A self-hosted gateway operating as infrastructure for multi-agent workflows with explicit control planes | A named teammate ('Bot') is a role plus memory plus thread; the machine underneath is one persistent, shared computer per user account |
| Hosting model | Open source, self-hosted; runs on Mac, Linux, or Windows WSL2 | Self-hosted; one gateway is documented as one trust boundary, native plugins run unsandboxed | Fully managed; a persistent cloud computer provisioned per user account, never exposed to the user directly |
| Memory system | Identity in SOUL.md, facts in MEMORY.md/USER.md with hard character limits, sessions searchable in SQLite; also integrates the Honcho memory backend per the Substack comparison | Memory stored as Markdown files in the workspace — user-inspectable and version-controllable, requiring explicit writing to disk | Maintained per Bot; retains stable working preferences, important facts, and summaries rather than replaying every message; the underlying mechanism is not published |
| Skill / workflow authoring | Skills are Markdown files with structured frontmatter that the agent writes and revises itself; a Curator prunes/archives unused skills and the GEPA pipeline evolves skills offline against execution traces | Not detailed in the Substack comparison beyond markdown-file memory; positioned as infrastructure rather than a skill-authoring product | A skill is a reusable instruction set referenced with '/' in the composer, or captured directly via 'Teach a task' (records a browser demo up to 10 minutes, no audio) |
| Multi-agent coordination | Not designed as a multi-agent product; focused on one personal assistant per user | Supports multiple agents within a single gateway, with explicit approval gates and security boundaries | Group chats (2-6 Bots), asynchronous cross-Bot messaging, and ownership handoffs work without configuration, sharing one browser session and credential set across the whole roster |
| Setup & onboarding | Faster initial setup and deployment than OpenClaw, per the Substack comparison | Steeper learning curve and more setup friction; documentation says to treat it like infrastructure, and recent releases have caused configuration breakage ('upgrade fatigue') | A macOS app install completing in a few minutes: download, drag to Applications, authenticate, then onboarding introduces Bots, the shared computer, and routines |
| Approval & safety boundaries | Requires sandboxing and permission scoping, per the Substack comparison's security note | Explicit approval gates and security boundaries; docs state one gateway per trust boundary and warn native plugins run unsandboxed | A documented FINISH WITHOUT ASKING / STOP AND ASK ME boundary set on each Bot, plus Auto Review (Require Approval rules always win over Always Allow) and a takeover flow that hands the human the machine for passwords, 2FA, and CAPTCHAs |
| Ideal use case | Individual users handling personal tasks, reminders, and light research; native support for 17+ messaging integrations | Business workflows, multi-channel operations, approval-gated processes, and compliance scenarios | Delegated specialist roles built one at a time — Chief of Staff, Research Analyst, Sales Outbound, Expense Manager — coordinating through routines and group chats |
| Cost / token economics | Self-hosted token costs the operator pays directly to whichever model runs it — tracked daily costs across autonomous agents ranged from $0 (local models) to $8.70 (Claude Opus 4.7), roughly $261/month | Same self-hosted, pay-your-own-model-provider economics as Hermes; the article frames autonomous reflection as what compounds the expense | Subscription-gated managed access; exact pricing is not stated in either Grok Bot X post analyzed here |