The self-hosted AI agent category has six products people actually consider in 2026: Smartopol, Hermes, OpenClaw, NanoClaw, IronClaw, OpenJarvis. Here's what each one claims, what each one actually ships, and where Smartopol lands.
no weasel words · no "other agents" · competitor marketing quoted, checked against the code
| capability | smartopol | hermes | openclaw | nanoclaw | ironclaw | openjarvis |
|---|---|---|---|---|---|---|
| Self-hosted | yes | yes | yes | yes | yes | yes |
| Self-evolving (nightly reflection) | cortex dreaming | simulated | ||||
| Native hierarchical memory | layered memory pyramid | keyword search only | basic | basic | encrypted KV | basic |
| Cross-channel single identity | 6-digit pairing | partial | per-channel | |||
| Multi-channel | Discord · TG · Web · WS · Terminal | 14 claimed | 20+ claimed | WA · TG | limited | local |
| Model-agnostic | 15+ providers + OAI-compat | Claude-first | Claude-first | Claude-first | Claude-first | Claude-first |
| Sandboxed plugin system | signed, isolated | script plugins | script skills | basic | basic | basic |
| Prompt-injection hardened memory | regex + fenced + evergreen | partial | CVEs in 2026 | yes | ||
| OS-level sandboxing | kernel-enforced | claimed | weak defaults | container | zero-trust | |
| Audit log (tamper-evident) | cryptographically chained | claimed | ||||
| Self-update (signed, atomic) | yes | pip install -U | manual | manual | manual | |
| Multi-user RBAC | 9 permissions | per-agent | yes | |||
| Real benchmarks published | 18M tokens · 65–87ms | |||||
| License | free-like-VLC | MIT | MIT | MIT | commercial | MIT |
| Pricing | free forever | free | free | free | paid only | free |
| Commercial support available | yes | yes |
The closest positioning rival. Their homepage: "an autonomous agent that lives on your server, remembers what it learns, and gets more capable the longer it runs." That is also exactly what Smartopol is.
pip install -U. Smartopol: signed-release, atomic-binary-swap, mid-conversation-resume.smartopol bench. Hermes: none.Hermes and Smartopol are philosophically aligned and both free to use.
Hermes chose channel breadth + MIT + community governance.
Smartopol chose memory depth + hardening + closed source for the long run.
Pick by which half matters to you.
Pitch: "the AI that actually does things. Clears your inbox, sends emails, manages your calendar, checks you in for flights. All from WhatsApp, Telegram, or any chat app you already use." Strongest axis: channel + skill marketplace breadth. Weakest axis: security posture.
npx openclaw · ~5 minutes claimedOpenClaw won the race to 5,700 skills. Smartopol won the race to skills that aren't malicious. If your threat model allows an unaudited marketplace, OpenClaw's ecosystem feels richer. If it doesn't — a smaller, curated, OS-sandboxed surface is the entire point.
Pitch: "lightweight, production-ready, security-focused alternative to OpenClaw. Docker MicroVM sandbox. Single command." Minimalism play — ~3,900 LoC — "small enough to fully understand and audit."
NanoClaw if value function is "smallest possible auditable surface."
Smartopol if value function is "the most capable private agent with responsible defaults." Both are valid. They are different products.
The enterprise zero-trust agent. Pitch: "Zero-trust sandboxing with encrypted memory. SOC 2 compliance tooling. Healthcare, finance, legal." Pricing around $20/month + LLM costs.
curl | sh.IronClaw wins the enterprise RFP checklist as a default vendor.
Smartopol wins the engineer deciding on a Tuesday afternoon — the one who wants something real on their machine by the end of the day.
A local-first personal agent framework. Earlier-stage than the others — primarily YouTube demos and secondary coverage. Positioning: "runs on your own device, not in the cloud."
Where OpenJarvis leads: pure local-first framing · strongest privacy positioning in the group · earliest-to-market for "JARVIS"-style assistant concept.
Where Smartopol leads: every other axis. Smartopol is also local-first, and adds channels, memory pyramid, dreaming, plugin sandbox, multi-user RBAC, commercial support, real benchmarks.
Honest read: OpenJarvis is more idea than product today. Revisit in six months.
Complementary, not substitutes. The typical Smartopol user runs Claude Code inside a Smartopol terminal session — that is a supported pattern (the claude_session tool exists for this). Smartopol becomes the long-lived brain. Claude Code is one of its hands.
Aider and Goose are coding-first agents with narrow scope — edit a repo, run tests, commit. Smartopol is a general-purpose agent that can also do coding — but can also remember your mother's birthday, schedule reminders, post to Discord, run Home Assistant automations, and wake up at 04:00 to reflect on the day.
If you only want a coding agent, Aider or Goose are excellent. If you want one agent for everything — Smartopol.
smartopol bench after install.hermes-agent.nousresearch.com/docs.If anything here is wrong, email hello@smartopol.ai with a link to the counter-evidence — we'll update the page same day.
Smartopol is currently in private beta. Invitations go out in waves, before public launch.
No newsletter. No nurture sequence.