OpenClaw vs Hermes Agent: The Great AI Gateway Split of 2026

For most of the past three years, choosing an open-source AI agent gateway was simple: you picked OpenClaw. With 377,000 GitHub stars, 1,600+ contributors, and a plugin ecosystem that connected to virtually everything, it was the default. The question was never whether to use OpenClaw — it was which version to pin.

Then March 2026 happened. Nine CVEs dropped in four days, two scoring CVSS 9.9. A supply-chain attack called ClawHavoc seeded 1,184 malicious skills across ClawHub, with roughly 12% of marketplace listings turning out to carry malware. Over 135,000 exposed OpenClaw instances were catalogued globally. The total CVE count for the project now exceeds 60.

In the fallout, a younger project called Hermes Agent — built by Nous Research — went from "interesting newcomer" to the fastest-growing open-source agent framework of 2026, pulling in 181,000 stars in just four months. The community did not just notice Hermes. It split in two.

What Is OpenClaw?

OpenClaw is a local-first AI agent gateway that routes sessions between users and language models. It is session-native, MIT-licensed, and — since February 2026 — governed by the OpenClaw Foundation after creator Peter Steinberger transferred the project before joining OpenAI.

The numbers tell a story of maturity: 78,800 forks, 137 releases, and integrations with over 20 external services out of the box. OpenClaw's plugin architecture is battle-tested, its community is massive, and its v4.0 roadmap — expected mid-2026 — promises multi-agent orchestration, Plugin SDK v2, vector-backed memory, and a wave of enterprise integrations.

If you need breadth today — connecting an agent to Slack, Jira, Salesforce, your internal APIs, and a dozen other services in one config file — OpenClaw remains unmatched.

But maturity cuts both ways. Long-time users have complained about chronic update instability (breaking changes between minor versions are a recurring theme) and persistent memory failures that lose agent context across sessions. These were annoyances before March 2026. After the security crisis, they became reasons to look elsewhere.

What Is Hermes Agent?

Hermes Agent is Nous Research's MIT-licensed agent framework, built around a philosophy that sounds simple but turns out to be genuinely different: agents should get better at their jobs over time.

The core innovation is a closed learning loop. Hermes agents accumulate "skills" — learned task patterns — through a self-improvement pipeline. After completing roughly 20 tasks, Nous Research reports that agents show a 40% improvement in task completion speed. (That benchmark is self-reported and not yet independently verified — take it with appropriate caution, but the architectural mechanism is real and observable.)

The latest release, v0.12 "Curator," leans further into this with self-improving skill curation. Hermes also features persistent dual memory (working memory plus long-term knowledge), deny-by-default permissions, checkpoint and rollback for agent state, and a 1.2-second cold start time — roughly half to a third of OpenClaw's typical 2–4 seconds.

What Hermes does not have is OpenClaw's sprawling ecosystem. The integration library sits at 6 to 16 connectors depending on how you count, versus OpenClaw's 20+. There is no marketplace equivalent to ClawHub (which, in light of ClawHavoc, might actually be a feature). And with only 11 releases to date — three of which users report were broken on launch — the release cadence is rapid but not yet reliable.

The March Inflection Point

The security crisis did not happen in a vacuum. It landed on a community already frustrated by OpenClaw's operational rough edges. When kilo.ai analyzed over 1,300 Reddit comments in the aftermath, the split was clear: roughly 35% of respondents were staying with OpenClaw, about 30% had switched to Hermes, and 20% were running both in parallel.

The switching narrative crystallized around a few key points. Hermes has zero CVEs on its record. Its deny-by-default permission model means agent capabilities must be explicitly granted rather than implicitly available. There is no centralized skill marketplace to compromise. And the checkpoint/rollback system means a bad agent run can be unwound rather than triaged after the fact.

One widely-shared Medium writeup by developer @sathishkraju walked through a full migration from OpenClaw to Hermes, documenting both the security improvements and the integration gaps. It became a reference point for the "switcher" community.

But the story is not one-sided. About 15% of Reddit commenters flagged concerns about possible astroturfing in Hermes advocacy — though no evidence of coordinated promotion has been found. And the practical reality is that Hermes's smaller ecosystem means real work for teams that depend on integrations OpenClaw provides out of the box.

Head to Head: Where Each Wins

Security. Hermes leads decisively. Zero CVEs versus 60+, deny-by-default versus opt-in restrictions, no marketplace attack surface. If your threat model includes supply-chain risk from third-party plugins, this is not close.

Ecosystem and integrations. OpenClaw wins on breadth. Twenty-plus connectors, a massive contributor base, and years of community-built tooling give it coverage Hermes cannot yet match. Enterprise teams with complex integration requirements will feel this gap.

Learning and adaptation. Hermes's self-improving skill system is architecturally unique in this space. However, it ships with self-learning off by default — a major gotcha that many new users miss. And there is a known flaw: the self-evaluation pipeline can encode errors into skills, meaning an agent that learns a bad pattern will confidently repeat it.

Performance. Hermes is faster on cold start (1.2s vs 2–4s) and claims faster task completion at scale. The v0.11 to v0.12 jump alone showed a 57% cold-start improvement.

Stability. Neither project covers itself in glory here. OpenClaw's breaking changes between minor releases are legendary. Hermes has shipped three reportedly broken releases out of eleven total. Both projects need better release discipline.

Governance. OpenClaw's move to a foundation with a public RFC process gives it structural transparency. Hermes is maintained by Nous Research, a single company — well-regarded, but a different governance profile. Both are MIT-licensed.

The Verdict

This is not a story where one tool replaces the other. It is a use-case split.

Choose OpenClaw if you need maximum integration breadth today, if your team has invested in the plugin ecosystem, or if you are betting on the v4.0 roadmap delivering enterprise-grade orchestration by late 2026. The Foundation governance model also matters if organizational policy requires transparent open-source stewardship.

Choose Hermes Agent if security is your primary constraint, if you want agents that compound their effectiveness over time, or if you are building greenfield and do not need OpenClaw's legacy connector library. Just remember to turn on self-learning — and monitor what your agents are teaching themselves.

Run both if you can absorb the operational overhead. A meaningful chunk of the community is doing exactly this: OpenClaw for its integration reach, Hermes for security-critical workloads.

The March 2026 crisis did not kill OpenClaw — 377,000 stars and a foundation-backed roadmap are not going away. But it broke the assumption that OpenClaw was the only serious option. For the first time, the open-source agent space has a genuine two-horse race. Where you place your bet depends on whether you are optimizing for breadth or for depth — and how much the security question keeps you up at night.