Hermes Agent reports No. 1 OpenRouter rank after v0.13.0

• NousResearch's release post and Teknium's feature summary introduced Hermes Agent v0.13.0 as the "Tenacity Release," centered on finishing long-running tasks with a multi-agent Kanban board, a persistent /goal loop, and broader extension points.
• The release notes linked by NousResearch's full notes post, plus WesRoth's rundown, say v0.13.0 also shipped Checkpoints v2, gateway auto-resume, default-on data redaction, and fixes for eight P0 security issues.
• Teknium's credential-pools tip surfaced one of the more operationally useful additions, rotating across multiple API keys for the same provider, while Teknium's Pareto Code note showed Hermes wiring in OpenRouter routing features inside the same stack.
• Adoption moved fast: NousResearch's ranking screenshot claimed Hermes Agent reached No. 1 on OpenRouter's global app rankings ahead of OpenClaw, and OpenRouter's congrats post pointed readers to the live OpenRouter apps leaderboard.
• Early user commentary from Teknium's migration retweet and Teknium's session-adaptation retweet framed Hermes as lower-maintenance and better at carrying context across sessions than OpenClaw, though those comparisons came from users, not a controlled benchmark.

You can skim the official v2026.5.7 release notes, open the credential pools docs, check the OpenRouter app rankings, and, if you want the next rollout clue, Windows native beta docs were already live before the OpenRouter ranking spike.

The official framing was unusually specific: make the agent finish what it starts. In Teknium's launch post, that broke down into four visible product bets:

1. Multi-agent orchestration through a Kanban-style board.
2. Enforced goal completion through /goal.
3. Disk-usage optimizations.
4. Extensibility for custom LLM providers and gateway channels.

WesRoth's walkthrough added the lower-level pieces the launch post skipped: heartbeats, zombie detection, per-task retry budgets, Checkpoints v2, auto-resume for interrupted sessions, default-on redaction, and eight closed P0 vulnerabilities. The release notes live in the linked GitHub tag.

The new board is the most concrete subsystem in the release. According to WesRoth's walkthrough, users can drop tasks onto a shared board where multiple Hermes workers pick them up, hand them off, and close them out.

That setup matters less as a UI tweak than as a durability layer. The support mechanics listed in that thread are the whole point:

• Heartbeats for active workers.
• Zombie detection for stalled jobs.
• Per-task retry budgets.
• Checkpoints v2 for state persistence.
• Gateway auto-resume for interrupted sessions.

The /goal command, which WesRoth's post also called the Ralph loop, pins the agent to a target across multiple turns so it does not drift off the original request. That sits neatly next to kylejeong's Autobrowse demo, where Hermes compressed a Hacker News browsing task from 102 seconds to 35, from 23 turns to 8, and from $1.46 to $0.28 by learning to execute JavaScript directly on the page and save that path as a skill.

The quiet operator feature in this cycle is credential pooling. Teknium's docs link post said Hermes can store multiple API keys for the same provider and rotate across them for stability, with the implementation documented in the credential pools guide.

That landed alongside a broader provider story. WesRoth's release summary said third-party model providers can now be added as plugins, and Teknium's routing post showed Hermes exposing OpenRouter's Pareto Code path for auxiliary tasks through its configuration docs.

Hermes also kept making its skill layer more configurable around the edges. Two days before the release, Teknium's blank-slate profile post added hermes profile create <name> --no-skills, which lets users start an agent without bundled skills instead of pruning a default profile after the fact.

The easiest adoption signal to verify is the leaderboard itself. NousResearch's screenshot showed Hermes Agent at 271B, ahead of OpenClaw at 245B, and OpenRouter linked to the public apps ranking page.

That ranking spike immediately turned into comparison chatter. testingcatalog's post called out Hermes overtaking OpenClaw, while kimmonismus's reaction framed it as a new community favorite.

The more useful signal came from migration reports. In Teknium's retweet of a migration report, one user said moving a workflow from OpenClaw to Hermes meant "zero maintenance," and in Teknium's retweet of another user, a second user said Hermes adapted better across sessions. kilocode's comparison post gave the cleanest explanation for why those reports diverge: Hermes packages a gateway around a learning agent, while OpenClaw packages an agent around a messaging gateway.

Windows support was already in motion before the ranking posts hit. Teknium's beta announcement said native Windows builds were available for early testing, with setup steps in the Windows native guide.

The screenshot in that beta post exposed the current shape of the stack better than the announcement text did. It showed Hermes Agent v0.13.0 on PowerShell with 18 toolsets, 84 skills, and categories spanning GitHub, MLOps, note-taking, smart home control, red teaming, and software development.

That same screenshot also surfaced a small but revealing default: agent.api_max_retries is set to 3, with the inline tip explaining it controls failed API-call retries before Hermes gives up. It fits the broader v0.13.0 theme, reliability features are no longer tucked away as backend plumbing, they are becoming first-class operator knobs.