242,000 Stars and Counting: Hermes Agent Now Out-Stars Claude Code and OpenAI Codex Combined Momentum on GitHub

The most-starred project in the AI agent space is no longer owned by a frontier lab. As of September 6, 2026, NousResearch/hermes-agent sits at 242,313 GitHub stars — comfortably ahead of Anthropic’s Claude Code (144,235) and OpenAI’s Codex (121,904), and nearly triple the count of OpenHands (86,338), the previous open-source standard-bearer for autonomous agents. The September 6 edition of LLM Daily flagged it as “a leading open-source AI agent framework,” noting 575 new stars arriving in a single day — and the repository was pushed to again less than an hour before this article was written.

The numbers around the project have become difficult to ignore. The repo, created on July 22, 2025, has accumulated 49,811 forks and 40,238 open issues — the latter a testament to how heavily it is used, not how broken it is. It is written in Python, MIT-licensed, and describes itself simply as “the agent that grows with you.” In roughly fourteen months, it has gone from a Nous Research experiment to the reference point for what a full-featured, self-hostable AI agent environment looks like.

The Pantheon Release: agents as a society

The current surge traces to v0.21.0, tagged v2026.8.31 and published on August 31, 2026. The scale of the release itself is a headline: since v0.20.0, the project merged roughly 2,475 pull requests across ~5,800 commits, changed ~5,680 files, closed ~2,100 issues, and drew contributions from more than 760 contributors. The release notes call it “The Pantheon Release” — and the theme is unmistakable: multi-agent orchestration moved from plumbing to product.

The centerpiece is Bot Mode, now bundled and default-on in the desktop app. Every agent profile gets a name and a deterministic avatar face; bots can be organized into Discord-style group chats alongside the human user, @-mentioned from the composer, and given rooms with their own names and pictures. As the release notes put it: “Before, ‘multi-agent’ meant plumbing; now it looks like a chat app full of coworkers.”

A companion feature, hermes peer, gives any Hermes agent a handle-based direct message channel to any other — across profiles and gateways. A research bot can hand findings to a coding bot, and the reply lands in each agent’s durable, inspectable Bot Chat rather than vanishing into a fire-and-forget pipeline.

Scheduled agents that remember

Perhaps the most quietly consequential change is what happened to cron jobs. Scheduled agents historically suffered from amnesia: every run started from zero. In v0.21.0, cron agents load and update persistent memory like any interactive session; a continuity=true flag carries each run’s output into the next, so a monitoring agent can deduplicate against what it already reported. Every job gets a durable notepad scratchpad, monitor-mode jobs skip the LLM call entirely when nothing has changed, and scheduled output can land in a bot’s chat — where the bot actually responds. In the release’s own words: “Your 9am briefing job now knows what it told you yesterday.”

Subagents got the same graduation. delegate_task now supports live orchestration — list running children, steer one mid-flight with a course correction, or stop it early and keep the partial result — with optional JSON-schema validation on child outputs, per-delegation cost reporting, and raised defaults of 250 iterations and 10 concurrent children.

MCP as a command center, not config archaeology

The Model Context Protocol surface was rebuilt into a single desktop page: drag-in “paste anything” import, background health checks that nudge re-authentication before a tool call fails, a fleet-wide cost and usage overlay with schema token estimates and 30-day usage per server, and hermes:// deep links that install an MCP server with explicit confirmation. Managing twenty MCP servers, the notes quip, “used to be config-file archaeology; now it’s a dashboard.”

The CLI kept pace: a Ctrl+P fuzzy command palette, a /status view showing reasoning mode, pending approvals, and context usage, and a status bar with live cache-hit percentage, latency, and tokens per second. The agent can now also drive the desktop app’s built-in browser — navigating, clicking, and reading pages directly — and six new providers landed, including Meta’s Model API (Muse Spark), Tencent TokenPlan, and Nebius Token Factory, alongside catalog additions like GLM-5.3-Flash, qwen3.8-max/flash, and MiniMax M3.

Security hardening is notable precisely because agent frameworks are now attack surfaces: protected agent-instruction files (AGENTS.md, skills, memory stores) always require write approval, so a prompt-injected agent cannot quietly rewrite its own standing orders, and a redaction sweep closed secret-leak paths across terminal errors, .env reads, checkpoints, and ACP logs.

Why the star count matters — and what it doesn’t

GitHub stars measure attention, not revenue, and Anthropic and OpenAI would happily trade stars for the enterprise contracts their CLI tools convert. Claude Code and Codex ship as thin clients to frontier models; Hermes Agent is the opposite bet — a self-hosted, MIT-licensed environment where the user owns the harness, the memory, the provider choice, and the audit trail. Its rise signals where developer sentiment sits: on extensibility, privacy, and escape from per-seat lock-in.

The project’s commit distribution also tells a health story. While teknium1 — the OpenHermes model author and co-founder of Nous Research — leads at ~13,972 commits, the next tiers — OutThisLife (~3,379), kshitijk4poor (~2,101), and hundreds of others — show a genuine community, not a single-vendor repo. For an ecosystem that many assumed would consolidate around two or three proprietary CLIs, 242,000 stars is a loud counter-argument: the agent layer is still anyone’s game, and open source is playing to win.