Hermes Agent's Velocity Release: v0.15.0, NVIDIA Partnership, and 140K GitHub Stars

TLDR: Hermes Agent shipped three releases this week — v0.15.0 “The Velocity Release,” v0.15.1, and v0.15.2. The core agent loop was slashed 76% (16K → 3.8K lines), session_search is now 4,500× faster at zero cost, Kanban gains a full Swarm orchestrator, and NVIDIA officially featured Hermes for self-improving local agents on RTX PCs and DGX Spark. The project crossed 140,000 GitHub stars in under 3 months and is now the most-used agent on OpenRouter.

Three Releases in Three Days

The last week of May saw an unprecedented burst of shipping. Here’s what landed:

v0.15.0 (May 28) — The Velocity Release

1,302 commits, 747 merged PRs, 560+ issues closed, and 321 community contributors. The theme is speed — startup, search, and agent loop.

Big Refactor. run_agent.py was the single file that drove the entire conversation loop — it had grown to 16,083 lines over months of feature additions. It’s now 3,821 lines across 14 cohesive modules under agent/. That’s a 76% reduction. The change is invisible to users — every existing feature, skill, and config should work identically — but it unlocks faster iteration and clearer contribution pathways.

Cold-start Performance. Another second shaved off startup time alongside a 47% reduction in per-turn function calls (399K → 213K over 31 turns). On Termux (Android), cold start dropped from 2.9s to 0.8s. Even hermes --version is 63% faster (701ms → 258ms), beating Codex CLI head-to-head.

session_search Rebuilt. The old session_search loaded transcripts into context and asked the LLM to summarize — ~90 seconds per query with real token cost. The new version is pure SQLite FTS5: ~20ms for discovery, ~1ms for scroll navigation. Three clean calling shapes — Discovery, Scroll, Browse — each optimized for a different retrieval pattern. Completely free to use.

Kanban Multi-Agent Platform (104 PRs). The Kanban system graduates from a simple board to a full orchestrator:

• hermes kanban swarm creates Swarm v1 dependency graphs for multi-agent workflows
• Triage channels auto-decompose complex tasks into Kanban items
• Per-task model overrides — route vision work to vision-capable models and cheap summarization to economy models
• Scheduled tasks with cron integration
• Worker introspection endpoints (/workers/active, /runs/{id}, /inspect)
• Worktree-per-task isolation prevents git conflicts between parallel agents

Session Orchestrator in TUI. Multiple live sessions are now manageable from a single TUI window with an overlay session switcher. Switch between agent instances working on different tasks without leaving the terminal.

Bitwarden Secrets Manager. One bootstrap token replaces all per-provider API keys. Configure once, rotate centrally, and every Hermes provider authenticates through Bitwarden.

ntfy — 23rd Platform. Self-hostable push notifications with zero account requirements. Send alerts from cron jobs, monitors, or any Hermes task to any device via the ntfy protocol.

Skill Bundles. /<name> now loads multiple skills at once. New additions include code-wiki, openhands, and web-pentest skill bundles.

Promptware Defense. Brainworm-class prompt injection attacks are blocked at three chokepoints: threat pattern matching in tools, recalled memory scanning, and delimiter marker detection. A new security-guidance plugin provides audit trails.

v0.15.1 (May 29) — The Patch Release

28 commits, 21 merged PRs. The headline fix: the Dashboard infinite-reload loop that affected Docker and fresh installs (the identity probe returned 401 by design, causing a stale-token reload loop). Fixed by adding allowUnauthorized to fetchJSON.

Other notable fixes:

• Docker --insecure is now an explicit env opt-in (HERMES_DASHBOARD_INSECURE=1)
• MCP bare command resolution (npx, npm, node) in Docker now resolves against /usr/local/bin
• Skills page sidebar and source pills restored (stale useMemo dep)
• Kanban workers properly handle SIGTERM and receive images in task bodies
• .md files delivered correctly again (media delivery switched to denylist-only)
• /yolo works mid-session for per-session bypass
• Web URLs pass through redaction unchanged

v0.15.2 (May 29) — Packaging Fix

A quick follow-up: plugin.yaml manifests are now bundled in wheel and sdist releases. The fix ensures plugin-based installations work out of the box.

NVIDIA Features Hermes for Local AI Agents

On May 13, NVIDIA published a blog post showcasing Hermes Agent on RTX PCs, RTX PRO workstations, and DGX Spark — marking a major milestone for local-first AI agents.

The post highlights four standout capabilities:

1. Self-Evolving Skills — Hermes writes and refines its own skills, saving learnings from complex tasks and user feedback into persistent skill documents.
2. Contained Sub-Agents — Short-lived, isolated workers for sub-tasks keep context small and manageable, ideal for local models with limited memory.
3. Reliability by Design — Every skill, tool, and plugin is curated and stress-tested. “Hermes just works — even with 30 billion-parameter-class local models — without constant debugging.”
4. Same Model, Better Results — Hermes is an active orchestration layer, not a thin wrapper. The agentic fabric enables persistent on-device agents rather than task-by-task execution.

The post also highlights Qwen 3.6 models running locally on NVIDIA hardware — the 35B variant fits in ~20GB memory while surpassing 120B-parameter models, and the 27B dense model matches the accuracy of Qwen 3.5 397B at 1/16 the size.

DGX Spark, with its 128GB unified memory and 1 petaflop AI performance, runs Hermes + local models all day. The recommended playbook is available at build.nvidia.com/spark.

140K Stars and Counting

Hermes Agent is now the fastest-growing agent framework on GitHub — 140,000+ stars in under 3 months. It’s the most-used agent on OpenRouter, and the community growth is accelerating: v0.15.0 alone saw 321 community contributors.

How to Update

pip install --upgrade hermes-agent

Verify:

hermes --version
# Should show: v0.15.2 (v2026.5.29.2)

If you’re running from a git clone, switch to the PyPI package:

pip install hermes-agent
hermes doctor

What’s Next

With the core refactor landed, the team is focused on deeper Windows support, expanded OAuth providers, and pushing the Kanban Swarm platform toward stable. The session_search rebuild and run_agent.py decomposition open the door for plugin-based agent loop customization — expect that pattern to mature in v0.16.0.

Track development on GitHub and join the community.

---

Hermes Tutorials is an independent community resource. Not affiliated with Nous Research.