Hermes Agent Crosses 90K GitHub Stars, v0.15.0 Ships All 69 Plugin Manifests

Hermes Agent crossed 90,000 GitHub stars today, roughly two months after launch. The milestone was surfaced by @DamiDefi, whose post racked up 263 likes and 59 retweets -- the highest engagement on the topic in the last 24 hours. It quoted @akshay_pachaar's Hermes Agent Masterclass, which sits at 5,988 likes and over 4.8 million impressions since May 13.

v0.15.0 milestone: 69 plugin manifests in the PyPI wheel

Smelter Labs AI runs a daily digest tracking the most substantial commits to the Hermes Agent repo. Today's entry -- the 7th in the series -- flags a packaging fix: the v0.15.0 PyPI wheel now ships all 69 plugin.yaml manifests that were previously missing. The wheel contained Python code but zero plugin descriptors, meaning downstream consumers had to source manifests from the GitHub repo directly.

This is the kind of packaging hygiene that separates "works on my machine" from reliable distribution. Plugin manifests are the metadata layer that Hermes uses to discover and register tools at startup -- without them, installed plugins are invisible to the agent.

Step 3.7 Flash + Hermes Agent

Nous Research posted about StepFun's new model release, calling it "a punch at a small size" and expressing interest in seeing how it performs with Hermes Agent. The post drew 152 likes and 14 replies. StepFun responded confirming that Step 3.7 Flash was built specifically for agentic workloads:

Small models purpose-built for tool calling and multi-step reasoning continue to be the default recommendation for local agent deployments. A sub-$0.50/M token model that doesn't hallucinate tool calls is more useful than a 400B frontier model that gets confused after three turns.

Community highlights

@lazy_shrey is running Hermes Agent with Mimo v2.5 Pro to manage a home server -- daily tech news digest at 9 AM, server health checks, and cron job management. This is the "personal infrastructure" use case that Hermes was designed for: replacing a patchwork of shell scripts and monitoring dashboards with a single agent that handles scheduling, alerting, and system administration.

@tonysimons_ posted a Hermes Agent Tip of the Day on browser_vision -- the tool that takes a screenshot of the active browser page and sends it through AI vision for tasks that don't live in the DOM. Captchas, visual puzzles, canvas-rendered UI -- anything a CSS selector can't reach.

@ElkimXOC discovered that Hermes Agent ships with Honcho, a persistent memory plugin that reasons about everything the user and agent do together. "The idea is that over time it will truly know you as a person," they wrote. Memory that doesn't just store -- but reasons about what to store -- is the direction the whole space is heading.

@Emily_brown_22, replying to a thread about skills as parameters, captured a broader shift: "Hermes Agent and OpenClaw workflows are moving from handcrafted toward measurable, optimizable procedural memory." This is the core bet behind SKILL.md and semantic skill retrieval -- treating agent instructions as data you can measure and improve, not static prompts you write once and forget.

The throughline across these posts is the same: Hermes Agent's users aren't just running demos. They're building persistent infrastructure -- home servers, daily automations, multi-model pipelines -- and memory is the feature that makes it stick.