Why is Hermes Agent getting so much attention?

Is Hermes Agent really getting that much attention right now?

Hermes Agent is getting exceptional attention right now, far beyond what we normally see from a new open-source AI agent.

Nous Research's GitHub repository currently sits at roughly 238,000 stars and 48,000 forks. The speed matters more than the absolute number. Hermes had only 862 stars around its public launch, reached roughly 99,000 after eight weeks, and was already around 190,500 by June 11. That puts the increase since launch at roughly 275×.

The first explosive phase has cooled, which is normal after a GitHub breakout. Yet Hermes is still adding thousands of stars during strong weeks. GitStarClub recently recorded roughly 3,700 new stars in a single week, enough to keep Hermes among GitHub's fastest-growing large repositories.

That separates Hermes from the thousands of agent projects that briefly trend and disappear. The project reached a quarter-million-star order of magnitude within months and is still attracting developers after the initial launch wave.

Point in Hermes Agent's growth	Approximate GitHub stars	What it tells us
Around public launch	862	Tiny early developer project
Eight weeks later	99,000	Viral breakout
June 11	190,500	Attention survived the first launch wave
Currently	~238,000	Growth has slowed but remains substantial

Are people actually using Hermes Agent, or just starring it on GitHub?

Hermes Agent's usage looks very real: OpenRouter currently ranks it as the largest publicly tracked application on the platform by cumulative token consumption.

OpenRouter's public rankings show roughly 47 trillion tokens attributed to Hermes Agent. Claude Code is around 12.8 trillion and Kilo Code around 9 trillion. On that specific public dataset, Hermes has generated roughly 3.7 times Claude Code's cumulative token volume.

The trajectory is even more interesting. OpenRouter reported more than 17 trillion Hermes tokens on June 12. The cumulative total has since climbed to around 47 trillion, almost 2.8× higher. Its current trending panel also shows roughly 14.5 trillion Hermes tokens during the latest measured period, up about 32%.

We should be careful with what those numbers prove. Autonomous agents can consume huge amounts of context because one job may involve planning, browser calls, terminal commands, subagents and retries. Forty-seven trillion tokens certainly does not mean 47 trillion human prompts or millions of daily users.

Still, token volume answers the narrower question very well. People are doing far more than bookmarking the repository. Hermes installations are running enough actual work to place the project at the top of OpenRouter's public application rankings.

OpenRouter application	Cumulative tracked tokens	Hermes relative to it
Hermes Agent	~47T	1.0×
Claude Code	~12.8T	~3.7×
Kilo Code	~9.0T	~5.2×
Zazen	~6.8T	~6.9×

Why did Hermes Agent take off so fast?

Hermes Agent took off because it bundles nearly everything people wanted from a personal AI agent into one open system that can stay running after the chat ends.

Hermes can execute terminal commands, browse websites, manipulate files, schedule jobs, call subagents, use vision and connect to messaging services. Nous currently advertises more than 40 built-in tools. You can run the agent on your laptop, but the more interesting setup is often a cheap VPS where Hermes stays alive while you talk to it through Telegram, Slack, Discord or another interface.

That changes the experience quite a lot. You can message Hermes from your phone, ask it to inspect a server, leave it working remotely and receive the result somewhere else. The agent also keeps persistent memory and reusable skills between sessions, so the setup can accumulate procedures instead of starting from a blank prompt each time.

Nous removed another big friction point by keeping Hermes model-agnostic. Users can connect OpenAI, Anthropic, Nous, OpenRouter, local endpoints and many other providers. OpenRouter currently records 433 different models used through Hermes, so people are clearly taking advantage of that flexibility.

The timing was excellent too. Developers were already comfortable with coding agents, MCP servers, Telegram bots, self-hosted models and always-on assistants. Hermes packaged those ideas into something that feels more like one personal runtime.

That combination explains more of the growth than any magical capability. Hermes arrived when enough pieces of the agent stack had matured, then made them unusually easy to assemble around one persistent assistant.

Did OpenClaw help Hermes Agent become popular?

OpenClaw clearly helped create the audience Hermes Agent is now competing for.

OpenClaw had already taught a large group of technical users why they might want a self-hosted agent connected to messaging apps, files, browsers and shell commands. Hermes entered after that mental model was established, so Nous did not have to explain the entire category from scratch.

The migration tooling makes the competitive relationship unusually explicit. When Hermes detects an existing OpenClaw installation during setup, it can offer to import it. The dedicated hermes claw migrate workflow can bring across settings, memories, skills, API keys and other parts of the user's environment.

Companies rarely build migration tools for products whose users they do not care about winning.

OpenClaw's security problems also made the trade-offs around persistent agents much more visible. Its GitHub security page has published multiple advisories, including issues around command execution and environment handling. That created demand for projects that could keep the same always-on-agent idea while taking permissions, isolation and approvals seriously.

We found no reliable dataset showing that hundreds of thousands of people directly switched from OpenClaw to Hermes, so claims of a mass exodus would go too far. What we can see is a much simpler pattern: OpenClaw educated the market, Hermes targeted that market directly, and Nous made switching unusually easy.

Is Hermes Agent really self-improving?

Hermes Agent genuinely learns from previous work, although today's self-improvement is mostly about memory and reusable procedures rather than an AI rewriting its own intelligence.

Suppose Hermes spends 20 minutes figuring out how to deploy one of your projects. It can turn that successful procedure into a reusable skill. The next time a similar task appears, the agent can load those instructions instead of rediscovering every step.

Hermes also stores persistent information about the user and searches previous conversations when older context becomes useful. Those mechanisms give the agent something most disposable chat sessions lack: continuity.

Nous is experimenting with a more ambitious layer called Hermes Agent Self-Evolution. The separate project uses DSPy and GEPA to generate variations of skills, run evaluations and keep versions that perform better.

The current implementation is much narrower than the marketing phrase "self-evolving agent" can suggest. Nous's own roadmap lists skill-file optimization as implemented, while tool-description optimization, system-prompt evolution, code evolution and a continuous improvement loop remain planned.

The learning claim holds up, but in a practical form. Hermes can remember what worked, package procedures and reuse them. Recursive autonomous improvement of the entire agent remains a research project.

Does Hermes Agent actually get better the longer you use it?

Hermes Agent can become noticeably better adapted to one user over time, but we still lack good evidence showing a huge measurable performance jump after months of use.

Nous deliberately keeps its permanent memory fairly small. The main memory file is limited to roughly 2,200 characters and the user profile to around 1,375 characters. Larger procedures are stored separately as skills and pulled in when needed.

That design avoids a common memory problem. Dumping every old conversation into every new prompt would eventually produce an expensive mess. Hermes instead tries to keep a compact picture of who the user is while storing detailed procedures elsewhere.

Imagine someone who repeatedly uses Hermes for deployments. Over time the agent might know where the repositories live, which server hosts each service, which commands normally work, what checks should run after deployment and which mistakes happened before. That can remove a surprising amount of repeated explanation.

What we still do not have is a serious longitudinal benchmark. Nobody has publicly shown something like 1,000 new Hermes users, measured their task success in week one, then measured the same workflows three months later.

Hermes clearly becomes more personalized. Claims that it becomes dramatically more intelligent every week are still ahead of the evidence.

Why are people so excited that Hermes Agent can use almost any AI model?

Hermes Agent's model freedom gives users a way to keep their agent even when they change which AI model they prefer.

This is becoming increasingly valuable because the best model keeps changing. A user might prefer one model for coding today, another for research, a cheap Chinese model for repetitive automation and a local model for sensitive work.

Hermes lets those models sit behind the same memory, skills, schedules, tools and integrations. Switching models does not require rebuilding the personal agent from zero.

Current OpenRouter data gives us unusually good evidence that users care about this. Hermes has already generated traffic across 433 different models. That is much stronger than a feature page saying the product is "model agnostic."

It also changes where the durable value sits. The model can become a replaceable component while the user's accumulated environment stays in place.

That architecture protects Hermes from one of the biggest risks facing AI applications. Nous does not need to build the world's best foundation model every six months. If another company releases a much better one, Hermes can benefit from it almost immediately.

What are people actually using Hermes Agent for right now?

The strongest Hermes Agent use cases today are surprisingly specific: give the agent a recurring stream of information, a clear job and somewhere useful to put the result.

A recent discussion in the Hermes community included a user who forwards travel confirmations to an agent-owned inbox. Hermes builds the itinerary and reportedly caught a train booked for the wrong day. The same setup takes food photos through iMessage, logs nutrition and combines Garmin, sleep, exercise and food data into one personal journal.

Another user described giving Hermes access to email and calendar so it can send a morning briefing. The same person built a language tutor that remembers previous progress and changes the curriculum accordingly.

A construction-company operator recently described using Hermes for bidding with a retrieval system and for tracking construction progress through an ERP. Another operator running a conference-listing business uses scheduled Hermes jobs to find new conferences, check whether old events have announced next year's dates and scrape changes to agendas, speakers and sponsors.

Home automation is another recurring theme. One recent user created a dedicated Hermes profile that manages Home Assistant scripts and can answer questions such as why a particular light suddenly turned on.

Across these examples, the pattern is remarkably consistent. The useful setups behave like small pipes: confirmations come in and an itinerary comes out; calendar plus inbox becomes a morning briefing; new listings become buying recommendations; conference pages become structured updates.

The vaguer prompts are much less convincing. Recent community threads still contain people asking what Hermes can really do beyond notes, grocery lists and generic research. Some users report that broad content-writing workflows still require heavy rewrites even after trying several models.

That gap tells us where Hermes is strongest today. The software gives people powerful building blocks, but the user still has to define a job worth automating.

Is Hermes Agent actually better than Claude Code or OpenClaw?

Hermes Agent is already one of the strongest general-purpose agent harnesses we can benchmark, but it does not beat Claude Code, Codex or OpenClaw on every model.

WildClawBench gives us a useful comparison because it runs the same 60 human-authored tasks through several real agent environments. These are long jobs averaging many tool calls rather than tiny coding questions.

The harness itself makes a surprisingly large difference. On the same task set, changing the agent framework while keeping the model fixed can shift the final score by well over 10 percentage points and in some cases close to 18 points.

Hermes performs especially well with several non-OpenAI models. MiMo V2 Pro scores 48.1% with Hermes, compared with 40.2% through OpenClaw, 35.3% through Codex and 29.9% through Claude Code. GLM 5 also performs best through Hermes in the published comparison.

GPT-5.4 goes the other way. Codex scores 56.8%, ahead of Hermes at 50.7%.

There is a real Hermes advantage in some combinations, yet the benchmark gives us no universal winner. The underlying model and the harness interact.

That is also why Hermes and Claude Code are slightly awkward competitors. Claude Code remains heavily centered on software development. Hermes is trying to become a persistent general-purpose agent that can code, browse, message, remember, schedule jobs and run remotely.

Model	OpenClaw	Claude Code	Codex	Hermes Agent
GPT-5.4	50.3%	48.4%	56.8%	50.7%
GLM 5	42.6%	31.0%	38.9%	46.4%
MiMo V2 Pro	40.2%	29.9%	35.3%	48.1%
MiniMax M2.7	33.8%	32.0%	35.8%	37.1%

Is Hermes Agent shipping too fast?

Hermes Agent is shipping fast enough to keep people constantly rediscovering the product, and that speed is already creating visible stability costs.

The project changes unusually quickly. A major release earlier this year involved more than 300 contributors and substantial changes to the agent architecture. Since then, Hermes has kept adding desktop interfaces, profiles, new providers, messaging integrations, browser capabilities, security controls and infrastructure backends.

That pace feeds attention because someone returning after a few weeks can find a noticeably different product.

The other side is visible directly on GitHub. The repository currently shows more than 5,000 issues and more than 5,000 pull requests. Raw counts in a project this popular include duplicates, low-priority requests and automated activity, but the order of magnitude still tells us that maintainers are handling an enormous stream of changes.

Recent bug reports make the problem concrete. One issue found profile-scoped credentials that caused managed tools to appear unavailable. Another found that auxiliary tasks could hit a hard timeout that made slower self-hosted models difficult to use. A separate cron bug created a new session row on every execution rather than cleanly reusing a job session.

None of those bugs kills the project. Together they show the price of extreme velocity.

Hermes currently feels closer to rapidly moving infrastructure than a mature utility you install once and forget for three years. For developers who enjoy experimenting, that pace is part of the attraction. People running critical automations will probably care more about stability than the next feature drop.

Is Hermes Agent actually cheap to run?

Hermes Agent can be very cheap with the right models, while heavy autonomous workflows can burn through an absurd number of tokens.

The software itself is MIT-licensed. Nous says Hermes can run on a VPS costing around $5 per month, and local models can reduce API spending further.

Compute for the agent runtime is therefore rarely the expensive part. Inference usually dominates.

WildClawBench illustrates the range. Under its controlled tasks, Hermes averaged around $0.11 per task with MiniMax M2.7, $0.26 with MiMo V2 Pro and roughly $0.44 with GPT-5.4 or GLM 5. Those tasks are standardized benchmark jobs, so real-world costs can vary enormously.

Community reports show how quickly usage can grow. In a recent discussion about token efficiency, one user said a fresh Hermes setup consumed 147 million tokens in a day. Another reported roughly 107 million tokens during one week despite feeling they had barely used the agent. Those are individual anecdotes rather than representative averages, but they fit the architecture: skills, long context, browser calls, tool loops and subagents add up quickly.

Cheap models change the equation dramatically. Several active users now describe pairing Hermes with inexpensive DeepSeek-family models for routine work while reserving expensive frontier models for harder tasks.

That flexibility is one of Hermes's practical strengths. A $5 server does not guarantee a $5 agent, but users have unusually fine control over where the expensive intelligence gets used.

Is Hermes Agent safe enough to leave running all day?

Hermes Agent has serious security controls, but we would still isolate it before giving the agent broad access to email, credentials, files and shell commands.

Nous currently documents eight security layers, including caller authorization, approval for dangerous commands, file-write protections, container isolation, credential filtering for MCP processes, prompt-injection scanning, cross-session isolation and input sanitization.

The project's GitHub security page currently shows no published Hermes security advisories. That is encouraging, although it cannot tell us that the code is vulnerability-free.

There have already been security-related engineering problems. Users have reported known high-severity vulnerabilities in some pinned dependencies, and the repository has tracked other hardening issues around package versions, credentials and agent isolation.

Nous's own security policy gives the most sensible advice: use a real isolation boundary when the agent will process untrusted material. The documentation recommends non-root execution, hardened containers or other sandboxed environments, restricted network exposure and careful review of third-party skills and plugins.

That advice becomes especially important with a product like Hermes because its usefulness comes from giving the agent real capabilities. A chatbot that can only produce text has a limited blast radius. Hermes can potentially read files, execute commands, open web pages and interact with external systems.

The practical answer is straightforward. Running Hermes continuously can make sense, but giving a powerful agent unrestricted access to a valuable production machine still looks reckless today.

What is still overhyped about Hermes Agent?

The overhyped part of Hermes Agent is the idea that we already have a dependable autonomous employee.

The independent benchmark data is enough to kill that interpretation. On WildClawBench's difficult real-world tasks, Hermes combinations generally land somewhere around the high 30s to low 50s in completion score. Even the strongest agent setups in the broader benchmark remain far from perfect reliability.

That level is impressive for software operating tools over long sequences, but it is nowhere near the standard we expect from a trusted human operator.

Community behavior points to the same limitation. Even these days, popular Hermes discussions keep asking variations of "what should I actually use this for?" The latest examples include users saying they struggle to find reliable workflows beyond notes, research and simple organization.

A product with a universally obvious killer use case does not generate that question quite so often.

The strongest setups usually involve humans narrowing the job first. Someone defines which emails matter, which websites to check, where outputs belong, what tools can be used and what counts as success. Hermes can then automate a surprising amount inside those boundaries.

Give the agent a fuzzy instruction like "manage my company" and the experience gets much less dependable.

The excitement around persistent agents is justified. The idea that Hermes has already solved autonomous knowledge work is several steps ahead of reality.

Can Hermes Agent keep growing after the GitHub hype fades?

Hermes Agent's GitHub explosion is slowing, but its current usage is still growing fast enough that the project has probably moved beyond a short-lived hype cycle.

The star curve makes the slowdown obvious. Hermes jumped from 862 stars near launch to roughly 190,500 by June 11. It currently sits around 238,000. That means it has added roughly another 25% since reaching an already enormous scale, but the early near-vertical phase has clearly ended.

Actual model usage has behaved differently.

As we saw above, OpenRouter had already counted more than 17 trillion Hermes tokens in June. The current cumulative figure is around 47 trillion, while Hermes also remains first in OpenRouter's public app rankings and is still growing strongly in its latest trending window.

That divergence is healthy. GitHub discovery does not need to accelerate forever if existing users keep running more work through the agent.

The unknown number is retention. We still lack credible public data for monthly active Hermes installations, three-month retention, jobs completed per active user or how many people install Hermes and abandon it after a weekend.

Those numbers will eventually tell us much more than GitHub stars.

The best case for Hermes is that users gradually build enough memory, skills, cron jobs, integrations and personal workflows that switching the agent off becomes annoying. If that starts happening at scale, Hermes will have something much more durable than open-source hype.

Why is Hermes Agent getting so much attention?

Hermes Agent is getting so much attention because it makes the idea of a persistent personal AI agent feel genuinely usable right now.

Nous brought together several ideas that developers had already been experimenting with: persistent memory, reusable skills, browser access, shell commands, scheduled tasks, subagents, messaging apps, remote servers and interchangeable AI models. Hermes then wrapped those pieces around one simple idea: keep the agent around while the models underneath it change.

That is a powerful proposition these days.

A developer can run Hermes on a cheap remote machine, message it from a phone, let it work with real tools, teach it a procedure and keep that procedure for the next model upgrade. OpenClaw had already shown people why this category was interesting, and Hermes lowered the friction for people who wanted a more extensible alternative.

The attention is also being reinforced by actual behavior. People are using Hermes for coding, server administration, inbox processing, travel planning, personal tracking, research, scheduled data collection and business workflows. Independent benchmarking shows that the Hermes harness can materially improve some models' ability to finish long agent tasks.

There is plenty left to fix. Reliability is still too low for blind autonomy, token consumption can get ugly, rapidly changing code creates bugs, and many new users still struggle to identify a workflow worth automating.

Those weaknesses do not really contradict the current excitement. They explain why Hermes feels important before it feels finished.

The attention is mostly deserved. Hermes has become one of the clearest demonstrations of where AI agents are moving: away from isolated chats and toward persistent software that remembers the user, stays connected to tools and survives the constant rotation of underlying AI models.

That future is still rough around the edges. Hermes is getting attention because, for the first time, a very large group of developers can actually run a version of it themselves.