Hermes’ phenomenal growth did not come from some proprietary technology that OpenClaw could not replicate in principle. Rather, at the critical window when the personal-agent category was taking shape, Hermes most precisely closed the loop on a "challenger growth system": it captured the user base OpenClaw had already educated, and built "Delegation Trust" — a differentiator closer to real, felt experience than the "self-improvement" narrative. As professional execution agents keep getting stronger, users still need an always-on steward they can actually trust with the job.
Open OpenRouter’s public app leaderboard and Hermes Agent ranks first across the entire platform by cumulative token volume, at 34.9 trillion (34.9T) tokens, while also ranking #1 in Productivity, Coding Agents, Personal Agents, and CLI Agents — a commanding lead over well-known agents like OpenClaw and Claude Code.
Figure 1 · Hermes Agent’s historical snapshot on OpenRouter (captured August 13, 2026)
OpenRouter can't see traffic through direct official APIs (native Claude or Codex subscriptions), but as the largest AI routing platform around, its leaderboard still means something. And on OpenRouter's own Coding Agents leaderboard, Hermes actually leads Claude Code & Codex by a wide margin — though it's unclear how much of that is genuine coding work versus delegated or background calls. Hermes' clearer edge is elsewhere: background automation, message-based tasks, always-on monitoring, and lightweight scheduling. As a product built by a Web3 team, Hermes has hit a level of distribution, community, and usage success that far exceeds expectations:
Why has Hermes managed to overtake in OpenRouter inference volume?
Where does the real dividing line between Hermes and OpenClaw actually lie?
In its relationship with Claude Code and Codex, how does Hermes stay "differentiated and coexisting" rather than competing head-on?
Before OpenClaw, the agent space already had mature infrastructure, but it suffered from a fundamental limitation: its unit of adoption was the "development project" or "enterprise workflow," not the "individual user." Early frameworks shared a common trait — they targeted developers and produced code or configuration. They built the infrastructure for agents without ever delivering the agent itself. The engineering bar was so high that these tools stayed permanently stuck at the "developer tool" stage, never closing the loop into a "personal asset" a regular person could own. The "personal agent product layer" facing end users directly was essentially empty.
Figure 0 · The six-layer Agent tech stack (Model → Protocol → SDK/Framework → Orchestration/Runtime → Execution Infrastructure → Deployment/Governance)
OpenClaw (@openclaw) didn't reinvent the agent loop or task scheduling — its real contribution was productization. LangChain answers "how do you build an agent"; OpenClaw answers "how do you own one." It skipped the middle layers of the stack and turned scattered framework capabilities into a complete product a person could configure and run long-term — shifting the unit of adoption from project to individual, across six dimensions:
Identity personification: giving the agent a persistent name and identity, breaking the stateless, tool-like feel of an API call.
Everyday entry points: using high-frequency messaging apps like Telegram/WhatsApp as the interface, replacing the complexity of a command line or IDE.
Persistent presence: running as a long-lived background process, shifting from passive "standing by" to active "being there."
Entitled permissions: deeply embedding the user’s file system, browser, terminal, and real-world action capability into the agent’s operating scope.
Extensible capability: turning accumulated workflows into reusable capability through Skills, Memory, and community plugins.
Mindshare ownership: the most fundamental shift — users moved from "using an AI tool" to "owning a personal digital companion."
OpenClaw’s popularity spawned a wave of imitators. These products solved real user pain points: cumbersome installation, difficult environment setup, missing WeChat/Feishu channels, compatibility with domestic models, fast cloud deployment, enterprise permission management, automatic updates, and security isolation. Each had its own user base and a reasonable business logic. But almost none built independent brand mindshare — because the question they answered was "how do you make OpenClaw easier to use," not "where should the personal agent evolve to after OpenClaw." The position of narrative challenger remains extremely scarce across the entire personal-agent market.
Nous Research(@NousResearch) originated in a Discord-based open AI research community in 2022 and formally incorporated in 2023. Its core founding team includes Dillon Rolnick(@dillonrolnick), Jeffrey Quesnelle(@theemozilla), Karan Malhotra(@karan4d), Teknium(@Teknium), Jai (@JSupa15) and Shivani Mitra(@shivani_3000), and its business spans:
The Hermes model series: Nous’ flagship open-weight model brand, long focused on post-training, instruction tuning, and agentic capability, with a large developer adoption base on Hugging Face.
DisTrO (Distributed Training Over-the-Internet): dramatically reduces cross-node communication overhead in distributed training, breaking the assumption that GPUs must sit in a single, tightly-interconnected data center — enabling cross-region, heterogeneous-compute collaborative training.
The Psyche decentralized training network: takes DisTrO further, using Solana to coordinate globally distributed compute nodes so GPUs across different networks and hardware environments can jointly participate in large-model training.
Hermes Agent: Nous’ consumer-facing personal agent product, integrating Hermes models, tool calling, Memory, Skills, messaging channels, and always-on operation into a persistent agent.
In April 2025,@NousResearch closed a $50M Series A led by crypto VC giant Paradigm (@paradigm), at a post-money token valuation of $1 billion. Before that round, the company had already raised roughly $20M in early-stage funding from investors including Distributed Global(@DistributedG), North Island Ventures(@NorthIslandVC), and Delphi Digital(@Delphi_Digital).
Nous built a full loop: Hermes (models), DisTrO (distributed training), Psyche (decentralized compute), and Hermes Agent (the product people actually use). Hermes Agent wasn't a trend-chasing fork — it was Nous' first move from years of supply-side work (data, models, training) into the demand side (real users, real workflows). That's a deeper starting point than an ordinary clone, but whether it turns into a durable, hard-to-copy edge still comes down to retention and revenue data no one has yet.
At the packaging level (model + tools + Memory + orchestration), Hermes and OpenClaw don’t differ dramatically, and complex professional work still flows to Claude Code or Codex. Hermes’ breakout growth was not driven by a technical generation gap — it precisely closed a systemic growth causal chain: through seamless migration tooling, it directly captured the user base OpenClaw had already educated and that was suffering from operational pain, skipping cold-start entirely and immediately acquiring high-intensity, high-virality Power Users — the core engine of its early growth.
The Product Leap: Building "Delegation Trust"
Hermes’ core product hypothesis is solving "operational burden transfer" — promising that "after something breaks, the system absorbs the fix internally." To this end, Hermes built a framework tightly binding three kinds of trust to three kinds of ownership:
Reliability trust: ensuring tasks keep moving and recover from failure (persistent Kanban, /goal mode, self-healing tools).
Safety trust: preventing privilege escalation, accidental deletion, or data leakage (Approvals workflow, sandboxing, strict permission boundaries).
Verifiable trust: proving a task is genuinely complete (Completion Contracts and Grounded Citations).
Conceptual Distinction: "Self-Improvement" (Narrative Edge) vs. "Autonomous Recovery" (Experiential Difference)
In Hermes’ product narrative, "self-improvement" and "autonomous recovery" carry meaningfully different levels of product value:
Self-improvement: fundamentally process adaptation built on Memory and Skills. Since competing products have similar underlying infrastructure, Hermes’ real edge lies more in being first to integrate this into a default system with lifecycle management — capturing the narrative advantage of "it grows with you," rather than an unproven, insurmountable technical moat.
Autonomous recovery: this is currently the most worth verifying experiential difference. Thanks to structured error returns and automatic provider fallback, Hermes can absorb failures inside the system. This "doesn’t keep bothering the user" system-level stability is a more direct, perceptible product strength — though its quantified edge over OpenClaw still awaits head-to-head comparative data.
Architectural Dividend: The Ability to Delegate to and Supervise Professional Agents
Hermes’ core value isn’t executing every professional task itself, but acting as an orchestrator that handles requirement completion, task decomposition, routing/monitoring, and final acceptance. Through built-in skills that delegate to external CLIs like Claude Code/Codex for the actual work, the community has converged on a practice pattern of "Hermes as orchestrator + external CLI as worker" (via the /goal mechanism and tools like oh-my-hermes), reflecting an architectural advantage: orchestrating professional agents to raise the ceiling on task complexity it can handle.
Reducing Hermes’ success simply to "Web3 background" is an oversimplification. What Web3 actually gave Nous is an "organizational operating system" that most other AI startups struggle to obtain simultaneously, letting it enter the mainstream market with the smooth experience of a standard AI product:
Patient risk capital: crypto-native money tolerates long timelines and parallel bets, so Nous could build models, training, Runtime, and Cloud at once instead of rushing to one revenue path.
A ready-made user market: Crypto AI users already run servers, APIs, and self-hosted setups, killing the cold-start problem and fueling early usage and Skills contributions.
User-sovereignty values: a stance of self-hosting, openness, portability, and anti-lock-in, translated directly into MIT licensing, multi-provider support, BYOK, and portable Memory/Skills architecture.
Community R&D and verticalization: built on global remote collaboration and open-source culture, users spontaneously become contributors, skill authors, and designers of vertical use cases.
Hermes shows almost none of this at the front end — no wallet, no token, no Solana required. But Paradigm's capital, Psyche, and the Crypto AI community are all still running the show backstage. Crypto-native org, crypto-invisible product: keep what's valuable, cut what blocks adoption.
Why OpenClaw Rejects Crypto, and Why Hermes Hides It
The OpenClaw-vs-Hermes split on Crypto isn't really "rejection vs. embrace" — it's two paths to the same goal: users owning their own AI. Both resist Big Tech monopolies and champion sovereignty; they just disagree on means.
OpenClaw (Local-first Sovereignty): resists financial speculation and defends "local-first" sovereignty. After early encounters with fake-token scams, it adopted zero tolerance toward Crypto. Through pure open source and local execution, it defends user sovereignty without touching a blockchain, firmly rejecting any financialization at the product level.
Hermes/Nous (Cryptoeconomic Sovereignty): an engineering-driven stance where Crypto is purely an underlying coordination tool. Introducing a blockchain is a pragmatic response to a specific engineering challenge (e.g., the Psyche network using Solana to coordinate heterogeneous compute), not the construction of a financial narrative for end users.
This section addresses a more fundamental question: when Claude Code and Codex can already handle most professional execution tasks at high quality, what’s the reason for Hermes to exist as an independent product?
Mode A: Direct Collaboration (rated average): the user is used to manually generating prompts in an LLM and handing them off for execution, manually shuttling results and reviewing them. Single-shot output quality is high, but the user still owns all project management and multi-agent coordination. For this hands-on type of user, Hermes’ automation is perceived as "an opaque middle layer" rather than something that actually reduces their workload.
Mode B: Delegated Management (rated high): the user treats Hermes as an always-on orchestrator and only hands it a final goal. Hermes handles task decomposition, delegates subtasks, tracks GitHub/CI status, and automatically triggers rework. Community practice (e.g., oh-my-hermes) shows that Hermes’ core value is precisely replacing the tedious work of cross-agent coordination and project management.
Within this framework, Hermes and Claude Code/Codex aren’t substitutes for one another — they sit in different layers: the latter provides execution quality at Layer 3, while the former provides persistence, cross-session state, and cross-agent coordination at Layer 2. Hermes’ value isn’t evenly distributed across all users — it likely concentrates heavily among advanced users running cross-agent, cross-system, long-running asynchronous tasks. That’s a more precise claim than the blanket assertion that "personal agents’ second mindshare has already formed," and it’s more useful for guiding commercialization and product priorities.
Architecture Overview (derived from official documentation and community research):
Figure 2 · Hermes Agent’s full technical architecture (User Entry → Gateway → Control Core → Provider Layer → Execution Layer → Orchestration Layer → State Layer → Governance Layer)
Based on official documentation and community research, this architecture spans the full path from user interaction to learning governance, and shows three key design traits:
System-level support for autonomous recovery: the "control core" explicitly includes context compression, provider fallback, and interrupt/state saving — architectural backing for self-healing when a task fails.
A "delegate, don’t replace" execution logic: the "tools and professional execution" layer places Claude Code, Codex, and other external CLIs on equal footing with Hermes’ own native tools (Terminal, Browser, etc.), confirming its role as an orchestration hub.
A governed form of "self-improvement": the "learning, maintenance & governance" layer includes nodes like Curator and Skill/Command Approval, showing that its experience accumulation is a governed process with human-in-the-loop checkpoints — not a fully automatic black box.
Evaluating Hermes’ economics by simply comparing its token cost against a direct Claude Code/Codex subscription leads to a misleading conclusion, because that comparison ignores Hermes’ core value: replacing the project management, context-shuttling, and cross-agent coordination work a user would otherwise do by hand.
User Value Formula Hermes User Value = Time Saved on Manual Coordination + Async/Unattended-Operation Value + Cross-System Automation Gains − Token & Tool Costs − Manual-Intervention Costs − Failure & Safety Risk
Hermes’ economics are therefore not absolute — they depend heavily on the user’s "delegation depth":
High delegation depth (economics hold up): if Hermes can turn a task that would otherwise require hours of manual babysitting into genuinely unattended execution, the overall time cost and efficiency gain stay net positive even if token cost is somewhat higher.
Low delegation depth (economics collapse): if the user still has to intervene frequently to fix errors and put out fires, Hermes becomes little more than a pure token burner and failure amplifier.
This mechanism precisely explains why different user segments rate Hermes’ economics so differently, and points to what actually matters for validating the business logic: quantifying the "unattended completion rate" and "manual interventions per task," not simply comparing model API unit prices.
Hermes Agent is open-sourced under the MIT license and positioned as a growth engine for the ecosystem. The real commercial loop is concentrated in Nous Portal, whose core value proposition is "one subscription, all your API keys," spanning three modules: (Source: portal.nousresearch.com/info)
Model routing: aggregates 300+ models (inference via OpenRouter and direct provider connections).
Tool Gateway: built-in access to Firecrawl (web search), FAL (image generation), Browser Use (cloud browser), Modal (sandboxed execution), and OpenAI Audio (TTS).
Hosting: ready-to-use Hermes Cloud instances (billed daily for uptime, excluding inference and tool-call costs).
Nous’ actual revenue depends heavily on which path the user takes, and shows a clear structural split:
MIT licensing drove Hermes' growth, but it's also a commercialization trap. If it's free to self-host, the paid tier needs a reason to exist — and that reason isn't clear yet. The bigger risk is value leakage: if cloud providers keep bundling Hermes in as an optional Runtime, Nous could end up like Linux or Kubernetes — massive adoption, but the money goes to whoever sells the compute and hosting, not the project itself. MIT bought ecosystem growth at the cost of controlling distribution. As long as users can self-host or run it through a third-party cloud, usage can't be forced into revenue — leaving Nous with real ecosystem influence and no guarantee of a payoff to match it.
Why aren’t Claude Code, Codex, WorkBuddy, Trae, OpenClaw, and Hermes the same kind of product? Because all of them are end products sitting atop the same six-layer infrastructure, serving different types of users and tasks:
Hermes hasn’t chased the mass market — it has precisely targeted four types of high-density Power Users, who form the bedrock of its viral, phenomenal spread:
Self-hosting and infrastructure enthusiasts: comfortable with VPS/Docker/SSH, treating Hermes as the natural control layer atop existing infrastructure.
Multi-model arbitrageurs: refuse to be locked into a single vendor, and routinely switch between frontier or local models depending on the task.
Multi-agent coordinators: urgently need to automate complex, cross-platform, cross-tool workflows that used to be stitched together by hand.
Open-source and Crypto AI communities: strongly aligned with user sovereignty and decentralization, resonating deeply with Nous’ organizational culture.
This group is small in absolute numbers but carries extremely high token consumption, code contribution, and technical evangelism — the core engine driving early technology products across the chasm, achieving high-density validation and word-of-mouth spread.
Hermes’ edge on the "task ownership / orchestration" dimension currently shows up mainly in the breadth of cross-channel, cross-life-context scheduling. Its lead on this dimension is very likely to be narrowed faster than its lead on other dimensions like "execution quality" or "always-on presence."
Short-term symbiosis: raising the execution ceiling
Hermes orchestrates, delegating implementation to Codex and architecture/review to Claude Code — a pattern the author has observed, not a benchmarked industry rule. The stronger those agents get, the more complex a task Hermes can pull off. Hermes routes and signs off; the professional agents execute.
Long term: a real threat to Hermes’ independent value
Long term, it's a threat. Model vendors are moving into orchestration faster than expected — Anthropic's Claude Managed Agents already does parallel multi-agent work, and OpenAI now pitches the Codex App as a "command center for agents." Codex's orchestration inside software engineering is already maturing, in some ways past Hermes — it's no longer just an executor.
Hermes still wins on breadth — channels, models, projects. But Codex may already out-compete it within software engineering specifically. The real question: can Hermes lock in users' project state, approval rules, Skills, and cross-agent workflows before model vendors just build it in natively?
Examining how big tech is responding to the personal-agent wave first requires clarifying its product boundaries: persistent personal-agent hosting aimed at individuals (e.g., Tencent QClaw, ByteDance ArkClaw) is a fundamentally different positioning from general-purpose work agents aimed at office/enterprise use (e.g., WorkBuddy, Trae):
The big-tech path (retail-ized / office-ized): — one-click deploy, preset templates, local integration. The catch: no matter how many outside models they support, users assume the real goal is funneling them into the company's own cloud and models (Volcano Engine, Bailian).
The Hermes path (open Runtime / infrastructure-ization): ArkClaw and Tencent Cloud both now offer Hermes as an optional plugin or template. Big tech keeps hosting, billing, and security; Hermes is a pluggable high-end option, not a replacement. Open ecosystem, commercial cloud, coexisting.
Big tech is redirecting its core resources toward professional workflows and enterprise control platforms with clear requirements, easy acceptance criteria, and direct monetization (e.g., Trae, WorkBuddy). These tasks integrate deeply with proprietary ecosystems like WeChat, DingTalk, and Feishu and convert directly to revenue. Meanwhile, a "model-neutral, self-hosted steward that can call a competitor’s models" sits in real tension with big tech’s core interest in a closed ecosystem loop, and so is treated only as an ecosystem supplement.
Web3 didn't make Hermes a smarter agent — it gave Nous a capital structure, an early user base, and a value system most AI startups don't have. Hermes' pitch to Crypto AI is simple: let Crypto be the infrastructure, not the interface users have to deal with.
Hermes has gone from Crypto AI research brand to widely-used open-source agent, with real inference volume and a genuine second mindshare — though that's still concentrated in OpenRouter and developer circles, not a GitHub-stars or community-scale win over OpenClaw. It got there without some technology OpenClaw couldn't copy — just by capturing power users early, building "Delegation Trust" (a real hypothesis, still unproven at scale), and coexisting with big-tech agent products instead of fighting them.
Five takeaways for Crypto AI:
Crypto can be an operating system, not a feature. Capital, an early user base, and values don't require exposing wallets or tokens. Crypto-native org, crypto-invisible product — that's the playbook.
Decentralized infra needs a demand-side hook. Supply-side networks like DisTrO or Psyche can't prove value alone. Hermes Agent is Nous' actual bridge to real users.
The moat can be trust, not just model quality. Whether users hand over long-term responsibility matters as much as raw execution — and it's underrated.
Big tech isn't zero-sum. ArkClaw running Hermes as an optional Runtime shows open Runtimes and cloud platforms can coexist — though the cloud layer may capture most of the value.
The endgame is task ownership, not single-shot execution. The most valuable layer may be whoever holds the goal and routes the work, not whoever executes fastest.
OpenClaw made owning a personal agent a real category. Hermes is trying to make delegating to one real — through persistent state, recovery, and multi-model orchestration. It's proven distribution and usage. It hasn't proven better task success, less babysitting, or real retention. The actual test: as Claude Code and Codex get better at orchestration themselves, will users still trust their goals to an open Web3-built Runtime — and keep paying for it?
Disclaimer: Thanks to the Hermes team and @brucexu_eth for their feedback during the writing of this piece. This report was written with the assistance of Claude Opus 5, ChatGPT-5.5, and Qwen 3.7. The author has made every effort to fact-check and verify the information for accuracy, but errors or omissions may remain — readers' understanding is appreciated. This content is intended solely for information synthesis and academic/research exchange.