Hermes’ rise on OpenRouter, in IOSG’s reading, was about product timing and trust
In an analysis written by Jacob Zhao, IOSG argues that Hermes’ breakout growth did not come from some exclusive technical invention that OpenClaw could not replicate in principle. The firm’s central claim is narrower and more product-focused: Hermes arrived at a key moment for the personal agent category, picked up a user base that OpenClaw had already educated, and then closed what IOSG describes as a challenger-style growth loop by offering a more convincing sense of delegatability.
The report says that as specialist execution agents keep improving, users still want a general-purpose manager that stays online, can be trusted over long periods, and acts as a persistent coordinator rather than just another model endpoint.
On OpenRouter’s public application leaderboard, Hermes Agent ranked first across the platform with 30.5 trillion tokens of usage, according to the report. It also ranked first in Productivity, Coding Agents, Personal Agents, and CLI Agents, ahead of OpenClaw and Claude Code. IOSG notes that the snapshot was taken on Aug. 4, 2026, and that the leaderboard is dynamic.
It also adds an important caveat. OpenRouter’s data does not cover all direct official API consumption across the industry, including native subscriptions such as Claude or Codex. Even so, IOSG says the platform is large enough in AI model routing and aggregation to make its rankings a useful directional signal.
The report does not argue that Hermes has displaced the top execution tools in every domain. For higher-end professional work — complex code generation, architecture design, and high-value data analysis — IOSG says many users still rely on Claude Code and ChatGPT. Hermes, though, remains stronger in background automation, message-based entry points, always-on monitoring, and lightweight task scheduling. That mix, the paper says, explains why a Web3-built agent product managed to generate unusual traction in reach, community activity, and usage intensity.
OpenClaw changed the unit of adoption from projects to individuals
IOSG frames OpenClaw as the product that made personal agents legible to the market. Before OpenClaw, the agent stack was already reasonably mature on the infrastructure side, but its adoption unit was usually a development project or enterprise workflow, not an individual user.
Earlier frameworks were aimed at developers and usually produced code or configuration. They built agent infrastructure, but they did not deliver an agent as a product. The engineering threshold stayed high, so the category remained stuck in the world of developer tools. What was largely missing, in IOSG’s telling, was a true product layer that turned the technology into something a person could treat as a long-term asset.
The article lays out a six-layer view of the agent stack: model layer, protocol layer, SDK and development framework layer, orchestration runtime layer, execution infrastructure layer, and deployment and governance layer. OpenClaw’s contribution, IOSG says, was not to reinvent the agent loop or task scheduling at the bottom of the stack. It was to wrap those scattered capabilities into a coherent product.
LangChain answered the question of how to build an agent. OpenClaw answered the question of how to own one, according to the report. IOSG says that shift showed up in six product-level moves: persistent identity and personality, daily-use interfaces such as Telegram and WhatsApp rather than IDEs or command lines, a resident always-on state, explicit operating permissions over files, browsers, terminals and real-world actions, extensible capabilities through Skills and Memory, and, most importantly, a shift in user mindset from using an AI tool to owning a dedicated digital companion.
The report also argues that OpenClaw’s popularity created a wave of imitators, many of which solved real pain points: complex installation, environment setup, missing access through channels such as WeChat and Feishu, compatibility with domestic models, rapid cloud deployment, enterprise permissions, auto-updates, and security isolation. But those products generally failed to establish distinct mindshare, IOSG says, because they answered the question of how to use OpenClaw more easily, not where the personal agent category should evolve after OpenClaw.
Why Hermes, and why Nous Research
To explain why Hermes emerged as the main challenger, IOSG spends substantial time on Nous Research. The company began in a Discord-based open-source AI research community in 2022 and formalized as a company in 2023. The founding team named in the report includes Jeffrey Quesnelle, Karan Malhotra, Teknium, and Shivani Mitra.
IOSG breaks Nous’ business into four major components. The first is the Hermes family of models, which the report describes as the company’s best-known open-source model brand, focused for a long time on post-training, instruction tuning, and agent capabilities, with wide adoption on Hugging Face. The second is DisTrO, or Distributed Training Over-the-Internet, aimed at reducing cross-node communication overhead and making collaborative training across geographies and heterogeneous hardware more practical over ordinary internet bandwidth. The third is Psyche, a decentralized training network that extends that approach by using Solana to coordinate globally distributed compute nodes. The fourth is Hermes Agent itself, which packages Hermes models, tool use, Memory, Skills, messaging channels, and long-running operation into a resident personal agent.
IOSG says Nous Research raised a $50 million Series A in April 2025 led by Paradigm, at a post-money token valuation of $1 billion. Before that round, the company had raised about $20 million in earlier financing from investors including Distributed Global, North Island Ventures, and Delphi Digital.
The article’s point is not just that Nous had funding. It argues that Hermes Agent emerged from a broader supply-side stack already in place: models, training, decentralized coordination, and open weights. In that framing, Hermes Agent was not a quick fork built to chase attention. It was Nous’ first major extension from supply-side capability into demand-side usage, real users, real tasks, and real workflows. That gave Hermes a deeper starting point than a standard clone.
Hermes turned OpenClaw’s operational pain into a growth engine
IOSG does not claim Hermes has a huge architectural gap over OpenClaw at the packaging level. Both are still combinations of models, tools, memory, and orchestration. The report says the breakout came from a system-level growth chain instead: Hermes offered migration tools, captured users that OpenClaw had already educated, and spoke directly to the operational burden those users had begun to feel.
That user handoff, in IOSG’s view, was Hermes’ earliest and most important growth engine.
The core product bet behind Hermes is what the report calls a transfer of operational responsibility. In practical terms, the promise is that when things go wrong, the system should absorb and repair the failure rather than pushing the burden back onto the user. IOSG organizes this around three kinds of trust:
- Reliability trust, which covers continued task progress and failure recovery through tools such as persistent Kanban, /goal mode, and tool self-healing.
- Safety trust, which covers overreach, accidental deletion, and data leakage through approval flows, sandboxing, and strict permission boundaries.
- Verifiable trust, which covers proof that work was actually completed through Completion Contract and Grounded Citations.
The paper then separates two concepts that are often blurred together in product messaging. Self-improvement, as IOSG defines it, is basically process adaptation built on Memory and Skills. Since competing products can assemble similar infrastructure, Hermes’ edge here is more about being early to package it as a default system with lifecycle management than about an unassailable technical moat.
Autonomous recovery is different. IOSG treats this as the more meaningful experiential difference today. Because Hermes uses structured error returns and automatic provider fallback, the system can absorb some failures internally. That, the report says, creates a more tangible product advantage because users feel the benefit directly when they are interrupted less often.
Hermes is strongest when it orchestrates, not when it replaces specialist agents
One of IOSG’s main arguments is that Hermes should not be judged as though it were trying to execute every specialist task by itself. Its core value, the report says, lies at the orchestration layer: filling in missing requirements, breaking down work, routing tasks, monitoring progress, and handling final acceptance. Through built-in Skills, Hermes can delegate execution to external CLIs such as Claude Code and Codex, and the community has already developed a practice pattern around that stack, including /goal workflows and the oh-my-hermes collaboration tool.
That is why IOSG describes Hermes as a control plane rather than a direct replacement. Specialist agents raise the ceiling on what can be done. Hermes raises the ceiling on what can be coordinated.
The report presents two operating modes to explain where Hermes is useful and where it is less so. In the first, which IOSG calls direct collaboration, the user still manually writes prompts in an LLM, sends them to an executor, carries output across tools, and reviews results by hand. Output quality may be high, but project management and multi-agent coordination remain manual. For users who prefer that level of involvement, Hermes can look like an extra opaque layer rather than a meaningful reduction in workload.
In the second mode, delegation management, the user treats Hermes as a standing coordinator and provides only the final objective. Hermes then breaks down work, assigns subtasks, tracks GitHub and CI status, and triggers rework automatically. Community examples such as oh-my-hermes, IOSG says, suggest that this is where Hermes’ value is clearest: replacing the cross-agent coordination and project-management overhead that users would otherwise have to do themselves.
Under that framework, Hermes is not in a simple substitute relationship with Claude Code or Codex. The specialist tools deliver execution quality. Hermes delivers persistence, cross-session state, and cross-agent coordination. IOSG adds that the value is unlikely to be evenly distributed across all users. It may be concentrated instead among advanced users running cross-agent, cross-system, asynchronous tasks over long periods.
The report includes a broad technical architecture view of Hermes Agent covering user entry points, a gateway, control core, provider layer, execution layer, orchestration layer, state layer, and governance layer. Based on official documentation and community research, IOSG pulls out three structural ideas. First, the control core includes context compression, provider fallback, and interrupted-state persistence, which together support recovery. Second, the execution layer places external tools such as Claude Code and Codex alongside Hermes-native tools such as Terminal and Browser, reinforcing the notion that Hermes dispatches and supervises rather than replacing specialist workers. Third, the learning and governance layer includes curator functions and Skill or Command approval, which suggests that accumulated system behavior is managed through governance with human intervention, not through a fully autonomous black box.
A crypto-native organization, with crypto mostly hidden from the product
IOSG argues that saying Hermes succeeded because it came from Web3 is too blunt. In the report’s view, Web3 gave Nous something closer to an organizational operating system, not a user-facing feature set. That operating system let the company bring a conventional AI product experience to market while still benefiting from structures that many AI startups do not have all at once.
The article lists four such advantages. One is patient capital from crypto-native investors, which supports long-duration bets, high uncertainty, and several parallel tracks at once, allowing Nous to work on models, training, runtime, and cloud together without narrowing too early to a single validated revenue line. Another is an existing user market: crypto AI users are often already comfortable with Telegram, servers, APIs, and self-hosting, which lowers educational friction and can produce intense early usage, tutorial sharing, and Skills contributions. A third is a user-sovereignty value system, expressed in the product through an MIT license, multi-provider support, BYOK, and portable Memory and Skills. The fourth is community-led development through global remote collaboration and open-source culture, where users can become contributors, Skill authors, and de facto product designers for vertical use cases.
At the same time, IOSG points out that Hermes keeps crypto largely out of sight for end users. To use the agent, memory system, skills, and automation functions, people do not need to connect a wallet, buy a token, or understand Solana. Yet in the background, Paradigm’s backing, Psyche, decentralized training, and crypto AI communities are still part of the picture.
The report summarizes this arrangement as “crypto-native in organization, crypto-invisible in product.” In other words, Nous keeps the parts of crypto it sees as most useful — capital, global communities, user-sovereignty values, and coordination capacity — while stripping away the parts most likely to get in the way of mainstream use, including wallets, tokens, speculative narratives, and onchain operational friction.
IOSG also compares Hermes with OpenClaw on crypto. It says the split is not best understood as one side embracing crypto and the other rejecting it on ideological grounds. Both products, from the standpoint of product outcomes, favor openness, user control, and reduced platform lock-in. The difference is in how they implement those values. OpenClaw, which the report describes as local-first sovereignty, rejects financialization at the product layer and, after early fake-token scams, maintains zero tolerance toward crypto. Hermes and Nous, which IOSG describes as cryptoeconomic sovereignty, treat blockchain as a practical coordination layer for engineering problems, such as using Solana inside Psyche to coordinate heterogeneous compute.
The business question is not API price but delegated work actually completed
When IOSG turns to monetization, it argues that a simple comparison between Hermes’ token costs and subscriptions to Claude Code or Codex misses the point. The product’s value, the report says, lies in replacing user time spent on project management, context shuffling, and cross-agent coordination.
The paper expresses that as a user value formula: value equals manual coordination time saved, plus the value of asynchronous and unattended operation, plus gains from cross-system automation, minus token and tool costs, minus manual intervention cost, minus failure and safety risk.
That means the economics depend heavily on depth of delegation. If Hermes can turn work that previously required hours of active supervision into truly unattended execution, then even somewhat higher token costs may still make sense. If users still have to intervene repeatedly to fix errors and rescue tasks, Hermes becomes a token sink and a fault amplifier instead.
IOSG says this explains why opinions on Hermes’ economics vary so sharply across user groups. The key metrics are not model API sticker prices, but unattended completion rate and the number of manual interventions required per task.
As for the commercial stack, Hermes Agent itself is open-sourced under the MIT license and positioned as an ecosystem growth engine. The more direct monetization layer sits inside Nous Portal, whose pitch is essentially one subscription for multiple classes of API keys. IOSG says it has three core parts: model routing across 252 models through OpenRouter and direct providers, a tool gateway integrating Firecrawl, FAL, Browser Use, Modal, and OpenAI Audio, and managed Hermes Cloud instances charged on a daily runtime basis, excluding inference and tool-call fees.
The report also highlights a structural constraint. By choosing MIT licensing, Hermes may gain adoption while losing control over value capture. If Hermes becomes a widely integrated optional runtime inside cloud providers, IOSG says it could face a problem similar to Linux or Kubernetes, where a large share of commercial value is captured by hosting and compute providers rather than by the open runtime itself. As long as users can freely choose self-hosting with their own APIs or third-party cloud deployments, large usage numbers do not automatically become direct revenue for Nous.
Hermes sits between open-source agent products, model vendors, and cloud platforms
IOSG places Hermes in a three-sided market structure involving OpenClaw, specialist tools such as Claude Code and Codex, and big-tech managed agent products. It argues these products are not all attacking the same job or the same user base.
For Hermes, the report says the product is not trying to be a mass-market assistant. Instead, it has found four dense groups of power users: self-hosting and infrastructure operators who already know VPS, Docker, and SSH; multi-model arbitrage users who do not want to be locked into one provider; multi-agent coordinators trying to automate complex workflows across tools and platforms; and members of open-source and crypto AI communities whose values line up closely with Nous’ culture. Those groups may be small in absolute number, but IOSG says they consume a lot of tokens, contribute code, and evangelize effectively, making them especially important in early growth.
Claude Code and Codex are described in the article as both suppliers and threats. In the near term, they are symbiotic with Hermes. Hermes routes work and handles acceptance. Codex can implement code, while Claude Code can help with architecture and review. As specialist agents get better, Hermes can support more complex workflows.
Longer term, IOSG sees pressure from model vendors moving upward into the control layer. It points to Anthropic’s Claude Managed Agents, which support parallel multi-agent orchestration, and OpenAI’s positioning of Codex App as a “command center for agents,” with support for multi-agent parallelism, automation, and long-running background operation. In IOSG’s view, that means Codex already has a fairly mature multi-agent control function inside the software engineering boundary, and in some places may exceed Hermes.
Hermes still has an edge, the report says, in cross-channel, cross-model, and cross-project personal control. But the central competitive question is whether Hermes can turn project state, approval rules, Skills, Memory, and cross-agent workflows into assets users do not want to move before model-native products absorb those same functions as standard features.
On the big-tech side, IOSG distinguishes between always-on personal agent hosting products such as Tencent QClaw and ByteDance ArkClaw, and broad office or enterprise work agents such as WorkBuddy and Trae. The distinction matters because the product boundaries are different. The article says one-click deployment, preset templates, and local ecosystem integrations can lower the barrier for cloud-managed agent products, but users still tend to believe those platforms ultimately want to steer usage toward their own clouds and models.
At the same time, IOSG notes that both ByteDance ArkClaw and Tencent Cloud have integrated Hermes Agent into their cloud consoles as an optional plugin or dedicated template. That amounts to a multi-runtime strategy: the platforms keep their own hosting, billing, security, and enterprise controls, while Hermes acts as a pluggable advanced component. In the report’s framing, that is not necessarily hostile competition. It can also be a complementary relationship between open runtimes and commercial cloud control planes.
Meanwhile, large internet companies are putting more emphasis on general office agents with clearer demand, easier acceptance criteria, and more direct monetization potential, such as WorkBuddy. Those products can tie more tightly into ecosystems such as WeChat, DingTalk, and Feishu and convert usage into revenue more directly.
IOSG’s closing view: Hermes offers a more mature path for crypto AI, but the hard test is still ahead
In its final section, IOSG argues that Web3 did not make Hermes a smarter agent by itself. What it changed was Nous’ capital structure, organizational form, seed-user base, and value system. From that perspective, Hermes represents a more mature crypto AI path: crypto as infrastructure and organizational coordination, not as an interface the user is forced to confront.
The report says Hermes has already moved from a crypto AI research brand into a global open-source agent product and established large attributable inference activity along with what it calls a clear second mindshare. Even so, IOSG says that mindshare is still concentrated in the OpenRouter ecosystem and among global developers. It has not yet translated into a comprehensive overtake of OpenClaw in GitHub stars or total community scale.
IOSG returns to the same point it made at the start. Hermes does not have a piece of magic technology that OpenClaw cannot copy. What it has done is study a challenger’s opening well: inherit high-intensity users, build a stronger sense of delegatability, pair that with a self-improving narrative, and redefine what users may expect from a long-term personal agent.
The paper ends with five implications. First, crypto can function as an organizational operating system rather than a product feature. Second, decentralized infrastructure needs a demand-side entry point if it is going to form a closed commercial loop, and Hermes Agent is Nous’ key validation of that move from training infrastructure to end-user demand. Third, a moat can be built on delegation trust, not just model capability. Fourth, relationships with cloud giants are not automatically zero-sum, because being integrated as a runtime can itself be a viable path, even if it creates a risk of value being captured elsewhere. Fifth, the endgame may move away from raw one-shot execution quality and toward ownership of the task, accumulation of trust, maintenance of long-term context, and orchestration of specialist workers.
OpenClaw, in IOSG’s framing, made “owning a personal agent” into a clear product category. Hermes pushes the category toward a more complete model of long-term delegation through persistent state, task recovery, evidence-based acceptance, multi-model supply, and specialist-agent routing. The unresolved question is whether users will still hand over long-term objectives and trust to an open runtime with Web3 roots once Claude Code, Codex, and cloud platforms continue to strengthen their own control layers — and whether they will keep paying for that privilege.