Personal AI Agents Are Dominating LLM Usage — And They're All Open Source
Something interesting is happening in the OpenRouter global rankings. The leaderboard that tracks real-world LLM token consumption across hundreds of thousands of users tells a story that most industry narratives are missing: personal AI agents — not coding copilots, not enterprise chatbots — are now the largest consumers of LLM inference on the planet.
The Numbers
Here’s what the rankings look like today:
| Rank | Agent | Category | Tokens |
|---|---|---|---|
| 1 | Hermes Agent (Nous Research) | Personal Agent | 271B |
| 2 | OpenClaw | Personal Agent | 245B |
| 3 | Kilo Code | Coding Agent | 149B |
| 4 | Claude Code (Anthropic) | Coding Agent | 79.2B |
| 5 | Descript | Video/Podcast | 48.9B |
The top two spots belong to personal agents. Four of the top five are open source. The only proprietary entry is Claude Code at #4. These aren’t vanity metrics — OpenRouter routes real inference traffic, and token counts reflect actual sustained usage across the platform.
What’s Actually Running
To understand why these numbers matter, you need to understand what these agents actually do — because they’re nothing like ChatGPT.
Hermes Agent, built by Nous Research and released in February 2026, is a self-improving personal agent. It doesn’t just respond to prompts — it builds reusable skills from experience, improves them through use, and maintains persistent memory across sessions. It ships with 40+ built-in tools (web search, browser automation, vision, code execution), supports scheduled automations and subagents, and runs on seven terminal backends including Docker, SSH, and Modal. It’s model-agnostic: swap providers with a single command.
OpenClaw takes a different approach with the same philosophy. It connects to 25+ messaging platforms — WhatsApp, Telegram, Discord, Signal, iMessage, Slack, Teams, Matrix, LINE, and more — turning whatever app you already use into a control surface for an AI agent. It runs locally, is MIT licensed, and was recently the subject of a freeCodeCamp deep-dive. The agent manages files, browses the web, executes commands, sends messages, and maintains long-term memory — all from your own hardware.
Both share a critical design principle: they meet you where you already are. No new app to learn, no new interface to adopt. Your Telegram becomes your agent’s interface. Your terminal becomes its workspace.
Why Personal Agents Outrank Coding Tools
Kilo Code and Claude Code are excellent tools — Kilo supports 500+ models across VS Code, JetBrains, and CLI; Claude Code is Anthropic’s own CLI agent. But they serve a bounded use case: writing and editing software. You invoke them when you need code, and they stop when you’re done.
Personal agents don’t stop. They’re always-on background processes — running cron jobs, monitoring inboxes, checking calendars, sending scheduled messages, transcribing voice memos, coordinating across platforms. That “always-on” quality is what drives the massive token counts. Hermes Agent’s 271B tokens aren’t from a few intense coding sessions — they’re from thousands of users running persistent agents 24/7.
This is the shift the industry hasn’t fully named yet. We went from “tools you use” to “agents that run.” The interaction model moved from synchronous (open app → type prompt → get response → close app) to asynchronous (agent runs continuously → surfaces what matters → takes action when needed). It’s the difference between checking your email and having a secretary who handles your email.
The Guardrails Gap: Where’s NemoClaw?
One conspicuous absence from the rankings deserves attention: NVIDIA’s NemoClaw.
Launched at GTC in March 2026 with significant fanfare, NemoClaw is literally built on top of OpenClaw. It adds enterprise-grade guardrails via NVIDIA’s OpenShell runtime, security policies from NVIDIA Agent Toolkit, and optimized local inference using Nemotron models on DGX and RTX hardware. NVIDIA’s own press release called OpenClaw “the operating system for personal AI” and positioned NemoClaw as the safe, enterprise-ready way to run it.
NemoClaw has NVIDIA’s full brand weight, engineering resources, and distribution channels. And it’s nowhere on the leaderboard. The base OpenClaw project sits at #2 with 245 billion tokens. The enterprise-hardened version? Not even registering.
This tells us something important about what drives adoption in 2026. NemoClaw optimizes for compliance: content filtering, audit trails, policy enforcement, controlled behavior. The agents dominating the rankings optimize for capability: more tools, more integrations, more autonomy, less friction. When individual users — developers, researchers, power users — choose their own agent, they consistently choose the one that does more, not the one that restricts more.
This isn’t unique to AI agents. It’s the same pattern we’ve seen in every technology wave: enterprise wrappers that add governance layers rarely outperform the underlying open-source project they’re wrapping. Docker vs. enterprise Kubernetes distributions. Linux vs. enterprise Linux. The base layer wins on adoption; the enterprise layer wins on contracts. Both can succeed, but the token counts make clear which one developers actually run.
The Open Source Pattern
Four of the top five agents are open source. This isn’t a coincidence — it’s a structural advantage for this category.
Personal agents require extraordinary trust. They connect to your messaging apps, read your files, access your calendar, and send messages on your behalf. They maintain persistent memory of your conversations and habits. Handing that level of access to a closed-source cloud service is a hard sell for anyone who thinks about it for more than a minute.
Open source also enables the model-agnostic architecture that defines this generation of agents. Hermes Agent, OpenClaw, and Kilo Code all work with any LLM provider — local models via Ollama, hosted models via OpenRouter, or direct API access to Anthropic, OpenAI, Google, and others. This flexibility is only possible when the agent layer is open and the model layer is a swappable dependency.
What Comes Next
The OpenRouter rankings are a real-time signal of where AI usage is actually going — not where VCs or press releases say it’s going. And the signal is clear:
- The interface is shifting from browser tabs to messaging apps and terminals
- The interaction model is shifting from synchronous prompting to ambient, always-on agents
- The trust model is shifting from “cloud service you use” to “infrastructure you run”
- The business model is shifting from proprietary platforms to open-source with bring-your-own-model
We’re watching the most significant UX transition in AI since ChatGPT launched. The chatbot era gave everyone a taste of what LLMs can do. The agent era is about making them actually do it — persistently, autonomously, and on your terms.
The rankings say the future belongs to the lobster and the research lab. The enterprise wrappers will follow, as they always do. But the users are already here.
Hermes Agent: github.com/nousresearch/hermes-agent OpenClaw: github.com/openclaw/openclaw Kilo Code: github.com/Kilo-Org/kilocode OpenRouter Rankings: openrouter.ai/rankings