AI coding agents, by actual usage

Short answer: nobody can honestly rank these on quality, and this page does not try. What it does instead is report a number almost nobody publishes: how many tokens each agent actually pushed through OpenRouter on 2026-07-31. That produces an order no vote would have produced — Hermes Agent first at 970B tokens, more than twice Claude Code's 422B, and Kilo Code third, ahead of OpenClaw.

The limitation belongs here rather than at the bottom, because it decides how you should read everything below. OpenRouter counts only apps that opted into attribution, and only traffic that went through OpenRouter. Agents people mostly run on a vendor subscription or a direct API key — Claude Code and Codex above all — are undercounted by construction. This is a ranking of what gets routed through one broker. It is not a ranking of what gets used.

The ranking, 31 July 2026

Tokens routed through OpenRouter's coding category over the preceding 24 hours, as displayed that day.

The tail on the same day: GitLawb 5.03B, Letaido 4.94B, mimocode 2.97B, goose 2.63B, Agent Zero 2.42B, OpenSquilla 2.37B, Qwen Code 2.24B, Crush 1.71B. The dashes above are entries we have not verified well enough to describe, so we are not describing them. All of it together comes to roughly 2.2 trillion tokens in a day, and Hermes Agent alone is about 44% of that.

What this measures, and what it doesn't

Two constraints, both structural, and both large enough to change the order.

First: appearing here is a choice a maintainer made. OpenRouter's attribution documentation is explicit about it:

This header is required for app attribution — without it, no app page will be created and your usage will not appear in rankings.

Only public apps, meaning those that send headers, are included in rankings

You can watch that happen. The agent sitting fifth on this list, pi, is there because of a GitHub issue opened on 2026-04-19, titled "OpenRouter attention (Pi featured in rankings)," whose entire request was:

Add X-OpenRouter-Title header to OpenRouter requests, so that Pi can be featured in the OpenRouter rankings.

The issue is closed and the header shipped. Before April, pi's traffic existed and the leaderboard could not see it. That is the shape of every absence in this table.

Second: only OpenRouter traffic is counted at all. Claude Code is bundled with a Claude plan and talks to Anthropic by default; Codex CLI comes with a ChatGPT plan. The economics of both push almost all real usage somewhere this dashboard cannot see, so 422B and 19.3B are leftovers, not totals. Cursor is worse: its agent and Tab run on Cursor-hosted models and its bring-your-own-key support covers chat only, so 14.3B says close to nothing about Cursor's size.

What's surprising about the order

Three things, once you have accepted the caveats.

Hermes Agent's lead is enormous, and partly explained by the caveat itself. Nous Research's agent is deliberately model-agnostic and its default provider is OpenRouter, so it is structurally over-represented here for exactly the reason Claude Code is under-represented. That does not make 970B fake — a token is a token, and no other opted-in app is close. It does mean the honest comparison between those two is about shape rather than size. Our setup notes cover what it does to your machine, including three widely repeated claims about it that are wrong.

Kilo Code is the entry nobody predicts. Third at 241B, and on all-time cumulative tokens it is narrowly ahead of Claude Code — 7.2T against 7.09T, with Hermes at 28.6T and OpenClaw at 4.63T. Kilo Code is an MIT-licensed VS Code and JetBrains extension that picked up much of the Roo Code userbase when that project stopped. Almost no roundup mentions it.

GitHub stars do not predict any of this. Measured the same day: OpenClaw has 384,636 stars and sits fourth at 153B. Kilo Code has 26,639 and sits third at 241B. OpenHands has 82,633 and routes 15B. OpenCode has 191,313 and does not appear at all. Stars measure who bookmarked a repository; this measures who is paying for inference through one particular door. Different populations — and the first is what almost every "best agent" list quietly uses as its evidence.

One more: Roo Code was archived on 2026-05-15 and its repository has had no push since. Eleven weeks later it still routes 6.38B tokens a day, ahead of goose, Qwen Code and Crush. Software does not stop when its maintainers do.

Three different shapes in one list

The table flattens three genuinely different kinds of product into one column, which is the main thing wrong with reading it as a ranking. If you are trying to locate yourself, the split matters more than the order.

Terminal agents — Hermes Agent, Claude Code, pi, Codex, Oh-My-Pi, goose, Qwen Code, Crush. You describe an outcome, the process reads files and runs commands, you check the result. What decides whether you can safely walk away is not model quality but whether a deterministic policy layer sits between the agent and your filesystem; Claude Code's hooks are the most developed version of that, with one sharp edge that silently defeats most first attempts. OpenHands is the same shape with a harder boundary, running the work inside a Docker sandbox. And if you want this shape without a single vendor underneath it, OpenCode is the usual comparison — an agent nowhere in this table and not small.

Editor extensions and editors — Kilo Code, Cline, Roo Code, Cursor, Zed. The review surface is where you already work, and diffs land in front of you rather than behind you. Right container if you want to see each edit as it happens; wrong one if you keep wishing the assistant would go away for ten minutes. Cline against Claude Code is mostly temperament rather than quality, and Zed against Cursor comes down to whether you depend on a VS Code extension.

Agents you reach through a messaging app — OpenClaw is the clearest case, a local daemon fronted by Telegram, Discord, Slack and others; Hermes Agent has a gateway mode that does the same job. The distinguishing property is not capability. It is that the agent is not tied to the laptop you happen to be sitting at, which is why these get used at hours when nobody opens an editor. Our OpenClaw setup guide covers the first-hour failures and the security picture, which is worth reading before you expose anything.

The useful question is not which row is highest. It is whether you want to watch, to delegate, or to be reachable. Those three answers point at three different columns of the table, and a benchmark cannot arbitrate between them.

What's growing fastest

Week-over-week change on the same date, for the biggest movers: Sekai above 999%, Codebuff +184%, Portkey AI +154%, Hello Minds and Ethoswarm +131%, Codex +85%, Sahasra +73%, Nous Research API +66%, pi +34%.

Growth percentages are the least trustworthy figures on any dashboard, and most of these are arithmetic rather than news — "above 999%" means the base was near zero last week. Two are worth a second look because the base underneath is not small.

Codex at +85% is the only agent from a model lab in the growth list, and it is growing on a number already undercounted for the reason above. Whether that changes the Codex versus Claude Code decision is doubtful — that one is still mostly settled by which subscription you already pay for.

pi at +34% on a 125B base is the more interesting one, because it has only been visible here since April. A newcomer with 81,051 stars being fifth by usage is what this dataset is good at surfacing and opinion lists are bad at.

What this data can't tell you

Being clear about this is the reason the rest of the page is worth anything.

• It cannot tell you which agent is better. Token count reflects verbosity, context window size, retry behaviour and task length at least as much as it reflects how many people use something. An agent that re-reads the whole repository every turn will out-rank a frugal one at identical popularity. There is no capability grid anywhere on this page because we have no honest basis for one.
• It cannot tell you how many people use each one. One team running nightly batch jobs can outweigh thousands of humans typing.
• It cannot tell you whether anyone stayed. There is no retention, satisfaction or task-success signal in a token count.
• Absence means nothing. Aider, Amp, Copilot, Devin, Jules, Kiro and Replit are not in this table. Neither is OpenCode. That is a statement about attribution and default providers, not about usage.
• It is one day. Daily figures move. The all-time column moves slowly, which is why we quoted it too — and why the Kilo Code result is the sturdiest thing here.

If you want these products assessed individually rather than counted, that is what the rest of our guides are for. Each one is dated, and several conclude that the right answer is something we do not make.

Sources

Token figures were read off OpenRouter's public coding-agent rankings on 2026-07-31 and are 24-hour counts unless labelled all-time. GitHub star counts, licences and archive status were queried from the GitHub API the same day. Everything here decays; check the date before relying on it.

• OpenRouter — Coding Agents rankings · openrouter.ai/rankings
• OpenRouter docs — App Attribution, source of both quoted sentences about opt-in tracking
• earendil-works/pi issue #3414 — "OpenRouter attention (Pi featured in rankings)" · pi on OpenRouter
• NousResearch/hermes-agent · Kilo-Org/kilocode · openclaw/openclaw · cline/cline
• RooCodeInc/Roo-Code (archived) · OpenHands/OpenHands · anomalyco/opencode · can1357/oh-my-pi
• Claude Code docs · Codex CLI docs · cursor.com · zed.dev

Not affiliated with OpenRouter, Nous Research, Anthropic, OpenAI or any product listed. If something here is wrong or has gone stale, tell us and we'll fix it.