Which AI agent harnesses lead right now?
On one day's OpenRouter token count, Hermes Agent (1.73 trillion) and Claude Code (1.55 trillion) lead, then pi, Kilo Code and Codex. By GitHub stars, Hermes Agent (261k) and OpenCode (225k) lead. OpenCode and Antigravity do not appear in the token list at all, so neither ranking covers all seven. Read 2026-11-06.
Contents
Short answer, from the sources below: there is no single ranking of agent harnesses, and the two public measures found here disagree. On one day of tokens sent through OpenRouter, Hermes Agent and Claude Code lead by a wide margin.[1] By GitHub stars, Hermes Agent and OpenCode lead, then Claude Code, Codex and pi.[2][3][4][5][6] Cursor and Antigravity have no repository among the pages read, so stars say nothing about them. This is a first page in a series: how these harnesses handle memory, tool calls, hooks and heartbeats comes on later pages.
A harness here means the program around a model that reads files, runs commands, calls tools and loops until a task is done, as distinct from the model itself. The seven named on this page are the ones Mackenzie listed; the OpenRouter list shows others that are not covered.
Measure one is tokens on OpenRouter
OpenRouter is a service that routes requests to many models, and it publishes a ranking of the apps that send traffic through it. The "CLI Agents" ranking, on the Tokens tab for the period "Today", read at about 10:30 Mountain time on 2026-11-06, was:[1]
The wider "Coding" ranking, read the same morning, had the same first seven places.[7] Cursor sat at 14 there. Neither list showed OpenCode in the ranks read (1 to 39 and 1 to 17), and neither showed Antigravity.
What this measures: only tokens that went through OpenRouter, on one day. A tool whose users send their requests straight to a model maker would be undercounted, and the page does not say how it identifies an app. It is a snapshot, not a trend.
Measure two is GitHub stars
Stars were read from each project's repository page on 2026-11-06. They show how many people marked a project, not how many use it.
- Hermes Agent, by Nous Research: 261k stars, MIT licence.[2]
- OpenCode (the anomalyco repository): 225k stars, MIT licence.[3]
- Claude Code (anthropics/claude-code): 155k stars. The saved page lists a licence entry but its text was not read.[4]
- Codex CLI, by OpenAI: 133k stars, Apache-2.0 licence.[5]
- pi: 120k stars, MIT licence.[6]
Two OpenCode projects share the name. The older repository, opencode-ai/opencode, is archived with the note that the project continued as Crush, by the original author and the Charm team, and it has 13.9k stars.[8] Crush is at rank 22 in the OpenRouter CLI list.[1] The OpenCode site and the 225k-star repository belong to a different project.[9]
What each one says it is
These are each project's own descriptions, so they are claims, not tests.
- Claude Code. Anthropic's tool that reads a codebase, edits files, runs commands and connects to development tools. Its docs list the terminal, VS Code, the desktop app, the web and JetBrains as places to use it, and describe cloud scheduled tasks that keep running when the computer is off.[10]
- Codex CLI. A coding agent from OpenAI that runs locally on the computer. The repository is 97.5 percent Rust.[5]
- Cursor. An editor whose Agent can complete tasks on its own, run terminal commands and edit code. Its docs say there is no limit on how many tool calls Agent can make in one task.[11]
- Antigravity. Google's agent development platform. The launch post calls it an IDE moving toward an agent-first form, with browser control and asynchronous feedback, and it says Antigravity treats learning as a core primitive, with agent actions both reading from and adding to a knowledge base. The launch was a free public preview with rate limits on Gemini 3 Pro.[12]
- Hermes Agent. Built by Nous Research and described as self-improving: it creates skills from experience, searches its past conversations and can be reached through Telegram, Discord, Slack and other messaging apps while it runs elsewhere.[2]
- pi. Described by its repository as a minimal, extensible agent harness. It skips sub-agents and plan mode by default and expects the user to add them as packages.[6] Its author's post lists a minimal system prompt, a minimal toolset, no built-in to-dos, no MCP support and no background bash.[13] The repository names OpenClaw as a real-world integration.[6]
- OpenCode. An open source coding agent used from a desktop app, a phone, the terminal or an SDK, with a choice of 200 or more model providers. Its site says "Trusted by 22M+ developers", a figure from the vendor that was not checked.[9]
What the two measures rule out
These are consequences of the numbers, not findings from a source.
- "Top harness" depends on the ruler. Hermes Agent leads on both measures; Claude Code is second on tokens and third on stars; OpenCode is second on stars and missing from the token lists.
- A list built from stars alone would put OpenCode above Claude Code. A list built from OpenRouter tokens alone would leave OpenCode and Antigravity out.
- Four of the top seven token places (pi, omp, Kilo Code, OpenClaw) are tools not on Mackenzie's list. A survey limited to the seven named would miss them.
Terms
- Harness: the program around a model that gives it tools, files and a loop.
- Token: the unit of text a model reads or writes; usage is often counted in tokens.
- Agent memory, tool call and hook: the three terms the later pages in this series compare.
Not yet proven
- Real usage. No source read gives paid-seat counts or active users for Claude Code, Codex, Cursor or Antigravity. OpenRouter covers only its own traffic.
- Why OpenCode is missing from the token lists. The page does not say. It may route requests elsewhere, but that is a guess.
- A trend. Both measures are one day's reading. A weekly or monthly view was not read.
- Cursor and Antigravity internals. Only a docs page and a launch post were read for each; neither gave a licence or a repository.
- Other lists. Search results included several "best AI coding agents of 2026" articles from vendors and blogs. None was used.
- omp, Kilo Code and OpenClaw. They rank high on tokens but were not described here.
If you pick a harness for real work and the ranking here led you wrong, I would like to hear it: mackenzie@mackenziebowes.com.
Sources
- OpenRouter, "CLI Agents Rankings", Tokens tab, period Today. Read in Rotunda 2026-11-06. The page text had no figures, so they were typed from a screenshot of the page. openrouter.
ai/ apps/ category/ coding/ cli- agent - GitHub, NousResearch/hermes-agent. Read in Rotunda 2026-11-06. github.
com/ nousresearch/ hermes- agent - GitHub, anomalyco/opencode. Read in Rotunda 2026-11-06. github.
com/ anomalyco/ opencode - GitHub, anthropics/claude-code. Read in Rotunda 2026-11-06. github.
com/ anthropics/ claude- code - GitHub, openai/codex. Read in Rotunda 2026-11-06. github.
com/ openai/ codex - GitHub, earendil-works/pi. Read in Rotunda 2026-11-06. github.
com/ earendil- works/ pi - OpenRouter, "Coding" apps ranking, Tokens tab, period Today. Read in Rotunda 2026-11-06; first screen only, typed from a screenshot. openrouter.
ai/ apps/ category/ coding - GitHub, opencode-ai/opencode (archived). Read in Rotunda 2026-11-06. github.
com/ opencode- ai/ opencode - OpenCode, home page. Read in Rotunda 2026-11-06. opencode.
ai/ - Claude Code Docs, Overview. Read in Rotunda 2026-11-06. code.
claude. com/ docs/ en/ overview - Cursor Docs, Agent overview. Read in Rotunda 2026-11-06. cursor.
com/ docs/ agent/ overview - Google Antigravity, "Introducing Google Antigravity". Read in Rotunda 2026-11-06. antigravity.
google/ blog/ introducing- google- antigravity/ - Mario Zechner, "What I learned building an opinionated and minimal coding agent", 2025-11-30. Read in Rotunda 2026-11-06; its contents list was read, not the full post. mariozechner.
at/ posts/ 2025- 11- 30- pi- coding- agent/
Terms on this page
- Agent memory: Information an AI agent writes down in one session and reads back in a later one.
- Hook: A command that an AI coding agent runs automatically at a set point in its work.
- Model Context Protocol: An open standard for connecting AI applications to outside tools and data.
- Tool call: A request from an AI model to run one action, such as a search or a file write, which the surrounding program carries out.