The AI agent landscape: who uses which agents, where, why, and with what tools

The biggest moves are new names. Nobody searched for Claude Cowork a year ago; in August, 368k people did. OpenClaw went from 90 to 450k, and Hermes Agent from 30 to 823k.

The characters are the quiet giant. Character.AI and Janitor AI are searched more than any agent outside the big chat apps, and they almost never act on anything. Counting them as agents would make agents look more used than they are.

Tokens measure work: every page an agent reads and every line it writes is counted. OpenRouter is the one router that publishes them by app. Of the 231.3T tokens its top 100 apps sent in the 30 days to 13 October, apps tagged as coding sent 193.9T. An app can carry several tags, so the groups overlap: Hermes Agent, the largest, is tagged both coding and personal agent. ChatGPT, Gemini, Copilot and Cowork run on their makers' own models and never pass through it.

Microsoft read 37.5 million Copilot conversations from January to September 2025 and found two products in one. On the phone, health was the top subject at every hour of every month, and people asked for advice as much as for facts. On the desktop, work and career moved ahead of technology from 8 a.m. to 5 p.m. Late at night, philosophy and religion came up more often.

Google's figure points the same way: 63% of Gemini's users talk to it rather than type. The phone agent is a companion that is always there; the desktop agent is a colleague with access to the files.

In 2025 the big labs shipped agents as separate products: OpenAI's Operator, then agent mode and the Atlas browser; Anthropic's computer use; Google's Project Mariner. In 2026 they pulled them back into one desktop app each.

Anthropic made Cowork, an agent that works in your files and apps, generally available in the Claude desktop app on 9 April, for paying subscribers on macOS and Windows. OpenAI shut the Atlas browser on 9 August and moved agent mode into ChatGPT Work, an agent for documents, spreadsheets and slides inside the ChatGPT desktop app, beside Codex. Searches followed: "ChatGPT agent" fell from 50k a month to 15k, and "Claude Cowork" rose from nothing to 368k.

The browser agent did not go away; it became a feature of the browser people already have. Claude in Chrome rose from 590 searches a month to 135k while the stand-alone AI browsers shrank around it.

The new place of 2026 is your messages. An agent runs all the time on a computer you own or a server someone rents you, and you talk to it in WhatsApp, iMessage, Telegram, Slack or Signal: book this, remind me of that, check the deploy, answer the email. It texts back when it is done.

OpenClaw, started in November 2025 by Peter Steinberger under the name Clawdbot, is open source, connects any model to two dozen chat apps, and is now the most-starred agent on GitHub, at 391,934. Hermes Agent, from Nous Research, does the same from the terminal and sends more tokens through OpenRouter than any other app. Poke is the hosted version for people who will not run a server: you text a number. For now this is the tinkerers' place. For everyone else, the agent in their messages is Meta AI inside WhatsApp.

An agent is only as useful as what it can reach. Since late 2024 the common plug has been the Model Context Protocol (MCP): a tool written once as an MCP server works in Claude, ChatGPT, Gemini, Copilot, Cursor and the open agents alike. ChatGPT's apps are built on it.

Its building blocks are installed at a scale few libraries reach. The TypeScript SDK was downloaded 65M times in the week to 13 October, 6.1× its first week of January, and the Python one 61M times. The official MCP Registry lists 43,275 servers: 27,666 can be reached over the internet and 17,182 are installed and run locally (2,075 offer both). Nearly two in three are hosted.

The most-installed tool is a browser. Playwright's server was downloaded 6.4M times last week, 6.9× its January level, and Chrome DevTools' rose 12×. An agent that writes a web page wants to open it and look, the same way a person checks their work. Second is reading: Context7 hands the agent the current documentation for a library, so it stops writing code against last year's version.

mcp-remote is the bridge a local agent uses to reach a server hosted somewhere else, and its 6.7× rise is a sign of where tools are going: off the laptop and onto the vendor's servers, where Notion, Linear, Stripe and the rest now host their own. Those hosted servers are not on npm, so this chart sees less of them every month.

A landscape like this goes stale the month it is drawn. A hedwigAI workbook re-reads its sources on the schedule you set, keeps each source beside its number, and asks you before it adds a row or changes what a column means.