Community usage
Which AI coding agents and models developers actually use
Every agent launched from AgentsRoom by users who share their usage, counted by CLI and by model. No survey, no vendor benchmark: what people start when they have work to do.
Agent launches
81
newvs previous period
Users
5
CLIs used
4
Models used
12
Most launched coding CLIs
Share of all agent launches over the period, with the users who launched each one at least once.
- 1
Claude70.4%57 launches5 usersnew - 2
Codex24.7%20 launches2 usersnew - 3
Antigravity3.7%3 launches2 usersnew - 4
Grok1.2%1 launches1 usersnew
Most launched models
The model pinned on the agent at launch. Launches that leave the model to the CLI's default are counted in the CLI totals only.
- 1
OpusClaude45.7%37 launches3 users64.9% of Claude launchesnew - 2
GPT-5.6 LunaCodex17.3%14 launches2 users70% of Codex launchesnew - 3
Opus 5Claude11.1%9 launches3 users15.8% of Claude launchesnew - 4
SonnetClaude7.4%6 launches4 users10.5% of Claude launchesnew - 5
GPT-5.6 SolCodex4.9%4 launches1 users20% of Codex launchesnew - 6
defaultClaude2.5%2 launches2 users3.5% of Claude launchesnew - 7
Gemini 3.8 Flash (High)Antigravity2.5%2 launches1 users66.7% of Antigravity launchesnew - 8
GPT-6 AstraCodex2.5%2 launches1 users10% of Codex launchesnew - 9
Opus 4.8Claude2.5%2 launches1 users3.5% of Claude launchesnew - 10
Claude Sonnet 4.6Antigravity1.2%1 launches1 users33.3% of Antigravity launchesnew - 11
Grok 4.6Grok1.2%1 launches1 users100% of Grok launchesnew - 12
HaikuClaude1.2%1 launches1 users1.8% of Claude launchesnew
How agents are started
- New agent30%
- Backlog ticket30%
- Quick agent12%
- Trigger12%
- Team run10%
- Resumed5%
Reasoning effort
- Medium40%
- High21%
- Xhigh21%
- Low18%
Agent roles
- Custom37%
- Fullstack16%
- Frontend15%
- Architect6%
- Qa6%
- Bot5%
Where the numbers come from
AgentsRoom is a desktop command center for AI coding agents. Every time a user starts an agent, the desktop records the CLI, the model, the reasoning effort, the agent role and how it was started, and nothing else: no prompt, no project, no file, no name. Only users who keep "Share usage data" on contribute, anonymous launches count as launches but never as users, and the page refreshes every ten minutes.
The same numbers rank the Providers and Models boards of the in-app leaderboard.
Download AgentsRoom