ModelCompare
Compare

AI coding agent usage

What was the most used AI coding agent on ?

Codex CLI from OpenAI, with a usage score of 47 out of 100 across 3 of 3 sources. Claude Code was next at 44.

Usage score over time, week by week

Usage ranking

AI coding agents by usage on July 21, 2026

This ranks how much each agent is used and followed, not how good it is: a popular agent can be a worse fit for your work than a small one. Click an agent to see where its numbers come from. Each source's column shows the agent's scaled score (the leader is 100) over the raw number. 3 sources had data on this day. The score averages an agent's scaled values over the sources that can measure it (at least three).

  1. 1
    Codex CLI
    OpenAI · open source
    Usage score 47
    • npm 100 · 16M
    • GitHub 25 · 95k
    • VS Code 15 · 11M
    • npm: 10016,023,958 downloads of @openai/codex in the 7 days to Jul 21, 2026. That's the most of any agent in this source, so it scores 100. Source
    • GitHub: 2594,956 stars on github.com/openai/codex on Jul 2, 2026. The leader here is OpenClaw with 382k, so Codex CLI scores 95k ÷ 382k × 100 = 25. Source
    • VS Code: 1511,321,010 installs of the openai.chatgpt extension on Jul 6, 2026. The leader here is GitHub Copilot with 76M, so Codex CLI scores 11M ÷ 76M × 100 = 15. Source

    Usage score: (100 + 25 + 15) ÷ 3 = 47. Codex CLI website

  2. 2
    Claude Code
    Anthropic · closed source
    Usage score 44
    • npm 72 · 12M
    • GitHub 36 · 136k
    • VS Code 25 · 19M
    • npm: 7211,546,513 downloads of @anthropic-ai/claude-code in the 7 days to Jul 21, 2026. The leader here is Codex CLI with 16M, so Claude Code scores 12M ÷ 16M × 100 = 72. Source
    • GitHub: 36135,577 stars on github.com/anthropics/claude-code on Jul 3, 2026. The leader here is OpenClaw with 382k, so Claude Code scores 136k ÷ 382k × 100 = 36. Source
    • VS Code: 2519,278,058 installs of the anthropic.claude-code extension on Jul 1, 2026. The leader here is GitHub Copilot with 76M, so Claude Code scores 19M ÷ 76M × 100 = 25. Source

    Usage score: (72 + 36 + 25) ÷ 3 = 44. Claude Code website

  3. 3
    OpenClaw
    OpenClaw · open source
    Usage score 39
    • npm 16 · 2.6M
    • GitHub 100 · 382k
    • npm: 162,627,215 downloads of openclaw in the 7 days to Jul 21, 2026. The leader here is Codex CLI with 16M, so OpenClaw scores 2.6M ÷ 16M × 100 = 16. Source
    • GitHub: 100381,823 stars on github.com/openclaw/openclaw on Jul 5, 2026. That's the most of any agent in this source, so it scores 100. Source

    Not counted: VS Code, because it has no VS Code extension.

    Usage score: (16 + 100) ÷ 3 = 39, divided by 3 rather than 2 because every agent's average counts at least 3 sources. OpenClaw website

  4. 4
    GitHub Copilot
    GitHub · closed source
    Usage score 36
    • npm 9 · 1.5M
    • VS Code 100 · 76M
    • npm: 91,482,958 downloads of @github/copilot in the 7 days to Jul 21, 2026. The leader here is Codex CLI with 16M, so GitHub Copilot scores 1.5M ÷ 16M × 100 = 9. Source
    • VS Code: 10076,122,820 installs of the GitHub.copilot-chat extension on Jul 1, 2026. That's the most of any agent in this source, so it scores 100. Source

    Not counted: GitHub, because it has no public GitHub repo.

    Usage score: (9 + 100) ÷ 3 = 36, divided by 3 rather than 2 because every agent's average counts at least 3 sources. GitHub Copilot website

  5. 5
    OpenCode
    Anomaly · open source
    Usage score 20
    • npm 13 · 2M
    • GitHub 47 · 181k
    • VS Code 1 · 849k
    • npm: 132,037,983 downloads of opencode-ai in the 7 days to Jul 21, 2026. The leader here is Codex CLI with 16M, so OpenCode scores 2M ÷ 16M × 100 = 13. Source
    • GitHub: 47181,008 stars on github.com/anomalyco/opencode on Jul 1, 2026. The leader here is OpenClaw with 382k, so OpenCode scores 181k ÷ 382k × 100 = 47. Source
    • VS Code: 1849,452 installs of the sst-dev.opencode extension on Jul 14, 2026. The leader here is GitHub Copilot with 76M, so OpenCode scores 849k ÷ 76M × 100 = 1. Source

    Usage score: (13 + 47 + 1) ÷ 3 = 20. OpenCode website

  6. 6
    Gemini CLI
    Google · open source
    Usage score 10
    • npm 3 · 494k
    • GitHub 28 · 106k
    • npm: 3494,396 downloads of @google/gemini-cli in the 7 days to Jul 21, 2026. The leader here is Codex CLI with 16M, so Gemini CLI scores 494k ÷ 16M × 100 = 3. Source
    • GitHub: 28105,784 stars on github.com/google-gemini/gemini-cli on Jul 6, 2026. The leader here is OpenClaw with 382k, so Gemini CLI scores 106k ÷ 382k × 100 = 28. Source

    Not counted: VS Code, because there's no snapshot of its extension page from around that day.

    Usage score: (3 + 28) ÷ 3 = 10, divided by 3 rather than 2 because every agent's average counts at least 3 sources. Gemini CLI website

  7. 7
    Cline
    Cline · open source
    Usage score 8
    • npm 1 · 99k
    • GitHub 17 · 64k
    • VS Code 6 · 4.5M
    • npm: 199,014 downloads of cline in the 7 days to Jul 21, 2026. The leader here is Codex CLI with 16M, so Cline scores 99k ÷ 16M × 100 = 1. Source
    • GitHub: 1764,144 stars on github.com/cline/cline on Jul 1, 2026. The leader here is OpenClaw with 382k, so Cline scores 64k ÷ 382k × 100 = 17. Source
    • VS Code: 64,538,449 installs of the saoudrizwan.claude-dev extension on Jul 3, 2026. The leader here is GitHub Copilot with 76M, so Cline scores 4.5M ÷ 76M × 100 = 6. Source

    Usage score: (1 + 17 + 6) ÷ 3 = 8. Cline website

  8. 8
    Kilo Code
    Kilo · open source
    Usage score 3
    • npm 1 · 96k
    • GitHub 7 · 26k
    • VS Code 2 · 1.3M
    • npm: 196,446 downloads of @kilocode/cli in the 7 days to Jul 21, 2026. The leader here is Codex CLI with 16M, so Kilo Code scores 96k ÷ 16M × 100 = 1. Source
    • GitHub: 725,507 stars on github.com/Kilo-Org/kilocode on Jul 3, 2026. The leader here is OpenClaw with 382k, so Kilo Code scores 26k ÷ 382k × 100 = 7. Source
    • VS Code: 21,336,331 installs of the kilocode.Kilo-Code extension on Jul 21, 2026. The leader here is GitHub Copilot with 76M, so Kilo Code scores 1.3M ÷ 76M × 100 = 2. Source

    Usage score: (1 + 7 + 2) ÷ 3 = 3. Kilo Code website

  9. 9
    Roo Code
    Roo Code · open source
    Usage score 3
    • GitHub 6 · 24k
    • VS Code 2 · 1.8M
    • GitHub: 624,303 stars on github.com/RooCodeInc/Roo-Code on Jul 4, 2026. The leader here is OpenClaw with 382k, so Roo Code scores 24k ÷ 382k × 100 = 6. Source
    • VS Code: 21,842,390 installs of the RooVeterinaryInc.roo-cline extension on Jul 18, 2026. The leader here is GitHub Copilot with 76M, so Roo Code scores 1.8M ÷ 76M × 100 = 2. Source

    Not counted: npm, because it has no command-line package on npm.

    Usage score: (6 + 2) ÷ 3 = 3, divided by 3 rather than 2 because every agent's average counts at least 3 sources. Roo Code website

  10. 10
    Qwen Code
    Alibaba · open source
    Usage score 2
    • npm 1 · 90k
    • GitHub 7 · 26k
    • npm: 189,895 downloads of @qwen-code/qwen-code in the 7 days to Jul 21, 2026. The leader here is Codex CLI with 16M, so Qwen Code scores 90k ÷ 16M × 100 = 1. Source
    • GitHub: 725,693 stars on github.com/QwenLM/qwen-code on Jul 1, 2026. The leader here is OpenClaw with 382k, so Qwen Code scores 26k ÷ 382k × 100 = 7. Source

    Not counted: VS Code, because it has no VS Code extension.

    Usage score: (1 + 7) ÷ 3 = 2, divided by 3 rather than 2 because every agent's average counts at least 3 sources. Qwen Code website

  11. 11
    Crush
    Charm · open source
    Usage score 2
    • npm 0 · 3.9k
    • GitHub 7 · 26k
    • npm: 03,877 downloads of @charmland/crush in the 7 days to Jul 21, 2026. The leader here is Codex CLI with 16M, so Crush scores 3.9k ÷ 16M × 100 = 0. Source
    • GitHub: 725,963 stars on github.com/charmbracelet/crush on Jul 2, 2026. The leader here is OpenClaw with 382k, so Crush scores 26k ÷ 382k × 100 = 7. Source

    Not counted: VS Code, because it has no VS Code extension.

    Usage score: (0 + 7) ÷ 3 = 2, divided by 3 rather than 2 because every agent's average counts at least 3 sources. Crush website

  12. 12
    Letta Code
    Letta · open source
    Usage score 0
    • npm 0 · 10k
    • GitHub 1 · 2.6k
    • npm: 010,353 downloads of @letta-ai/letta-code in the 7 days to Jul 21, 2026. The leader here is Codex CLI with 16M, so Letta Code scores 10k ÷ 16M × 100 = 0. Source
    • GitHub: 12,553 stars on github.com/letta-ai/letta-code on May 24, 2026. The leader here is OpenClaw with 382k, so Letta Code scores 2.6k ÷ 382k × 100 = 1. Source

    Not counted: VS Code, because it has no VS Code extension.

    Usage score: (0 + 1) ÷ 3 = 0, divided by 3 rather than 2 because every agent's average counts at least 3 sources. Letta Code website

Not ranked that day, only one source has them: Hermes Agent (GitHub, 207k), OpenHands (GitHub, 79k), Aider (GitHub, 47k), Auggie (npm, 38k), Amp (npm, 18k), pi (npm, 7.6k), Droid (npm, 1.6k).

Method

How the usage score is made

No single public number measures how much each AI coding agent is used, so we blend the ones that exist. In each source the leader scores 100 and the rest are scaled to it. An agent's score is the average over the sources that can measure it. Sources that can't see an agent at all (Cursor has no npm package) are left out of its average, but the average always counts at least three sources, so an agent seen by only one or two can't outrank one that leads everywhere. An agent needs two sources to be ranked.

  • npmCLI downloads, 7 days · data from Jul 21, 2026

    Downloads of each agent's command-line package in the 7 days to the date, with one-off bursts (mirrors, bots) capped. Downloads count installs and CI runs, not people, and agents with their own installer are undercounted. Source

  • GitHubStars · data from Jul 6, 2026

    Stars on the agent's GitHub repo, from monthly snapshots. Stars measure interest, and build up over time. Source

  • VS CodeExtension installs, all time · data from Jul 21, 2026

    All-time installs of the agent's VS Code extension, from monthly snapshots. Cursor and Windsurf install from Open VSX instead, which isn't counted. Source

Browse

The most used AI coding agent, month by month

2026
2025

Pick the model behind your agent

Most of these agents let you choose the model. Send one prompt to up to six models and see which answer you'd actually use, with what each one cost.

Compare models

Data through October 21, 2026 for 20 agents. npm downloads from the npm registry; OpenRouter app rankings from OpenRouter (CC BY 4.0); GitHub stars and VS Code installs from monthly snapshots in the Internet Archive and the live APIs; survey figures as published by Stack Overflow and JetBrains.