Opt-in cohort
Aggregated first-party observations from public participants, not a representative survey of all developers.
What the data says
Codex CLI appears in the largest number of participating developers' records (140 of 199), while Claude Code accounts for the largest share of observed tokens (63%). Those measures answer different questions and should not be read as industry-wide market share.
Participating developers
199
Listed cohort members with usage in the 30-day window
Reach leader
Codex CLI
70% of the observed cohort
Token leader
Claude Code
63% of measured tokens
Token leader's top model
claude-opus-5-5
Heaviest model observed through that agent
Developer reach counts how many listed participants used an agent at least once. Token share measures how much of the cohort's observed work flowed through it. A tool can rank highly on one measure and lower on the other.
Codex CLI
267.62B observed tokens
Claude Code
580.03B observed tokens
OpenCode
23.11B observed tokens
Cursor
33.42B observed tokens
Hermes
9.71B observed tokens
Pi
3.70B observed tokens
Antigravity
594.6M observed tokens
Openclaw
201.3M observed tokens
Claude Code
135 participating developers
Codex CLI
140 participating developers
Cursor
46 participating developers
OpenCode
56 participating developers
The table keeps the two rankings side by side and adds the leading model observed through each agent. Participant percentages can sum to more than 100% because one developer may use several agents.
A developer is counted once for each agent they used during the window. The measure shows observed reach, not satisfaction, retention, paid seats, or a developer's preferred tool.
Token totals vary with task size, context length, model behavior, and logging coverage. They describe intensity inside this cohort, not revenue, installations, or adoption across the wider industry.
Window and coverage. The page requests the trailing 30-day public insights snapshot. The sample includes listed participants with at least one eligible usage row in that period. The displayed generation date marks the API snapshot used to render the page.
Developer reach. For each agent, distinct participating developers are divided by all listed developers with eligible usage in the window. People who use multiple agents appear in multiple rows, so reach percentages are not additive.
Token intensity and top model. Token share is the agent's portion of all eligible tokens in the snapshot. Top model means the model with the most tokens attributed to that agent; it does not establish user preference or model quality.
Limitations. Participation is opt-in and the cohort is not weighted to represent all software developers. Agent integrations differ in what their local logs expose, and some records cannot attribute usage to a specific model. Treat the results as a transparent cohort observation—not an industry census, popularity poll, or market-share estimate.
Aggregated model, agent, trend, skill, and tool-call totals from participating developers.
Explains what the CLI aggregates, what it does not collect, and the limits of the public cohort.
Hermes
18 participating developers
Pi
11 participating developers
Antigravity
2 participating developers
Openclaw
1 participating developers
| default |
| OpenCode | 56 | 28% | 2.5% | 23.11B | muse-spark-1.3-contributor-free |
| Hermes | 18 | 9.0% | 1.1% | 9.71B | deepseek-v4.1-flash |
| Pi | 11 | 5.5% | 0.4% | 3.70B | model-24d682bd5981 |
| Antigravity | 2 | 1.0% | <0.1% | 594.6M | gemini-3.8-flash |
| Openclaw | 1 | 0.5% | <0.1% | 201.3M | model-95d1f9678305 |
| Openai Api | 1 | 0.5% | <0.1% | 193.7M | gpt-realtime-1.5 |
| Kimi CLI | 1 | 0.5% | <0.1% | 90.3M | k3-256k |
| Deepseek | 1 | 0.5% | <0.1% | 32.5M | deepseek-flash |
| Autoclaw | 1 | 0.5% | <0.1% | 17.7M | tdpsk_deepseek-v4-flash-202605 |
| Factory Droid | 1 | 0.5% | <0.1% | 8.2M | claude-opus-4-6 |
| Goose | 1 | 0.5% | <0.1% | 29.3K | glm-5.3-flash |
| Kilo Code | 1 | 0.5% | <0.1% | 18.5K | free |
| Continue | 1 | 0.5% | <0.1% | 4.0K | qwen2.5-coder:7b |