LLM router adoption signals
Compare 28-day package downloads and growth. Downloads do not measure active customers or hosted traffic.
9 packages tracked · 2 unmeasured routers · Methodology
9 packages · Ranked by downloads
Mixed package types; compare with care| Rank | Router / package | Downloads | Change | Period & source |
|---|---|---|---|---|
| 1 | @ai-sdk/gateway ↗npm · AI SDK gateway provider | 94,295,187 | +4.8% | 2026-09-29 – 2026-10-26 ↗Retrieved 2026-10-27Previous period89,967,457 downloads2026-09-01 – 2026-09-28 |
| 2 | litellm ↗pypi · Python SDK and gateway | 88,946,984 | +5.8% | 2026-09-29 – 2026-10-26 ↗Retrieved 2026-10-27Previous period84,059,002 downloads2026-09-01 – 2026-09-28 |
| 3 | @openrouter/sdk ↗npm · Client SDK | 2,966,398 | -1.6% | 2026-09-29 – 2026-10-26 ↗Retrieved 2026-10-27Previous period3,014,755 downloads2026-09-01 – 2026-09-28 |
| 4 | portkey-ai ↗npm · Client SDK | 745,816 | +7.1% | 2026-09-29 – 2026-10-26 ↗Retrieved 2026-10-27Previous period696,586 downloads2026-09-01 – 2026-09-28 |
| 5 | tensorzero ↗pypi · Python client and embedded gateway | 24,128 | -27.8% | 2026-09-29 – 2026-10-26 ↗Retrieved 2026-10-27Previous period33,432 downloads2026-09-01 – 2026-09-28 |
| 6 | merge-gateway-ai-sdk-provider ↗npm · AI SDK gateway provider | 6,861 | -39.7% | 2026-09-29 – 2026-10-26 ↗Retrieved 2026-10-27Previous period11,384 downloads2026-09-01 – 2026-09-28 |
| 7 | routellm ↗pypi · Model-selection framework | 5,706 | -15.3% | 2026-09-29 – 2026-10-26 ↗Retrieved 2026-10-27Previous period6,740 downloads2026-09-01 – 2026-09-28 |
| 8 | @maximhq/bifrost ↗npm · Gateway launcher | 3,058 | -61.4% | 2026-09-29 – 2026-10-26 ↗Retrieved 2026-10-27Previous period7,929 downloads2026-09-01 – 2026-09-28 |
| 9 | @helicone/ai-gateway ↗npm · Gateway launcher | 129 | +65.4% | 2026-09-29 – 2026-10-26 ↗Retrieved 2026-10-27Previous period78 downloads2026-09-01 – 2026-09-28 |
What the ranking tells you
We track one publisher-documented package per measured router. Each count covers 28 complete days and compares with the preceding 28 days. npm reports package downloads; PyPI Stats excludes mirrors. Check the dates in each row before comparing snapshots.
Repeat installs, CI jobs and dependency installs count. In particular, Vercel’s gateway provider can be installed as a dependency without active Gateway usage. Packages serving multiple purposes, such as LiteLLM, do not isolate router usage. We do not sum ecosystems or convert downloads into users.
Counts are saved snapshots refreshed manually, not live telemetry. Ranking is descending within the selected filters; growth is percentage change, not a change in market share. A zero previous count has no percentage baseline. Publisher maintenance and runtime quality are separate checks; TensorZero’s repository is archived.
Not measured
Cloudflare AI Gateway: The reviewed integration uses a shared OpenAI client. Its downloads cannot be attributed to Cloudflare AI Gateway. A gateway-specific package has not been established in this review. Source ↗
Requesty: The reviewed quickstart uses shared OpenAI and Anthropic clients. Their downloads cannot be attributed to Requesty. A Requesty-specific package has not been established in this review. Source ↗
Beyond package downloads
Paid adoption and token spending answer different questions. The Ramp AI Index ↗ measures observed AI payments among businesses on Ramp, with token metrics from a separate product cohort. Its public index does not establish router-level market share and is not included in these ranks.