OpenClaw has been getting faster and smaller at the same time. The performance work is visible in agent turns. The dependency work is quieter, but it cuts npm size, install size, audit surface, and native package surprises.
The package grew while OpenClaw gained channels, providers, media, memory, and plugin SDK surface. Then we started moving heavier plugin dependency cones out of core. The full release rows and caveats live in the technical report.
v2026.4.149.8sv2026.5.281.9s
v2026.4.147.5sv2026.5.281.9s
v2026.4.14686 MBv2026.5.28581 MB
2026.3.3143.3 MB2026.5.2817.9 MB
* Root shrinkwrap landed in 5.22; 5.28 cuts the duplicate tree sharply, but a smaller nested dependency tree still remains.
Measured as unique package name/version roots in a fresh npm install with scripts disabled.
Latest release, down 19% from 5.27 in the same fresh-install audit.
Down 53% from the monthly high.
5.28 cuts nested openclaw/node_modules from 656.1 MiB to 259.7 MiB in the local audit.
Still not fully gone.
Down 27% from 5.27 because 5.28 no longer ships the internal dist/plugin-sdk/src/** declaration mirror.
The npm package grew from 82.9 MB unpacked to 182.6 MB unpacked while the surface area expanded.
Bedrock, Slack, OpenShell, Anthropic Vertex, Matrix, and WhatsApp move out of the core dependency path.
npm materialized a large nested tree with every canvas platform package.
17.9 MB published tarball, 300 installed package roots, and a 361.7 MiB fresh install in the local audit.
Compared with 5.27: cold turns down 14.5%, warm turns down 16.0%, install size down 52.8%.
The direction is simple: keep core small, move optional capabilities into plugins, make dependency ownership explicit, and measure the user-visible effects. The 5.27 and 5.28 performance rows are repeat-3 mock-provider workflow runs from release branches; the install row is a fresh local npm install with scripts disabled. Useful for spotting large shifts, not for pretending every millisecond is a benchmark law.
For methodology, caveats, per-release rows, and the shrinkwrap boundary audit, read the technical report.
Growth, here, looks more like molting than adding.