OpenClaw Change log & Dev Experience
Faster runs. Safer edges.
Three stable releases landed, but the bigger story is shape: OpenClaw got faster and smaller, made its modular parts more public, and showed up in an enterprise-stage conversation at Dell Tech World.
- Release window: May 22, May 26, and May 27 stable releases.
- Main story: the product is becoming both lighter and more legible — a faster core, a public ecosystem surface, steadier channels, and clearer proof paths for operators.
OpenClaw Weekly Dev Experience · slide 1 / 13
The numbers
Velocity stayed high
The week covered 2,616 commits from 198 authors, with three stable releases, 473 stable changelog bullets, 386 fix bullets, and 425 issue references.
- Top commit authors: Peter Steinberger 1,132; Vincent Koc 656; Shakker 124; Ayaan Zaidi 52; Dallin Romney 39; Gio Della-Libera 38; github-actions[bot] 36.
- Commit mix: 1,764 fix, 246 test, 145 refactor, 95 docs, 90 perf, 81 ci, 64 chore, 46 feat.
- Stable releases captured: 2026.5.22, 2026.5.26, and 2026.5.27; stable changelog sections included 473 bullets, 386 fix bullets, and 425 issue references.
OpenClaw Weekly Dev Experience · slide 2 / 13
OpenClaw is now a public ecosystem
not just one assistant
This week’s bigger story is the ecosystem page: OpenClaw is making the modular parts behind the product visible as a public reef of hosted agents, crawlers, SDKs, skills, libraries, and native tools.
- The page reframes OpenClaw as a federation: agents and services, local-first crawlers, agent infrastructure and SDKs, TypeScript libraries, native tools, and the rest of the builder surface.
- Key message for the room: the product is becoming more understandable because the pieces used to build it are becoming public, reusable, and inspectable.
OpenClaw Weekly Dev Experience · slide 3 / 13
Sharper claws, smaller core
performance became a product proof point
Peter’s performance writeup gives the deck a quantified spine: OpenClaw is moving optional weight out of core while making agent turns faster and installs easier to trust.
- Stable cold turn: 9.8s on 2026.4.14 to 3.4s on 2026.5.27 — 2.9× faster.
- Stable warm turn: 7.5s to 3.0s — 2.5× faster.
- Agent peak RSS: 686 MB to 635 MB — 7% lower.
- Published tarball: 43.3 MB on 2026.3.31 to 17.8 MB on 2026.5.27 — 59% smaller.
- Installed dependencies: 371 on the latest release, down 42% from the monthly high; main is already at 314.
- Direction: keep core small, move optional capabilities into plugins, and measure the user-visible effects.
OpenClaw Weekly Dev Experience · slide 4 / 13
Dave Morin made the enterprise case
at Dell Technologies World
The Dell Tech World fireside gave OpenClaw a useful public proof point: agentic building is moving from developer novelty into enterprise workflow, local AI infrastructure, and board-level platform strategy.
- Dave framed OpenClaw as the first AI tool that felt truly personal: a Mac/PC/iPhone-style creation moment, not another developer utility.
- The adoption story is broader than coders: 80% new builders, global ClawCon demand, and use cases the team “would have never dreamed of.”
- Start safely: sandbox the agent, give read-only access first, then feed workflow context through data adapters.
- North Star: support every platform, every model, partnerships, workflow-specific adapters/plugins, and reliability/security/observability enterprise buyers can trust.
Source: Dell Technologies World 2026 / theCUBE · YouTube MV9Zwa5r8Y8 · slide 5 / 13
Gateway work showed up as speed
Less rediscovery, fewer waits
A lot of this week was intentionally boring: fewer repeated scans, fewer rebuilt catalogs, and fewer hidden cleanup paths delaying visible replies.
- Startup and runtime paths reuse plugin metadata, channel/session lookups, auth env snapshots, usage-cost data, warnings, and scheduled-service imports.
- Visible replies separate user-facing sends from slower follow-up work, so chat feels less blocked by cleanup.
- Performance evidence moved into release and CI lanes instead of staying anecdotal.
- Rotated CPU profiles, bounded waits, package inventory checks, and smoke aliases make regressions easier to catch.
OpenClaw Weekly Dev Experience · slide 6 / 13
Transcripts became infrastructure
Meeting notes, replay, media provenance
Transcript work moved from feature copy into shared plumbing that channels, summaries, Codex mirrors, WebChat, and CLI/TUI replay can all use.
- Transcript-backed meeting summaries and source-provider chunks are now part of the core story.
- Cleaned user turns, media provenance, Codex mirrors, WebChat replies, and TUI replay all lean on a more consistent transcript path.
- The practical win: less disagreement between what the user saw, what the agent processed, and what the operator can inspect later.
OpenClaw Weekly Dev Experience · slide 7 / 13
Channels got steadier
Telegram, iMessage, WhatsApp, Slack, Discord
Channel delivery work kept aiming at the same bar: replies should land in the right place, progress should not lie, and late cleanup should not erase successful outcomes.
- Telegram preserves typing/progress context and forum-topic delivery better.
- iMessage handles attachment roots, duplicate local sources, and duplicate native exec approval prompts more carefully.
- WhatsApp group/media behavior and Slack final replies got recovery work.
- Discord voice improved playback and model picking.
- Wake-name handling is more tolerant without turning ambient speech into agent triggers.
- Realtime Talk runs can be inspected, steered, cancelled, or followed up from Web UI and Discord voice.
OpenClaw Weekly Dev Experience · slide 8 / 13
Codex and providers got less brittle
Auth, resume, schema guards, model coverage
The provider story is not just more logos. It is fewer dead ends when auth, schemas, runtime model selection, or long-running Codex runs behave differently by environment.
- Runtime models resolve earlier, shared app-server clients survive startup and spawned-helper failures, and native hook relay generations survive restarts.
- Resume, timeout, usage-limit recovery, dynamic tool-schema guards, and workspace-memory routing reduce stuck runs.
- OpenAI-compatible embedding providers are core.
- DeepInfra catalog browsing, Pixverse video generation, VLLM thinking params, Claude CLI OAuth overlays, xAI usage-limit surfacing, and bare Anthropic model IDs all got attention.
OpenClaw Weekly Dev Experience · slide 9 / 13
Security boundaries tightened
Reject unsafe inputs earlier
The strongest security work this week moved checks ahead of live runs. Bad inputs should fail before they become confusing agent behavior.
- Group prompt text is kept out of the system prompt, and system-event/fetched-file/tool-call text is wrapped or scrubbed more safely.
- Side-effecting command wrappers, unsafe Node runtime env overrides, bad numeric CLI options, no-auth Tailscale exposure, stale device tokens, and non-admin device-role approvals are rejected earlier.
- Browser snapshot reads honor SSRF policy, sender allowlists run before dispatch, and release/install trust checks run earlier.
OpenClaw Weekly Dev Experience · slide 10 / 13
Install and release paths got harder to wedge
Windows, Docker, packages, updates
The install story became less about hero debugging and more about packaging enough proof that routine updates stay routine.
- Root and OpenClaw-owned npm plugin packages ship generated shrinkwraps, package acceptance runs earlier, and docs/assets stay out of runtime tarballs.
- Release postpublish checks, beta smoke checks, shrinkwrap override pins, and package inventory checks make broken publishes harder to miss.
- Alpine, Docker templates, Windows Scheduled Tasks, Windows/macOS proof lanes, macOS runners, Testbox/Crabbox delegation, and trusted runtime fallback roots all got hardening.
- Community upgrade chatter still shows Docker-on-Windows, local models, and Codex OAuth as places where clearer recovery docs pay off.
OpenClaw Weekly Dev Experience · slide 11 / 13
Developer experience in one slide
state clarity is the next leverage point
The DX signal stays useful, but it should not crowd out the ecosystem story. Compress it to the recurring pressures: setup, provider state, model capacity, channels, and review loops.
- Strong beta and PR energy: contributors are testing releases quickly, reviewing changes, and helping each other debug local model limits.
- Recurring friction: setup/auth/model routing, subscription vs provider naming, local context budgets, upgrades, and Telegram/Discord reliability.
- Make selected model, auth route, provider source, rate-limit origin, and context budget visible from one canonical status path.
- Keep release proof close to the operator: what changed, what to test, and where recovery starts when a channel or plugin drifts.
OpenClaw Weekly Dev Experience · slide 12 / 13
OpenClaw got faster, smaller, and more public
now make the modular story obvious
Week 15 was the week the platform story became easier to show: the core got lighter, agent turns got faster, the ecosystem became public, Dave Morin carried the enterprise story at Dell Tech World, and the release train kept tightening trust boundaries.
- The release story: less hot-path drag, fewer core dependency surprises, and more fail-closed behavior before live runs.
- The ecosystem story: ~70 public projects plus a Dell Tech World video signal make the modular architecture legible beyond the assistant UI.
- The next useful bar: make the system explain what it is using, why it chose it, and what proof shows the fix is real.
OpenClaw Weekly Dev Experience · slide 13 / 13