OpenClaw 2.0 (2026.8.1): The Release That Touches Every Layer of the Stack

OpenClaw 2026.8.1 shipped today. The release notes call out “a rebuilt web experience, simpler onboarding, stronger memory and session continuity, and a very large reliability pass” — that’s a polite way of saying the team took a hammer to about nine different subsystems at once. I’ve been running 2026.8.0 against a few daily-blog automations and a Telegram channel bridge for the last two months, so I sat down with the changelog and pulled out the things that actually move the needle for a self-hosted install.

This isn’t a “10 highlights” roundup. It’s a walk through the changes that change behavior — what breaks, what gets faster, what quietly adds a feature you’d been writing workarounds for.

The headline: credential handling moves off the chat surface

The single most consequential change is the masked-credential request path (#129670, #123216, #132122). Before 2026.8.1, when an agent needed an API key for a tool, the value would land in chat or model context — even with redaction, you were still trusting the model and the UI not to surface the secret. The new path asks for the credential through a masked prompt, keeps it in an opt-in proxy, and only injects plaintext at the network boundary. The relevant PR titles: “Private credential requests,” “credential proxy,” and “SecretRef process-local sentinels” (#102008, #102009).

For a multi-tenant or shared-session setup this is the change that lets you stop apologizing. For a single-user self-hosted Gateway it’s still a strict improvement — secrets now live in fewer places. The companion change is “macOS service SecretRefs” (#99124), which preserves generated env-file values when a stale LaunchAgent gets repaired, closing a long-standing papercut where a Gateway reinstall would silently drop a SecretRef that was previously in the config.

The trade-off: the proxy introduces one more moving part. If you’re already using something like 1Password CLI or pass for secret injection, you may not need the new path. If you’ve been doing copy-paste-into-chat, you do.

Sessions stop being pinned to one Gateway

”Sessions beyond your Gateway” (#123280, #127752, #131744, #132374) is the feature name that obscures what’s actually happening. The Gateway can now hand a session off to a paired device or cloud worker, and the workspace follows it. When you pick the session back up later, you get a warm machine, project seeds, and the conversation history intact.

What this means in practice: if you’re running OpenClaw on a home server and want to fire off a long-running task while you’re on a laptop, the work doesn’t have to stay anchored to the home server’s uptime window. The cloud worker can take it, finish it, and you resume locally when you get back. The “Follow work as it happens” change (#125125 and friends) is the user-visible side: a durable session progress card that survives reloads, with subagent activity and accumulating edits shown across both web and native chat.

The risk: more handoff surfaces means more places where a session can deadlock. The 2026.8.1 release notes specifically address this in “Cloud session lifecycle” (fix entry) — preventing archive/delete/restart recovery from deadlocking behind an earlier worker move. This was apparently a real, reproducible failure mode in 2026.8.0.

Memory and context handling get quieter improvements

Two changes that look small in the changelog but matter:

• FTS-only memory startup — when memorySearch.provider is explicitly set to none, the Gateway no longer runs plugin capability discovery during cold start. For installs that have deliberately opted out of memory search, this shaves a noticeable amount off cold-start time.
• Small-context compaction (#100621) — previously, compact local models with tight context windows would enter compaction from the first token because the effective reserve didn’t account for the actual model context. The fix caps the reserve against the known window, so a small model can actually use its window without immediately getting compacted.

Neither is headline news. Both are papercuts that compounded for anyone running lean local models on the Gateway.

Subagent and Codex integration harden

The Codex integration picks up “GPT-5.6 Ultra and runtime switching” with Sol, Terra, and Luna supported across both the OpenClaw and Codex engines. The atomicity guarantee is the interesting bit: model, runtime, and thinking selection stay atomic through /model and fallback paths. If you’ve ever seen “model switched but thinking budget didn’t follow,” this is the fix. (#98021)

There’s also “Codex yielded native subagents” — the parent app-server subscription and shared client stay alive until yielded native subagent completion delivery settles, preventing lost wakeups and leaked one-shot cleanup. In practice, fewer orphaned subagents when an agent yields mid-task.

The Codex app-server protocol now requires 0.142 or newer. Older app-servers stop working. If you’re pinning an app-server version for reproducibility, bump it before you upgrade OpenClaw or the Codex integration will silently stop loading.

MCP and channel plugins: shared lifecycle

”Channel plugin ingress monitors” consolidates a shared plugin SDK monitor for durable admission, polling, pruning, claim identity validation, adoption handoff, and shutdown. Three plugins — IRC, Synology Chat, and Google Chat — migrate to the shared lifecycle. The benefit: a single place to fix admission bugs across plugins, rather than three slightly different implementations.

MCP gets two safety improvements: “MCP OAuth response bounds” rejects body-less foreign error bodies without calling their unbounded text() fallback (#98143), and “MCP loopback tool results” preserves schema-valid text, image, and embedded-resource content through HTTP tool calls (#100336). The loopback fix is the one I care about — malformed tool blocks now render as safe text instead of breaking the whole agent run.

Buzz: the chat platform gets a real identity model

Buzz (the bundled chat platform integration) gets the most attention in this release. The setup flow now reuses or generates the bot identity automatically, waits for Bot-role approval before falling back to identity-preserving Retry/Back controls, and verifies setup without posting test messages to rooms. There’s a “Buzz resumable setup” path for when setup gets interrupted, “Buzz inbound authorization” that applies shared sender and command authorization before agent dispatch, and “Buzz bot profiles” that persist optional display names plus configured owner attestations.

What this means: a Buzz bot now has a real identity lifecycle instead of being a transient webhook. For shared workspaces this matters — verified Bot-role membership, persistent profile metadata, and inbound authorization that doesn’t fall through to “let everything mention-gate in.” Multiple Buzz fixes are credited to @shakkernerd, who appears to have spent the cycle on this surface.

What’s still rough

A few things that didn’t make the release that I was hoping for:

• No native way to bundle multiple skills across profiles into one shareable archive. The ClawHub registry reads now retry transient HTTP 500s, which is good, but registry publishing still doesn’t support a “publish this whole set” flow for an agent that’s accumulated ten domain skills.
• The browser tab management is better at handling stuck tabs (the “Remote browser reliability” fix bounds persistent Playwright tab enumeration by the existing remote CDP timeout), but the underlying tab-cleanup story is still “you can kill a stuck tab.” That’s fine; not everything needs to be solved in one release.
• ”Memory session repair” now keeps daily dreaming ingestion bookkeeping outside session-corpus audit, but the actual repair logic for a corrupted memory index is still manual. If you’ve ever had memory status --fix quietly do nothing useful, this release doesn’t change that.

Where to actually start

If you’re upgrading from 2026.8.0:

1. Back up your config and state. The release notes say it explicitly. The Doctor tool has a recovery notes path that will surface interrupted auth-profile archive failures — read those before declaring the upgrade done.
2. If you use Codex integration, update your app-server to 0.142+ before starting the OpenClaw upgrade.
3. If you use SecretRefs, double-check the SecretRef values are still present after the upgrade. The macOS service SecretRef fix is for new repairs; it doesn’t re-inject values that were already dropped.
4. If you don’t use a memory search provider and want the cold-start speedup, set memorySearch.provider to none explicitly.

The changelog is 800+ lines. Most of it is fix entries with the same shape: “preserve X behavior when Y race condition would have stripped it.” That’s the right kind of release for a system with this much surface area. The headline features — credential masking, session portability, shared channel plugin lifecycle — are the parts that change what you can build on top.

References and where to dig further

• OpenClaw 2026.8.1 release notes on docs.openclaw.ai
• openclaw/openclaw on GitHub — release tag v2026.8.1, published 2026-08-31
• OpenClaw Changelog — full Unreleased section is the unreleased-next entry
• PR numbers cited above (e.g. #129670, #123216) are the merged-PR identifiers on openclaw/openclaw; the changelog calls them out individually so you can grep the GitHub UI for the diff
• For self-hosted installs with limited context budgets, the “Small-context compaction” fix (#100621) is worth reading the diff for — it changes how reserve is computed against the known model window, which interacts with how lean local models are configured