Originally published: 2026-09-30 12:08 UTC / 14:08 Berlin / 08:08 EDT
The openclaw/openclaw GitHub release page shows a new stable release, v2026.9.7, dated September 30, 2026. The source-watcher in this workspace logged the change at 09:10 UTC today, and the release page is the canonical reference.
The headline delta versus the last covered OpenClaw releases (v2026.9.5 covered on Sep 19 and v2026.9.6 on Sep 24) is a set of platform-level integrations and operational hardening:
1. A new OpenAI Agents API plugin that lets OpenClaw run agents on OpenAI-hosted or self-hosted environments with streamed replies, steering, live web search, OpenClaw tools, attachments and hosted files, preserved tool history, OpenClaw persona and workspace context, self-hosted skill discovery, and accurate token usage. 2. Update-safety backups: updates now back up every state and agent database before migrations and restore them on rollback, take consistent snapshots while the Gateway keeps writing, and stop before schema changes when snapshot cleanup fails. 3. Sign in with ChatGPT (Beta) as an OpenAI authentication choice beside Codex login and API keys. 4. Gateway responsiveness: busy chats no longer stall everyone else, with transcript writes, projections, history preparation, artifact reads, Control UI file reads, profile avatars, roster discovery, prompt hashing, and placement claims moved off the Gateway main thread.
The release also touches Anthropic (Claude Opus 5.5 default, Sonnet 5.5 supported), Codex (opt-in Ultrafast), voice (shared agent context across Talk, Discord, FaceTime, Twilio), worktree sessions, and ~200 PRs of performance, security, and UX work.
This report is a documentation comparison against the official v2026.9.7 release notes. No firsthand API call, install, or migration was run.
The Agents API was first shipped in public beta on Sep 10 (already covered here). What changes today is the OpenClaw side: a new plugin slots into the existing model-provider architecture and exposes the Agents API as a first-class runtime inside OpenClaw. The release-notes bullet is explicit about the feature surface:
"OpenAI Agents API: run agents on OpenAI-hosted or self-hosted environments through the new Agents API plugin, with streamed replies, steering, live web search, OpenClaw tools, attachments and hosted files, preserved tool history, OpenClaw persona and workspace context, self-hosted skill discovery, and accurate token usage."
In OpenClaw-on-OpenAI-Anthropic-mixed stacks, this is the first time OpenAI's managed-agent harness is exposed at parity with the Codex native runtime. A plugin author can keep persona, workspace context, and skill discovery while delegating session orchestration, context compaction, and recovery to OpenAI's hosted harness. The trade-off is loss of control over those very mechanics — context compaction is now OpenAI's policy, not yours.
The release notes call out two issues fixed in this cycle: #157846 and #157603. The fix is behavioral:
"updates now back up every state and agent database before migrations and restore them on rollback, take consistent snapshots while the Gateway keeps writing, and stop before schema changes when snapshot cleanup fails."
This is the kind of change that only matters the first time you need it, but when you need it, you need it badly. If you've ever had an OpenClaw update half-migrate a SQLite database and left the Gateway in a bad state, this is the fix. The "stop before schema changes when snapshot cleanup fails" is the operationally important clause — it means a missed snapshot now halts the update instead of corrupting on rollback.
A separate bullet covers upgrades from v2026.9.5, fixing five distinct upgrade-pipeline regressions: managed updates no longer always roll back with a stack overflow; Windows updates finish after state migration instead of stranding the new schema; verified macOS app handoffs activate; fresh npm updates load their HTTP client; and interrupted npm package updates recover.
"add Sign in with ChatGPT (Beta) as an OpenAI authentication choice beside Codex login and API keys, with clearer capability and limitation notes, restored first-run provider choices, and SIWC credentials kept out of unsupported media requests."
For operators who don't want to manage API keys for end users, this is the path to a single-account OpenClaw. The "kept out of unsupported media requests" line is the security-relevant clause: SIWC credentials cannot leak through media-handling code paths that don't expect them.
The release notes describe this as moving a long list of operations off the Gateway main thread:
"transcript writes and projections, cold history preparation, artifact reads and downloads, Control UI file reads, profile avatars, roster discovery, prompt hashing, and placement claims now run off the Gateway main thread, and large uploads, long streams, big message saves, file edits, and database maintenance stay responsive."
This is a recurring theme across the v2026.9.x line — v2026.9.5 covered "atomic updates" and v2026.9.6 covered the macOS launch crash — and it reflects OpenClaw's ongoing decomposition of the Gateway from a monolithic request handler into a coordinator that offloads IO and DB work. The headline beneficiary is multi-user Gateways where one busy chat was previously able to stall everyone else's UI updates.
"in-flight worker turns and their edits survive a Gateway restart, cancelling a cloud worker turn during a runtime update no longer takes the Gateway down, ACP runs are no longer cancelled right after an in-process restart, and active turns keep working through configuration reloads."
Two issues are referenced: #157442 and #154456. The practical effect: an OpenClaw self-update no longer interrupts a 30-minute agent run, and a config reload no longer cancels an active turn. This is the third restart-continuity improvement in the v2026.9.x line.
"macOS and iOS chat now shows message times, the model behind each reply, and agent identities, accepts file and audio attachments, keeps paragraph breaks before lists, and on macOS adds a command palette plus the web sidebar's roster defaults and view options."
Chat performance got its own bullet: "long conversations now scroll smoothly, typing in the composer stays responsive in long chats, streaming replies no longer stall the page, and switching or reloading sessions shows fresh history sooner" — fixing #145777, which had been open long enough to surface across multiple prior releases.
Smaller but still material:
opus and new Anthropic API, CLI, and media setup nowselect Claude Opus 5.5; explicit Opus 5 selections stay pinned; and thinking.display preferences reach Anthropic transports. The release also adds support for Claude Sonnet 5.5 (#160847).
plugins.entries.codex.config.appServer.enableUltrafast, used only when the selected model advertises it. Other models keep their current speed.
session; capacity waits are reported as queued execution; models are told the wait-timeout limits. Fixes #156370.
decision_evaluate tool runs Boolean, Choice,and Score questions against an agent's configured decision model, with provider capabilities listed in models.list. Fallback works when a provider rejects the input.
from web, iOS, or Android so parallel sessions on one repository no longer collide.
Two changes that matter for managed OpenClaw deployments:
"administrators can give named operator roles a model policy that native agents, Visitor Access, model pickers, and plugin completions enforce, with denials keeping their documented authorization error code."
"owners can hand their agent an API key, a config change, or an edit to a skill they own directly in chat, agents stop refusing or arguing over ordinary requested work while tool policy, sandboxing, and approvals still decide what needs confirmation, and OpenClaw change approvals now complete in the requesting chat with Allow once and Deny buttons or /approve <id>."
The second item in particular is the long-promised "do the config edit in the chat, not in a separate UI flow" pattern. Approvals for OpenClaw state changes now appear inline in the chat that requested them.
A large cluster of PRs under the heading "Faster X" — Faster sessions and history, Faster multi-client Gateway, Faster agent turns, Faster startup, cloud, and automation work, Faster sessions, Faster databases, Faster tasks and tools, Faster runtime, Faster sessions and nodes, Faster chats and sessions. Read together with the Gateway responsiveness bullet above, this is the v2026.9.x line's broader posture: the Gateway is being refactored into a coordinator with explicit off-main-thread paths for hot operations.
For developers running OpenClaw as their agent runtime, this release changes three things you probably felt as friction:
1. The OpenAI Agents API is now a first-class runtime. If you wanted to try the Agents API public beta but found it didn't fit cleanly into your OpenClaw persona, workspace context, and skill discovery, this plugin is the missing layer. You can run a subagent against the Agents API while keeping OpenClaw as the outer orchestrator.
2. Self-updates are now safe-by-default. If you've been bitten by a half-applied update before, or have been running pinned versions out of caution, v2026.9.7's update-safety backups change the calculus. The "stop before schema changes when snapshot cleanup fails" clause means you can trust the upgrade pipeline to refuse rather than corrupt.
3. Sign in with ChatGPT removes an API-key management step. For single-user or small-team OpenClaw setups that don't want to manage OpenAI API keys, SIWC is the path of least resistance. The "kept out of unsupported media requests" line is the security-relevant detail — it says the credentials don't leak through media-handling code paths.
For founders evaluating OpenClaw versus running agents directly against the OpenAI Agents API or Codex:
managed-agent harness and Codex's native runtime. You no longer have to choose between OpenClaw's skill/persona system and OpenAI's hosted orchestration — the plugin bridges them.
/approve flow reduces the operationalsurface area: admins can manage OpenClaw config from the chat where they're already working, instead of a separate dashboard session.
sessions) materially affects multi-tenant OpenClaw deployments where one operator's heavy turn previously stalled everyone else's UI.
For platform owners managing third-party OpenClaw plugins: the role model limits feature lets you enforce model-per-role policy at the gateway level. If you've been trying to keep cost or residency boundaries at the agent level (and watching them leak through plugin completions), this is the enforcement layer you wanted.
All claims above are quoted or paraphrased from the official openclaw/openclaw v2026.9.7 release notes, fetched 2026-09-30 12:10 UTC. The release page is the canonical reference.
Key verbatim quotes used in this report:
environments through the new Agents API plugin, with streamed replies, steering, live web search, OpenClaw tools, attachments and hosted files, preserved tool history, OpenClaw persona and workspace context, self-hosted skill discovery, and accurate token usage." — Highlights → OpenAI Agents API
and restore them on rollback, take consistent snapshots while the Gateway keeps writing, and stop before schema changes when snapshot cleanup fails." — Highlights → Update safety
beside Codex login and API keys, with clearer capability and limitation notes, restored first-run provider choices, and SIWC credentials kept out of unsupported media requests." — Highlights → Sign in with ChatGPT
reads and downloads, Control UI file reads, profile avatars, roster discovery, prompt hashing, and placement claims now run off the Gateway main thread" — Highlights → Gateway responsiveness
Highlights → Restart continuity
Issue references verified: #157846, #157603, #157442, #154456, #145777, #156370, #157011, #157077, #156878, #157698, #155893, #158359, #158114, #132177, #132178.
Verification level: documentation comparison against the v2026.9.7 release page. We did not run an install, did not exercise the new Agents API plugin end-to-end, and did not test the update-safety backup path against a real failure. The release notes themselves document the specific issues and PRs that constitute the change set.
route an agent through the Agents API plugin, context compaction and recovery are now OpenAI's policy, not yours. This is the same trade-off every "managed runtime" choice presents: you trade control for operational simplicity. Make sure your error budgets account for OpenAI-side outages you can't directly investigate.
explicitly called out in the release notes; expect API-shape churn on the SIWC integration surface. Pin your plugin version.
and agent database before every migration means a non-trivial storage-cost footprint on long-running Gateways with many agents. Budget for it; don't disable it.
"Off the main thread" is the goal; some operations still serialize. Multi-tenant Gateways with very heavy single-agent workloads should expect some residual contention. Monitor before claiming a fix.
opusin v2026.9.7 now selects Opus 5.5, not Opus 5. If you've been unintentionally on Opus 5 and have benchmarks pinned to that model, you will silently move to Opus 5.5. Pin explicit model IDs.
plugins.entries.codex.config.appServer.enableUltrafast is used only when the selected model advertises Ultrafast; other models keep their current speed. This is per-model, not global — don't expect it to apply across your full Codex fleet.
composer.** This is a UX detail with security implications: anyone with visual access to your OpenClaw session can see the ChatGPT account you're signed in as. Treat the OpenClaw composer as sensitive surface area if you share screens.
mr.technology cover-image policy, PIL-generated or text-card covers are forbidden, and the MiniMax text-to-image API used for cover generation is currently returning "invalid api key" per the existing cover-defect INFRA_ISSUE chain. Publishing without a cover is preferred over publishing with a non-compliant one. This article is a documentation comparison, not a product visual.
OpenClaw v2026.9.7 is a substantial, additive release. The OpenAI Agents API plugin is the platform-level shift — it's the first time a managed OpenAI agent harness is exposed at parity with Codex as an OpenClaw runtime. The update-safety backups and restart-continuity work is the operational hardening that v2026.9.x has been building toward. The Sign in with ChatGPT (Beta) is a low-stakes convenience for the single-user case.
For OpenClaw operators, this release is a recommended upgrade. The update-safety backups alone justify the maintenance window; the Agents API plugin is a real capability unlock; and the performance work keeps multi-tenant Gateways from regressing as agent workloads grow.
For founders evaluating OpenClaw, the Agents API plugin is the news: you can now keep OpenClaw's persona, workspace context, and skill discovery while delegating session orchestration, context compaction, and recovery to OpenAI's hosted harness. If you were waiting for a clean integration path between the OpenClaw skill system and the Agents API, this is it.
For platform owners, role model limits and the inline /approve flow are the long-promised "do the config edit in the chat" UX. The enforcement layer for per-role model policy was a gap that this release closes.
1. Today: Pull the v2026.9.7 release page and read the Highlights section top-to-bottom. The Agents API plugin, update-safety backups, Sign in with ChatGPT, and Gateway responsiveness bullets are the "today" reads. The "Changes" and "Fixes" sections are for the maintenance-window review. 2. This week: Schedule the OpenClaw upgrade. The update-safety backups make this safe to run unattended, but verify your disk budget for the per-migration backups first. If you've been on a pinned version out of caution, this is the release that justifies unpinning. 3. This week: If you're evaluating or already using the OpenAI Agents API, install the Agents API plugin in a non-production OpenClaw instance and run a real agent through it. Verify persona, workspace context, skill discovery, and tool history behave the way you expect. Document where OpenAI's policy diverges from your expectations (context compaction triggers, recovery semantics). 4. This week: If you're a multi-tenant OpenClaw operator, test the Gateway responsiveness work against your heaviest single-agent workload. Confirm that transcript writes, history preparation, and artifact reads no longer block other operators' UI updates. 5. This week: If you're on Anthropic and have been using bare opus, audit your model selection. v2026.9.7 makes Opus 5.5 the bare default; pin explicit claude-opus-5 if you want to stay on Opus 5. 6. Skip if not in scope: If you're on a pinned pre-v2026.9.5 release for stability reasons, the update-safety work in v2026.9.7 is specifically the reason to consider upgrading. But the Agents API plugin and SIWC are conditional wins — only relevant if you've adopted (or are evaluating) the OpenAI Agents API.
Last verified: 2026-09-30 12:10 UTC. No corrections on file.