OpenClaw
OpenClaw is tracked as a living system dossier: a self-hosted channel gateway, operator surface, approval boundary, and public-publication handoff layer.
System dossier
Role: Owner-gated operational gateway for channels, previews, approvals, and public-write handoff. Research question: Can an always-on agent act inside everyday channels while keeping delivery, permissions, public sources, and publication approval auditable? Current observation: The v2026.9.6 macOS DMG was rebuilt and notarized after a launch crash; this is a release-maintenance fact, not evidence of broader runtime reliability.
Source ledger
Publishable sources attached to this record.
| # | Source | Role | Public status |
|---|---|---|---|
| 1 | docs.openclaw.aidocs | primary receipt | source_urls |
| 2 | docs.openclaw.aidocs | supporting receipt | source_urls |
| 3 | docs.openclaw.aidocs | supporting receipt | source_urls |
| 4 | docs.openclaw.aidocs | supporting receipt | source_urls |
| 5 | github.comrepo | supporting receipt | source_urls |
| 6 | docs.openclaw.aidocs | supporting receipt | source_urls |
| 7 | github.comrepo | supporting receipt | source_urls |
| 8 | docs.openclaw.aidocs | supporting receipt | source_urls |
| 9 | github.comrepo | supporting receipt | source_urls |
| 10 | github.comrepo | supporting receipt | source_urls |
| 11 | github.comrepo | supporting receipt | source_urls |
| 12 | docs.openclaw.aidocs | supporting receipt | source_urls |
Current state, as of Oct 9, 2026
Where this stands now
holding OpenClaw v2026.9.8 is released and adds configurable agent‑skill loading; macOS was rebuilt in v2026.9.6, but no evidence confirms Telegram checks were waived or the Android APK was omitted.
OpenClaw v2026.9.8 has been published and introduces configurable agent‑skill loading. The release notes do not detail any changes to Telegram integration checks or Android packaging. Earlier in the series, v2026.9.6 received a rebuilt, notarized macOS DMG to fix a launch crash, while the npm package remained unchanged. The system continues to operate as a gated operator surface for channel approvals and public‑write handoff, with the newest features focused on skill configuration rather than platform‑specific packaging changes.
What changed since the original text
- v2026.9.8 adds configurable agent‑skill loading. 1
- v2026.9.6 macOS build crashed at launch and was replaced by a rebuilt, notarized DMG on 2026‑09‑24. 2
Open questions
- Did the v2026.9.8 release explicitly waive Telegram integration checks?
- Was the Android APK intentionally skipped for v2026.9.8?
Watch next: A release note or upstream announcement confirming the status of Telegram checks and Android packaging for v2026.9.8 would update this assessment.
- Weekly update: v2026.9.8 Agent Skills: OpenClaw adds configurable agent-skill loading update, Sep 28, 2026
- Weekly update: v2026.9.6 macOS rebuild: OpenClaw macOS app rebuilt and notarized to fix launch crash in v2026.9.6 update, Sep 21, 2026
Rewritten automatically by the New Runtime knowledge desk from this record's weekly updates, releases and archive. The original text below is kept as written; where the two disagree, this section is newer.
Weekly knowledge desk
Weekly updates
5 reviewed updates to this system, newest first. Every update is checked against its evidence before it changes the record.
v2026.9.34 Safer Updates: OpenClaw 2026.9.34 hardens upgrades and critical runtime paths in an extended-stable release
This gateway-only extended-stable release brings selected changes from the end of September 2026, plus security, reliability, and performance fixes. Highlights include safer atomic agent file edits, reconciliation of official plugins during managed upgrades, model startup and account-selection fixes, and updates to Node runtime requirements and production dependencies. (2 earlier releases since the last dossier entry are covered by this one.)
New release v2026.9.34 - Safer Updates (released Oct 24, 2026). This gateway-only extended-stable release brings selected changes from the end of September 2026, plus security, reliability, and performance fixes. Highlights include safer atomic agent file edits, reconciliation of official plugins during managed upgrades, model startup and account-selection fixes, and updates to Node runtime requirements and production dependencies. Release source
Current observation: OpenClaw v2026.9.34 is a gateway-only extended-stable release with security and reliability fixes, safer managed updates and agent edits, and refreshed runtime requirements.
- OpenClaw openclaw 2026.9.34 external, Oct 24, 2026
v2026.10.5 Restart-safe Operations: OpenClaw 2026.10.5 makes updates and conversations recover more reliably across rest
The release improves update completion and recovery across devices and platforms, preserves interrupted turns and session data through Gateway restarts, and makes Gateway reads and startup lighter. It also adds an X mentions channel, voice-call features, worker-local inference, and Gateway access to ClawHub plugins and skills. (3 earlier releases since the last dossier entry are covered by this one.)
New release v2026.10.5 - Restart-safe Operations (released Oct 18, 2026). The release improves update completion and recovery across devices and platforms, preserves interrupted turns and session data through Gateway restarts, and makes Gateway reads and startup lighter. It also adds an X mentions channel, voice-call features, worker-local inference, and Gateway access to ClawHub plugins and skills. Release source
Current observation: OpenClaw v2026.10.5 adds more self-recovering updates and restart-safe conversations alongside a lighter Gateway, an X channel, and expanded voice and ClawHub features.
- OpenClaw openclaw 2026.10.5 external, Oct 18, 2026
v2026.10.2 Reliability: OpenClaw fixes upgrades, agent delivery, and Gateway responsiveness
Version 2026.10.2 makes upgrades from 2026.10.1 and several Windows, cloned-filesystem, and private-Node update paths complete cleanly. It also addresses agent reply and tool-run failures, Gateway hangs, secret-redaction gaps, and defaults OpenAI and Anthropic API-key routes to provider-native compaction. (2 earlier releases since the last dossier entry are covered by this one.)
New release v2026.10.2 - Reliability (released Oct 12, 2026). Version 2026.10.2 makes upgrades from 2026.10.1 and several Windows, cloned-filesystem, and private-Node update paths complete cleanly. It also addresses agent reply and tool-run failures, Gateway hangs, secret-redaction gaps, and defaults OpenAI and Anthropic API-key routes to provider-native compaction. Release source
Current observation: OpenClaw v2026.10.2 was published with update and recovery fixes, provider-native compaction defaults, and broad fixes for agent delivery and Gateway responsiveness.
- OpenClaw openclaw 2026.10.2 external, Oct 12, 2026
v2026.9.8 Agent Skills: OpenClaw adds configurable agent-skill loading
This release adds configurable agent-skill loading. The notes identify the release as v2026.9.8 but do not provide details about the change. (4 earlier releases since the last dossier entry are covered by this one.)
New release v2026.9.8 - Agent Skills (released Oct 3, 2026). This release adds configurable agent-skill loading. The notes identify the release as v2026.9.8 but do not provide details about the change. Release source
Current observation: OpenClaw v2026.9.8 was published, with Telegram integration checks waived and the Android APK skipped for this release.
- OpenClaw openclaw 2026.9.8 external, Oct 3, 2026
v2026.9.6 macOS rebuild: OpenClaw macOS app rebuilt and notarized to fix launch crash in v2026.9.6
The original v2026.9.6 macOS build crashed at launch. A rebuilt, notarized DMG was published on 2026-09-24 to replace the faulty binary. The npm package remains unchanged, and the new DMG can be downloaded directly or via the in‑app update from v2026.9.5. (8 earlier releases since the last dossier entry are covered by this one.)
New release v2026.9.6 - macOS rebuild (released Sep 23, 2026). The original v2026.9.6 macOS build crashed at launch. A rebuilt, notarized DMG was published on 2026-09-24 to replace the faulty binary. The npm package remains unchanged, and the new DMG can be downloaded directly or via the in‑app update from v2026.9.5. Release source
Current observation: The local gateway is running on the stable v2026.7.1 line, while upstream v2026.7.2-beta.5 pushes harder on state recovery, durable channel delivery, session branching, MCP Apps, structured approvals, meetings, wearable voice surfaces, and local inference.
- OpenClaw openclaw 2026.9.6 external, Sep 23, 2026
This dossier treats OpenClaw as a longitudinal operating system experiment, not as a generic agent product page. Public release notes show how the upstream runtime is changing. The local corpus supplies the operating lens: where OpenClaw belongs in a real newsroom, where it must stay bounded, and what kind of proof is required before an agent can touch a public channel.
No private discovery feeds, credentials, host details, raw logs, handoff paths, message identifiers, unpublished drafts, or access-policy internals are exposed here.
Running locally as the private gateway line; latest upstream beta is not the production baseline.
State recovery, durable delivery, branching, MCP Apps, approvals, meetings, Wear OS, and local inference.
OpenClaw may preview and publish only after explicit owner approval.
Current public verdict
OpenClaw is most useful here as a standing operator surface: the agent can live inside channels, keep a private conversation with the owner, show formatted previews, and perform selected public write actions only after explicit approval. That makes it valuable and dangerous in the same place.
The local rule is deliberately conservative: OpenClaw is not the product backend, not the source of truth, and not the first-pass newsroom processor. It is the approval, delivery, and heavy-research lane. Durable facts, source provenance, editorial state, and public claims belong in explicit project systems; OpenClaw can act on them only through a bounded handoff.
The assistant answers wherever a human happens to be typing.
The same agent layer can operate through Telegram, dashboard, mobile, and other channel surfaces.
A human asks the model to send something and then audits what happened.
Draft, source eligibility, private preview, explicit approval, delivery result, and ledger all become separate steps.
Security depends on obscurity, channel etiquette, or the model behaving well.
Allowlist, mention rules, approvals, tool scope, and owner-only commands become the real boundary.
Development Timeline
-
The agent became a standing gateway, not another chat tab
Local operating notes put OpenClaw on the always-on side of the lab: channels, dashboard, tools, and mobile reach, while durable newsroom state stays outside the runtime.
Why it matteredA channel agent changes when work can start: not only at the laptop, but wherever the owner can message the system.
Corpus lensThe same power is bounded: OpenClaw may coordinate work, but the repository, source registry, Supabase path, and public site keep the durable record.
Normality shiftAn agent runtime becomes infrastructure; the question moves from "can it answer?" to "what is it allowed to operate?"
-
Channel reliability became the product surface
The release focused on rough edges: misplaced replies, stuck sends, reconnects, model setup failures, admin defaults, Telegram progress, duplicate replies, and recovery after stuck channel messages.
Why it matteredFor an always-on agent, "the model answered" is not enough. The answer must land in the right channel, thread, and message context.
Corpus lensThis maps directly to the QWG publication gate: a private preview or receipt is useful only when delivery can be checked and not inferred.
Normality shiftDelivery becomes a first-class capability, not plumbing hidden behind the chat interface.
-
The control plane moved closer to daily work
Upstream v2026.7.1 expanded the Control UI, onboarding, official apps, model/provider support, Codex and connected coding-agent workflows, Telegram behavior, scheduled work, sessions, goals, workspace terminals, and gateway recovery.
Why it matteredOpenClaw is less a bot wrapper and more an operator cockpit for concurrent conversations, tasks, costs, files, models, and approvals.
Corpus lensThe local gateway is on this stable line, but the public dossier still treats release claims separately from workflow-tested publication behavior.
Normality shiftChannel-native agents need dashboards because the hard part is no longer only answering; it is supervising work already in motion.
-
The beta points toward remote work and stricter channel control
The latest observed beta emphasizes remote coding sessions, native automation parity, Android voice wake, headless Linux node capabilities, safer channel operation, guided setup, gateway/session recovery, and Linux desktop packaging.
Why it matteredThe runtime is pushing beyond "message the agent" toward distributed work surfaces and nodes that can carry context, sensors, and automation.
Corpus lensThis is not local production evidence yet. It marks what to test next: durable Telegram ingress, owner permissions, and recovery around interrupted handoffs.
Normality shiftThe agent gateway starts to look like a small operations layer spread across devices, not a single bot process.
-
The beta turns remote work into recoverable operations fabric
The newest observed beta adds quarantine-backed state safety, crash-recoverable SQLite snapshots, durable filesystem publication, shared channel ingress and dead-letter recovery, message-level session rewind and forks, ticketed MCP Apps, cross-surface questions and approvals, meeting transcript collection, Wear OS talk controls, guided setup, local-provider detection, and RAM-gated llama.cpp/Gemma inference.
Why it matteredThis reads like a response to the everyday failures of always-on agents: corrupted state, lost channel messages, stale conversation branches, approval bottlenecks, meeting context leakage, and setup friction.
Corpus lensFor QWG, the important question is not whether every surface exists. It is whether recovery, approvals, transcripts, and local inference can be proven without weakening the publication gate.
Normality shiftThe runtime is becoming a stateful operations layer: many surfaces, one recoverable task record, and explicit approvals around every risky edge.
-
Owner-gated publication became the hard rule
Local newsroom workflows now route public-channel publication through sanitized handoff packages, private receipt, requested RichMessage preview, explicit owner command, delivery verification, and idempotent ledger recording.
Why it matteredPublic posting is the irreversible edge. The system must separate draft creation from final publication authority.
Corpus lensDiscovery-only aggregators stay private; unresolved primary sources block publication; Codex prepares but does not send public Telegram messages.
Normality shift"Agent can publish" becomes "agent can publish only selected, eligible, previewed items after an explicit owner command."
Local Experiment Axes
Can the owner reach the agent from daily surfaces without making the bot public or noisy?
Do receipts, previews, and final messages land in the intended private or public destination?
Do allowlists, owner-only actions, tool scopes, and explicit approvals remain stronger than prompts?
Does the private preview show the exact visible post, buttons, and formatting that would be published?
Are private discovery feeds, tracking links, raw notes, and blocked source items kept out of public output?
Can the system prove a selected item was published once and reuse the recorded result instead of duplicating it?
Next Verification Work
The next OpenClaw work should stay operational and narrow:
- Verify the v2026.7.2 durable-ingress and channel-safety claims in a private test lane before treating them as workflow-tested.
- Run one receipt -> RichMessage preview -> explicit approval -> publish ledger rehearsal with non-public or dummy output.
- Test failure recovery for an interrupted preview or final delivery without exposing private source material.
- Keep OpenClaw out of routine source-card processing until latency, cost, and handoff reliability are boring.
- Reclassify only the tested axes; do not let an upstream release title upgrade the local verdict by itself.
September 2026 update
The v2026.9.6 macOS DMG was rebuilt and notarized after a launch crash, while the npm package did not change. This is a release-maintenance fact, not evidence of broad reliability. Launch, approval-gated delivery, restart recovery, and idempotent receipts still need separate smoke tests.