OpenClaw 2026.10.5: Worker Inference, X Replies and Upgrade Checks

OpenClaw released stable version 2026.10.5 on October 18, adding worker-local inference and a channel that answers X mentions. The release notes also describe restart-recovery improvements and an approval change that plugin authors should check before upgrading.

For teams running agents across machines, the useful question is where the agent loop runs and which machine holds authority. For teams considering a public-facing bot, it is who can trigger a reply and what that reply costs.

What worker-local inference changes

OpenClaw workers normally send model requests through the Gateway, its coordinating service. The new option runs the native agent loop and provider requests on a paired worker. Coding tools execute there too, while the Gateway retains transcripts and control over session access and permissions.

The versioned deployment guide makes this an explicit choice. Operators need compatible Gateway and node builds, worker hosting enabled, and a named paired-device profile configured for worker inference. An ordinary device placement continues using Gateway inference. Upgrading alone does not switch existing sessions.

Here, “local” describes where the agent loop and provider calls execute. The model endpoint can still be an external service. This feature by itself establishes neither offline operation nor that model weights run on the worker.

Credential handling needs particular care. The documented deployment requires an external credential custodian or provider proxy to hold the actual provider keys. The worker receives scoped, opaque authentication values for that service. OpenClaw does not supply the custodian or verify that the configured value really is a scoped token.

Code running under the same operating-system account can access that account's files, environment and process memory. A workspace grant does not provide operating-system isolation. Teams still need to control the worker account, network access and token permissions.

There are practical limits: Windows worker-local inference is deferred, and OpenClaw's additional nested-container worker mode is unsupported. Automatic compaction and retry remain disabled in this worker runtime. Losing the Gateway connection stops an active local provider request; reconnecting can settle the interruption, but that request does not resume.

X replies need a deliberately narrow setup

The new X channel turns mentions of a bot account into conversations and posts the answer as a public reply. By default, only allowlisted authors can trigger it. DMs, original posts and media uploads are unsupported.

An optional guest mode lets other people ask repository questions. Its restricted tools and separate sessions make it relevant to public project support, although operators must configure the required repository containment and queue settings. Guests cannot edit files or run commands through this mode.

Setup requires an X account and API authorization. X API charges are separate from model-token usage. The plugin's default estimated-spend limits are $100 per UTC day and $1,000 per billing cycle for each account. These are configurable safety limits, not a subscription price. Already-delivered streaming events can push recorded spend past a limit.

For a first trial, start with a small allowlist and a budget suited to the expected traffic. Check the reply quality and API spend before opening guest access.

Check transforming plugins before upgrading

The tool-policy migration guidance matters if a plugin changes tool arguments after a trusted policy has approved them inline. An ordinary hook must now request separate approval for its final arguments. Otherwise, the changed call is blocked.

For example, a plugin that changes an approved file destination needs approval for the replacement parameters. Unchanged calls remain allowed, subject to other checks. Test affected hooks in a staging environment so a deliberate policy change does not appear as an unexplained failure during normal work.

The release also reports fixes for interrupted conversations, first messages during workspace preparation and subagent recovery after Gateway restarts. Those are maintainer-reported improvements, not an independent reliability benchmark.

For an existing deployment, test one representative task through interruption and recovery before expanding its workload. For a new worker deployment, verify where requests go and how credentials are scoped. OpenClaw now offers more placement flexibility, but the deployment still determines how much authority each worker can exercise.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help developer-tool and enterprise AI teams turn architecture changes into clear technical stories, launch explainers and adoption guides.

Book a 30-minute call.