OpenClaw local-vision proposal reports CPU proof but keeps cloud-chat caveat

OpenClaw's proposed guided local-media setup now carries detailed September 15 evidence for real OCR and visual understanding on a Linux CPU host, alongside controlled transport checks. It extends managed llama.cpp setup rather than adding another downloader or process owner. The operator consequence is a possible path from hardware inspection through consent to verified local image routes. The proposal remains open, and the author explicitly separates local-route behavior from the broader chat model's handling of original images. That boundary is central to any privacy claim.

The facts

- PR #148193 asks users to review hardware, download and resource budgets before opting into local-media setup. - Activation requires both real OCR and vision checks; decline, cancellation or failed verification preserves the previous configuration. - The managed image routes do not download missing artifacts during inference or silently fall back to a cloud image provider. - The source explicitly warns that an existing cloud chat model can still receive original images under its native-vision policy. - Reported real-hardware coverage is Linux x86_64 CPU only; Metal, CUDA, Windows and Linux ARM remain unverified. - A later transport guard rechecks direct-loopback routing after asynchronous model preparation, before image bytes enter the shared adapter.

Why it matters

Local inference only becomes an operational feature when installation, routing and failure behavior agree. A model that runs on a laptop is not yet a setup flow that preserves existing chat and memory settings or refuses an accidental proxy path. The submitted evidence distinguishes real-model outputs from controlled HTTP sinks, a useful separation because transport tests cannot establish recognition quality. It also distinguishes unchanged media setup from an upstream index migration: the published-state compatibility cell reports core schema and chunking changes before checking that media setup causes no additional rebuild.

Current

Inspected on 2026-09-16. The OpenClaw stable-channel baseline is 2026.9.4 on npm latest; extended-stable 2026.6.35. The primary source was open (not merged) when captured. Mainline merges, open proposals and packaged releases are distinct availability states; the release baseline does not establish that a proposal has shipped.

Evidence

Primary evidence: openclaw/openclaw PR #148193 (https://github.com/openclaw/openclaw/pull/148193). Supporting context: openclaw/openclaw issue #148056 (https://github.com/openclaw/openclaw/issues/148056); OpenClaw — official product README (https://raw.githubusercontent.com/openclaw/openclaw/main/README.md). The source bodies and linked context were inspected. Test results quoted from contributors remain attributed reports, not independently reproduced experiments.

Operator take

Treat this as a reviewable local-processing direction, not a reason to announce that all images stay on-device. Inspect the entire chat and media route before making such a claim. If evaluating the branch, reproduce consent, failed verification, missing artifacts and proxy changes with synthetic images on the actual supported hardware. Keep CPU observations separate from GPU performance expectations. The final decision still belongs to maintainers on plugin-owned recipes, the local_image tool and supported backend scope; green CI can inform that decision without making it for them.

Caveat

The source reports successful hosted CI and source-bound transport checks, while a retained older paragraph still says no new CI success is claimed. This article follows the newer dated evidence and preserves its revision boundaries. Results are author-reported, not rerun here; product acceptance and broader hardware qualification remain pending.