Experimental Features

Settings > Experimental holds toggles for features that are still maturing. “Experimental” doesn’t mean unsafe — these are shipped and tested — but the surface area, defaults, or persistence shape may shift between releases. Toggling one off is always reversible.

Graduated features

The following features have graduated out of Experimental and are now always available. Their experimental toggles have been removed — the features are on by default and their settings live in their own sections:

Collaboration — multi-user collaborative workspace sharing. Now has its own Collaboration settings section visible by default.
Community — the community registry and trust system. Now has its own Community Registry & Trust settings section.
Claude Code Plugins — the Claude CLI marketplace integration. Now has its own settings section visible by default.

Legacy “off” values for these graduated features are ignored after upgrading and no longer hide their surfaces. Agent backend choices such as Ollama, OpenAI API, and Codex also live in Settings > Models as first-class model/provider controls rather than experimental toggles.

Claude Remote Control

Bridges your local Workbench workspace to a Claude Code session running on claude.ai (web or mobile). Drive an agent from your phone while looking at the diff on a laptop, or hand off a session between devices.

Has its own setup flow, troubleshooting steps, and constraints (some Workbench features are deferred while a bridge is live).

→ Full doc: Claude Remote Control

Work Tracking

Work Tracking has moved to r26D Company Settings. It is no longer an experimental feature.

Claude Code Usage

This is the only remaining experimental toggle. It gates the Anthropic OAuth Usage API integration: the chat composer’s usage meter normally shows local token totals (or, on Codex backends, live Codex account/rateLimits/* quotas), but for Anthropic / Custom Anthropic sessions the meter can additionally surface real Claude subscription quotas — 5-hour session, weekly all-model, weekly Sonnet/Opus, extra usage — by querying Anthropic’s usage API directly. Requires a Pro or Max plan with standard login (env-var auth tokens don’t have the OAuth scopes the API needs).

When this is off, the meter on Anthropic-family sessions renders in a greyed-out state that links here when clicked. Codex (Native and Subscription) / OpenAI / OpenRouter / Ollama sessions are unaffected — the meter stays live for them because it’s driven by Codex’s own rate-limit notifications and Workbench’s per-turn token records, not the Anthropic Usage API.

When this is off, Workbench never reads your Claude Code OAuth tokens for usage data — they stay strictly in the keychain / .credentials.json for the claude CLI’s own use. The Claude Code sign-in flow in Settings > Models remains available because agent credential failures need a recovery path independent of usage tracking.

Enabling this toggle opens a confirmation dialog that explains how the data flows and warns that this endpoint is undocumented, so programmatic access may not be permitted under Anthropic’s Consumer Terms. Disabling never prompts.

→ What the panel shows and how it’s cached: Usage & Metrics

OpenRouter balance in usage meter

Shows your live OpenRouter credit balance alongside local token totals for OpenRouter-backed sessions. Turn this off to avoid polling the OpenRouter credits endpoint.

Stability outlook

Toggle	What’s still moving
Claude Code Usage	Cache invalidation around tier changes; richer breakdown views; (non-Anthropic providers already render local-aggregate totals unconditionally).
OpenRouter balance in usage meter	Polling cadence and how the balance is surfaced in the meter.

If a toggle moves to “default-on” or graduates out of experimental, the release notes for that feature describe any one-time migration. Experimental toggles are otherwise independent, never bundled.