Wicked Studio puts plans, agent work and approval gates into one chat

Wicked Studio wants supervising coding agents to feel like one conversation. Its October 17 version 0.10.0 turns a job’s page into a chat: your messages, named agent turns, a proposed plan, approval questions and completed steps share the same thread. The documented workflow waits for a person to confirm the plan before work starts.

This is a console for wicked-crew, which orchestrates the coding CLIs you already use, rather than another model provider. The project documents container images for Claude Code, Codex, GitHub Copilot, OpenCode and Pi. Changes and Evidence links sit alongside the final response, while the underlying run record is available in a Details section.

Memory also gets its own Desk tab, where users can edit remembered information and review proposals. Steer remains the place for enforceable rules and policy proposals. A redesigned Needs you inbox collects decisions waiting for the operator.

A quota failure should offer a usable next step

Studio 0.10.1, released October 18, changes what happens when a helper cannot continue because it has exhausted its quota. Its release notes describe a gate that explains the cause and offers reassignment to another available helper or stopping the job. It removes approval actions that merely retry the unavailable worker. The patch also fixes finished conversations opening short of their newest message.

Crew 0.14.1 followed the same day with a related fix: a quota-exhausted worker’s benched status survives a daemon restart, and an ask is not assigned to a benched worker. These are practical operator improvements, without a claim that agents have become better at coding.

What running it actually requires

The documented setup needs Node.js 22 or newer, a container runtime and wicked-garden 12.40.0 or newer. Garden supplies required worker skills and is installed separately, for example with npx wicked-installer install wicked-garden. After that, npx wicked-crew serve starts the local console at http://127.0.0.1:7701. The documentation covers macOS, Linux and Windows; containers can use Docker, Podman, Rancher Desktop, Colima or OrbStack.

Agents sign in through the Agents tab using the operator’s own accounts and plans. Provider allowances and charges still apply. The author says credentials stay in a broker while containers receive placeholders; model traffic still goes to the signed-in providers. Workers operate on real repository worktrees. This is a single-host, loopback service today, with remote execution explicitly unshipped.

The review story needs careful reading. The documented default requires a distinct reviewer for the work, while an explicit reduced-assurance mode permits self-review. A different CLI seat is not necessarily a different foundation model or an outside audit. The README also warns that shipped workflow validators can establish only that a worktree changed; task-specific correctness checks need to be authored and pinned. Some phase gates can approve on that limited deterministic floor when a distinct judge is skipped.

Per-tool input governance is documented for Claude; other workers rely on container restrictions and phase-boundary checks. Acceptance testing can be waived or explicitly skipped with a recorded reason. AITrending reviewed the English releases and documentation without installing the software or reproducing its enforcement claims. The appeal is the concrete supervision workflow and visible receipts, not a demonstrated guarantee of correct or safe code.