Built-in Agent Tower

Choose View → Agent Tower to show or hide the dock. Drag its divider to resize it, or use Expand to make room for a longer conversation. Dock returns it to the side of the workspace.

When you narrow the window, the browser panel gives up space so the model and Tower remain visible. At very narrow widths, Tower fills the window. Your conversation keeps running through resizing, and widening the window restores your preferred panel widths.

Create separate named session tabs for different tasks. Rename a tab to describe the work, close it when you want it out of the way, and reopen it from the session list. Closing a tab does not delete its conversation.

Project mode works with the current document. Global mode keeps a conversation on the current computer; an explicit copy or move can bring it into a project.

Open Settings → Agents. Choose a supported subscription backend or configure an API/Ollama connection. Managed installation and repair show their progress and any error. Installation alone is not sign-in.

For Claude, complete the browser sign-in. If the browser shows an authentication code, paste it into Claude authentication code in Settings and choose Submit code. Wait for the verified connection state. Codex uses its provider sign-in flow. Refresh connection checks authentication and available models.

The model picker includes available connections and their discovered models. Thinking choices follow the selected model's capabilities.

Ask for a specific task and review the agent's questions and results. Current engineering automation covers Design, Drafting, eCAD, Mill, Plasma, Oxyfuel, Cartesian Laser, offline Fiber Galvo preparation, Router, and GcodePilot preparation/diagnostics. In Drafting, Tower can use compact native recipes for dependent part creation, authoritative body lookup, static clearance, and detached joint-motion studies; named inspection results appear in the same Browser rows as manual studies. Turn reports unavailable until native turning CAM exists. Physical machine control remains operator-owned.

Thinking summaries open during activity and collapse afterward. Ordinary tool cards stay compact; screenshot results open inline. Responses support Markdown, code, tables, links, and image references. Questions offer readable choices and a place for your own answer.

Long project conversations keep their full saved history while streaming updates and screenshots avoid repeatedly processing older messages. Closed tool details load when opened, and opening a long transcript fills older cards in batches so you can keep interacting with the workspace. Project status checks no longer reprocess the complete conversation, and geometry arriving from large rebuilds is retained in compact buffers instead of hidden full-message copies. This keeps long Tower modeling sessions bounded without deleting the transcript or scan/model source data. You do not need to clear the session to use this improvement. Ctrl/Cmd+S and Ctrl/Cmd+O continue to save or open a project while Tower has keyboard focus.

A green Send button starts a turn. During active work, the follow-up control is yellow to indicate steering; the red Stop control cancels the active turn.

Use Screenshot to capture and annotate the view, copy the result, then paste into Tower. You can also use + Images, paste, or drop multiple images. Remove unwanted previews before sending.

Project conversations accept PNG, JPEG, WebP, and GIF images: up to eight per message and 5 MiB per image. Screenshots returned by project tools are shown in the transcript. Saved project images remain accessible after reopening.

Claude and Codex subscription backends provide web search. API and Ollama backends do not have web search in this release.

Any backend with project archive tools can download a known public HTTP/HTTPS URL into the project. This supports images, PDFs, and other reference material, with source information retained. Downloads do not use your browser cookies or grant access to local/private network addresses.

Settings offers three shell-command choices:

• Disabled: shell commands cannot run.
• Ask every time: review the exact command before allowing it.
• Auto-approve: commands through Tower's shell tool run without another prompt.

Shell access is disabled by default and uses your computer's account permissions when enabled. A shared project cannot change that setting. Built-in project tools do not start physical machinery.

Claude and Codex retain accepted conversation context through Stop, errors, and reopening on the same computer. Tower reconnects to the current project tab automatically; you do not need to find or supply an internal document ID.

If a new provider session rejects the restored history as too large, Tower can summarize earlier history in separate batches and retry once before any work has executed. Completed recovery summaries are saved locally; interrupted recovery resumes from those completed batches instead of repeating the same work on every prompt. The workspace tool catalog also avoids repeated definitions to leave more room for project history. Progress remains visible, the pending message and its images are retained, and the saved transcript is unchanged. Stop also works during restoration.

Compact explicitly replaces older portable conversation content with a handoff summary after summarization succeeds. Clear asks before removing conversation history. Neither control deletes the design.

Undo, Redo, and same-project checkpoint restoration keep the agent's active turn and conversation. Opening a different project stops the old runtime and displays an interruption notice. A failed geometry calculation process can be replaced and the model rebuilt without treating that failure as successful modeling. If the whole document renderer stops, the tab offers Open recovery with the latest durable recovery time; it restores directly while other tabs stay available.

Use Connect an Agent if you prefer a separate external MCP client instead of the built-in Tower.