--- title: Chat, models & runs description: Provider readiness, searchable model selection, run controls, follow-ups, approvals, and transcript tool-call groups. section: Workspace order: 3 slug: chat --- ## Connect a provider Verde checks the four GUI-chat providers—Codex, Claude Code, OpenCode, and Cursor—when the app opens. If none is ready, the **Connect an AI provider** screen reports one of these states for each provider: - **Ready** — the CLI is installed and authenticated. - **CLI not found** — install the provider and make sure its executable is on the `PATH` inherited by Verde. - **Sign-in needed** — run the provider's login flow, then return to Verde. - **Could not verify** — the readiness check failed or timed out. After installing or signing in, choose **Check again**. **Not now** dismisses the screen without changing provider configuration. Amp is not listed because it runs as a terminal TUI rather than a native chat provider. See [Provider setup](/docs/providers) for the exact install and login commands. ## Choose a provider and model Create a chat with `Ctrl+T`, then open the model pill in the composer. The picker has a provider rail, model search, default-model badges, and `Ctrl+1` through `Ctrl+9` shortcuts for the visible results. Before the first message is sent, selecting a model from another provider also switches the thread's provider. Once a thread has messages, its provider is locked so the transcript and provider session stay consistent; create another thread to use a different provider. Verde loads the available models from OpenCode, Claude Code, and Cursor at runtime. Codex models come from Verde's supported model list. A provider that is missing or signed out remains visible with a readiness hint instead of silently accepting a prompt it cannot send. ## Configure a run The **Run** pill groups the settings that affect the next agent run: - **Reasoning** appears only when the selected provider and model expose a reasoning-effort control. - **Speed** appears for Codex and for Cursor models that support fast mode. Choose **Default** or **Fast**. - **Access** is always available. **Supervised** asks before risky commands; **Full access** lets the agent run commands without asking first. Access is saved per thread. Reasoning and speed options can change when you switch models because not every provider exposes equivalent controls. ## Send, steer, and stop Press `Enter` to send the composer. While an agent is working, type another message and press `Tab`: Codex tries to steer the active turn, while Claude Code, OpenCode, and Cursor queue the message as the next turn. The pending message stays pinned above the composer. Double-click that pin to pull it back into an empty composer for editing, then press `Tab` to queue it again. **Stop** aborts the active run. Verde keeps the draft with its thread, so changing panes does not discard unfinished text. The same paths are scriptable with `verde live chat send`, `followup`, `stop`, and `draft`. See [CLI reference](/docs/cli#chat-control). ## Attach images and mention files Paste an image with `Ctrl+V` (`Cmd+V` on macOS) while the chat composer owns focus. Verde stages it above the composer instead of converting it to text. Paste again to attach multiple images, remove individual previews before sending, and click a transcript image to open its full-size preview. Codex, Claude Code, and OpenCode accept local image attachments. Cursor accepts them only when its ACP session advertises image support. Remote Herdr Codex GUI sends do not currently accept local image attachments; Verde reports the limitation instead of silently dropping them. To reference a workspace file, type `@` followed by part of its name or path. Verde searches the imported workspace, shows up to eight matching files, and inserts the selected relative path into the prompt. Use the arrow keys and `Enter`, or click a result. ## History and provider threads Each workspace ends with a **History · N** row in the expanded sidebar. It opens the command palette scoped to that workspace's saved chats, grouped into **Today**, **This week**, and **Older**. Search there to find a thread, press `Enter` to open it in an existing chat pane, or `Ctrl+Enter` to open it in a new pane. Press `Tab` on a result for thread actions. Right-click a workspace—or use `Ctrl+Shift+P`—to import an existing Codex, OpenCode, or Claude Code provider thread. The import dialog lists recent provider sessions and also accepts a provider thread ID. Cursor thread import is not supported yet. Imported and Verde-created provider threads can be refreshed with **Sync thread** when they have a provider thread ID and no request is running. The same thread menu can open a linked thread in its provider TUI, return a TUI thread to chat, or archive the thread. Archiving removes it from the active sidebar without deleting its persisted transcript. ## Approvals and input requests Providers can pause a run for approval. Verde renders the request in the transcript and lets you approve or deny it without leaving the chat pane. Codex MCP yes/no confirmations use this same approval surface. Structured MCP forms are not supported yet; Verde marks the input request unsupported and declines it rather than guessing an answer. **Full access** reduces command approvals, so use it only for workspaces and instructions you trust. ## Tool-call groups Consecutive tool calls collapse into a summary row showing the number of calls and their completed, failed, or running state. Click the row to inspect the individual calls. A group containing a failure opens by default so the error is not hidden. Choose the default under **Settings → Transcript → Tool call groups**: - **Collapsed** — start each group closed unless it failed. - **Expanded** — start each group open. - **Remember last** — reuse the last state you chose. The same preference is available as `transcript.tool_call_groups` in [`verde.json`](/docs/config#verdejson). ## Design feedback from the browser For UI work, Design Mode can select an element or region in the browser and send its DOM context—and, where supported, a screenshot—straight to a chat or terminal agent. See [Design Mode](/docs/design-mode).