> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tempestai.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Chat Pane

> Headless chat with hosted LLMs or CLI agents from inside the workspace canvas. No terminal, no TUI.

# Chat Pane

The chat pane is a canvas node that renders a headless conversation with an LLM or an agent CLI — no terminal, no TUI. It sits alongside terminal-based agent sessions on the workspace canvas and is meant for quick asks, planning, and one-shot generations without the ceremony of a full agent tab.

Add one from the canvas node picker (**+ → Chat**), or convert a selection into a chat via the canvas context menu.

## Three backends, one UI

Every chat node picks one of three backends via the model dropdown. The choice is per-node:

| Backend                 | What it uses                                                | Auth                                        |
| ----------------------- | ----------------------------------------------------------- | ------------------------------------------- |
| **API** (BYOK)          | Provider SDK called from the Rust backend                   | [OS credential manager](/settings/api-keys) |
| **CLI Agent**           | The vendor CLI, driven headlessly by a per-turn Node bridge | Whatever the CLI itself is logged into      |
| **Warp** (experimental) | The `warpllm` Rust gateway, single-round-trip               | BYOK slot or provider env var               |

Streaming, message rendering, tool-call display, and copy-to-clipboard behave the same regardless of backend.

## Model picker

Clicking the model chip opens a portal picker with two sections:

* **CLI Agents** — grouped by agent (Claude Code, Codex, Gemini CLI, Opencode). Each agent has a shortlist of its own models. When the experimental Warp toggle is on, Warp shows up here too with a `provider/model` list (OpenAI, DeepSeek, Kimi, OpenRouter).
* **BYOK providers** — Anthropic, OpenAI, Google Gemini, Mistral, DeepSeek, xAI, Groq, OpenRouter, Ollama, LM Studio — each with a shortlist. The active provider is remembered across sessions.

A search box across the picker filters both sections at once.

## CLI agent bridges

The four CLI-agent backends (Claude Code, Codex, Gemini CLI, Opencode) all run through per-turn Node sidecars shipped inside the app under `src-tauri/resources/{agent}-bridge/bridge.mjs`. The Rust backend spawns the sidecar on `claude_stream_start`, streams NDJSON back over a `claude://{streamId}` Tauri channel, and the frontend maps it to typed chat events.

Under the hood each bridge shells the vendor CLI in a specific non-interactive mode:

* **Claude Code** — Anthropic's Claude Code CLI in stream mode
* **Codex** — `codex exec --json --sandbox workspace-write` (no per-tool approvals)
* **Gemini** — `gemini --experimental-acp` speaking JSON-RPC ACP (per-tool permission prompts surface as chat UI)
* **Opencode** — `opencode serve` on a free local port, consuming its SSE stream and permission POSTs

**Auth for CLI bridges is whatever the CLI already has on your machine.** Tempest does not hand a BYOK key to the CLI bridges — if you can `claude` in your terminal, you can chat with Claude Code in the pane.

## Warp (experimental)

Warp is an opt-in, single-round-trip gateway that routes any `provider/model` string through a Rust `warp_chat` command. Enable it in **Settings → Experimental → warpllm**. The Agents section of the model picker then shows Warp with its own model list.

Warp reads the API key for the selected provider from either:

1. The matching [BYOK slot](/settings/api-keys), or
2. An environment variable — `OPENAI_API_KEY`, `DEEPSEEK_API_KEY`, `MOONSHOT_API_KEY` (for Kimi), or `OPENROUTER_API_KEY`

Warp responses are not streamed; you'll see the whole reply arrive at once.

## Chat pane vs terminal sessions

|             | Chat pane                                            | Terminal session                                    |
| ----------- | ---------------------------------------------------- | --------------------------------------------------- |
| Interface   | Headless message stream                              | Full PTY / TUI                                      |
| Backend     | API / CLI bridge / Warp                              | Vendor CLI in a real shell                          |
| Persistence | Canvas node state                                    | Session store + PTY resume                          |
| Tool calls  | Rendered inline as event cards                       | Whatever the CLI's UI shows                         |
| Best for    | Planning, quick asks, embedding into a larger canvas | Long agent runs, file edits, tool-heavy tasks       |
| Isolation   | Same process as the app                              | Hephaestus sandbox (see [Agents](/agents/overview)) |

Both surfaces are first-class. A common workflow is to plan and iterate in a chat node, then hand the final prompt off to a terminal agent session in a fresh worktree.

## Related settings

* **Settings → API Keys** — BYOK slots consumed by both the API backend and Warp. See [API Keys](/settings/api-keys).
* **Settings → Experimental → warpllm** — reveals the Warp backend in the picker.