Agents
Bring your own coding agent — run Claude Code, Codex, Gemini, Cursor, Hermes, or Pi inside Dices.
Dices is a bring-your-own-agent cockpit: every chat is driven by a real coding agent, and you choose which one. Claude Code runs built-in, and a growing list of other agents plug in over a shared protocol — so you can use your own Anthropic API key for Claude, or your own official CLI for the other agents, with the model and provider you prefer.
How agents connect
Dices talks to agents two ways:
- Claude Code runs in-process via the Anthropic Agent SDK. There is nothing extra to install — it ships with Dices.
- Every other agent connects over the Agent Client Protocol (ACP), a stdio JSON-RPC protocol. You install that agent’s command-line tool once, and Dices launches and talks to it as a subprocess.
This means the agents you can run are exactly the ones you’ve installed locally, plus the always-available Claude Code.
Note: Whichever agent you pick, the work happens on your machine against your files. Dices is the cockpit; the agent is the engine.
Claude Code (built-in)
Claude Code is the default agent and needs no install. It does need credentials — specifically, your own Anthropic API key:
- Anthropic API key — create one at console.anthropic.com (billed to you), then paste it in Settings. Dices uses it to authenticate the Agent SDK directly, and stores it encrypted at rest.
Note: Dices never handles your provider credentials beyond your own API key — Claude uses your Anthropic API key; other agents authenticate in their own official CLI.
See Models & reasoning effort for choosing which Claude model each chat runs.
Other agents (over ACP)
To use another agent, install its official vendor CLI and sign in there, then let Dices detect it. Open Settings → More agents: Dices scans for installed agent CLIs and lists the ones it finds. Each detected agent gets a button to start a chat with it. Dices connects to it over the open ACP protocol — it never handles that provider’s credentials.
| Agent | Install / launch | Notes |
|---|---|---|
| Claude Code | (built-in) | Agent SDK, in-process. Auth via your own Anthropic API key. |
| Codex | codex-acp | Connects over ACP once the CLI is on your PATH; sign in via the Codex CLI. |
| Gemini | gemini --acp | Requires the official Gemini CLI plus a GEMINI_API_KEY in your environment. |
| Cursor | cursor-agent --acp | Cursor’s agent over ACP; sign in via the Cursor CLI. |
| Hermes | pip install 'hermes-agent[acp]' | Installs the ACP extra for Hermes; sign in via its CLI. |
| Pi | pi-acp | Pi’s ACP launcher; sign in via its CLI. |
Note: For every non-Claude agent, you sign in inside that agent’s own official CLI — Dices just launches it. You are responsible for complying with each provider’s terms of service.
Tip: After installing a new CLI, reopen Settings → More agents so Dices re-scans. If an agent doesn’t appear, confirm its command runs from a normal terminal first.
Provider badges
Because chats from different agents live side by side in the sidebar, Dices shows a provider badge on each session that distinguishes non-Claude sessions (for example, a Gemini chat) from built-in Claude Code chats. This makes it easy to tell at a glance which agent is driving which conversation.
What differs between agents
The agents are interchangeable for everyday coding, but a few capabilities are Claude-specific because of how the protocols differ:
- Reasoning effort — ACP exposes no live “reasoning” knob, so the effort control applies to Claude only. ACP agents simply ignore it. See Models & reasoning effort.
- Model selection — available models depend on the agent you’ve connected; the per-chat model picker shows what that provider offers.
Related
- Models & reasoning effort — pick a model per chat and tune effort.
- Permissions — control how much each agent can do on its own.