Blog/Grok Build, OpenCode & Kimi Code

ServerCC 1.8.0: Grok Build, OpenCode, and Kimi Code on your iPhone

2026-07-26

1.8.0 is on the App Store. It adds three coding agents: Grok Build (xAI’s CLI coding agent), OpenCode (open source), and Kimi Code (from Moonshot AI). With Claude Code and Codex, that makes five. Support works the same as for the first two: a terminal over SSH, session history, exact resume, and tmux persistence.

The three new agents

Grok Build is xAI’s official CLI agent, the grok command. ServerCC verifies it is the official CLI (the community grok-cli shares the binary name and is treated as not installed), reads its on-disk session store for history, and detects its device-auth sign-in URL so you can finish login from the phone.

OpenCode is an open-source coding agent that isn’t tied to one model provider: OpenAI, Google, or local models all work. ServerCC lists its sessions through the official CLI, filters them to the workspace (including sessions started in worktrees), and resumes any of them by exact session ID. OpenCode removed Anthropic OAuth sign-in in 1.3.0, so use it with an API key or another provider, or run Claude Code instead.

Kimi Code is Moonshot AI’s coding agent, the kimi command. History comes from its session index, resume is exact, and its device-code sign-in flow gets a login card too. Its TUI has no initial-prompt flag, so ServerCC starts the session and you type the first prompt in the TUI. Everything else works the same as with the other agents.

What all five agents get

All five agents go through the same flow ServerCC uses for Claude Code:

  • Choose which agents appear. Enable agents in Preferences → Agents (default stays Claude Code + Codex). Pages only render enabled agents, and health checks only probe them.
  • Install from the phone. The server page detects each CLI and its version, installs or upgrades with the official installer in one tap, and accepts custom binary paths.
  • Exact resume. Per-agent history tabs in every workspace. Tapping a row resumes that specific session, not “whatever ran last in this directory”.
  • Permission modes at launch. Grok Build: Default / AcceptEdits / Auto / DontAsk / Bypass / Plan. Kimi Code: Default / YOLO / Auto / Plan. OpenCode: an auto-approve toggle. Saved per workspace, with a confirmation before enabling unattended modes.
  • Sign in from the phone. When a CLI prints its auth URL (x.ai device auth, OpenAI sign-in, Kimi’s device code), ServerCC turns it into a tappable login card.
  • MCP server browser. Reads and writes each agent’s own config format (TOML for Grok Build, JSON for OpenCode and Kimi Code) and keeps fields it doesn’t recognize.
  • tmux persistence and notifications. Persistent sessions, reconnect after drops, and completion notifications work the same for all five.

The full feature matrix, including per-agent caveats, is in the Supported AI Agents docs.

What stays Claude Code-only

Some features depend on capabilities only Claude Code has: Agent View and supervisor-managed bg sessions, CLAUDE.md editing, and macOS Keychain unlock stay Claude-only. The Skills browser covers Claude Code and Codex for now; the other agents will be added once their skills layouts are verified.

Also in 1.8.0: parallel sessions on the current branch

A new per-workspace setting lets the current branch run any number of sessions at once, instead of one persistent session. Use it to open a second session on the same checkout without creating a worktree (each worktree still runs one session). It’s off by default and explains the trade-offs when you turn it on. For parallel work, use it together with persistent tmux sessions.

Update: since 2.1.0 this is no longer a setting. Any branch can run several sessions at once.

Get started

Update to 1.8.0, open Preferences → Agents, and switch on the agents you use. If you’re new to ServerCC, the Quick Start covers adding a server and starting a session, and Claude Code on iPhone walks through the full mobile workflow.