Supported AI Agents
ServerCC runs seven coding agents on your own server over SSH — from your iPhone or iPad. Version 1.8.0 added Grok Build, OpenCode, and Kimi Code alongside Claude Code and Codex; 1.8.1 adds Pi and OMP (oh-my-pi), with the same core workflow for every agent: install detection, one-tap launch, session history with exact resume, permission modes, and persistent tmux sessions.
| Agent | CLI | Permission modes | Notes |
|---|---|---|---|
| Claude Code | claude |
Default · AcceptEdits · Auto · Bypass | Full support, incl. Agent View, bg sessions, Skills, CLAUDE.md editing |
| Codex | codex |
Sandbox & approval modes | Skills & MCP browsers, experimental-feature toggles |
| Grok Build | grok |
Default · AcceptEdits · Auto · DontAsk · Bypass · Plan | xAI’s official CLI; sign-in card; official-CLI verification |
| OpenCode | opencode |
Auto-approve toggle | Open-source, provider-agnostic; worktree-aware history |
| Kimi Code | kimi |
Default · YOLO · Auto · Plan | Moonshot AI’s CLI; sign-in card; initial prompt not supported by its TUI |
| Pi | pi |
— (no permission prompts) | Mario Zechner’s minimal agent; installs via npm (Node ≥ 22.19) |
| OMP | omp |
YOLO toggle | oh-my-pi — extended Pi fork, single prebuilt binary; can reuse Claude Code/Cursor credentials |
Enable Agents
Pick the agents you use under Preferences → Agents. The default is Claude Code + Codex, so existing setups are unchanged; at least one agent always stays enabled. Server and workspace pages only render sections for enabled agents, and health checks only probe enabled agents — so agents you don’t use cost nothing.
Disabling an agent hides its sections and history but deletes nothing — re-enable it and cached sessions reappear. An agent with a live persistent or parallel session keeps its section visible even while disabled, so you can always get back to running work.
Install & Health Checks
The server detail page shows an install section per enabled agent: detected version, one-tap Install/Upgrade using each vendor’s official installer, and an optional custom binary path. Pi is the one exception on the installer front: its official installer is interactive, so ServerCC installs it via npm instead (Node ≥ 22.19 required). Detection also looks in the common install locations (~/.local/bin, ~/.grok/bin, ~/.opencode/bin, ~/.kimi-code/bin, ~/.npm-global/bin, ~/.bun/bin) — so a freshly installed CLI is found even when its PATH line only lives in an interactive-shell rc file.
The community "superagent" grok-cli installs a binary that is also called grok. ServerCC verifies the version output of the official xAI CLI and treats the community CLI as not installed.
Sessions: History, Resume, Reconnect
The workspace landing page shows a history tab per enabled agent. Tap any row to resume that exact conversation — ServerCC resolves each agent’s own session store (Grok Build’s on-disk session files, OpenCode’s CLI session list, Kimi Code’s session index, Pi and OMP’s per-project session files) and resumes by session ID, not "most recent". After a connection drop, non-persistent sessions continue the workspace’s latest conversation; tmux-backed persistent sessions reattach exactly and also work with all seven agents.
OpenCode’s history is project-scoped and includes sessions started in worktrees. One limitation: session history for Grok Build, Kimi Code, Pi, and OMP is read from their default home directories — a custom GROK_HOME/KIMI_CODE_HOME/PI_CONFIG_DIR/OMP_PROFILE exported only in shell rc files is not picked up.
Permission Modes
Each workspace remembers a launch permission mode per agent, set in workspace settings. Grok Build supports Default, AcceptEdits, Auto, DontAsk, Bypass, and Plan; Kimi Code supports Default, YOLO (auto-approves regular tools, still asks for sensitive files), Auto (fully unattended), and Plan; OpenCode has an auto-approve toggle, and OMP a YOLO toggle. Pi has no permission system: it always runs tools directly, so there is nothing to configure — for risky work, prefer a worktree. Unattended modes show a risk confirmation before they can be enabled.
Signing In from Your Phone
When an agent prints a sign-in URL in the terminal — Grok Build’s device-auth flow (x.ai), OpenCode’s OpenAI sign-in, or Kimi Code’s device-code flow — ServerCC detects it and shows a login card: tap to open the URL in the browser, finish there, and return to the session. Since 1.8.1 this works for Grok Build’s TUI too, which shows only a pairing code — ServerCC reconstructs the device-auth URL from it. Pi and OMP sign in inside their TUIs via /login (an interactive picker with no printed URL, so no card); OMP can also reuse Claude Code or Cursor credentials already on the server. Note that OpenCode removed Anthropic OAuth in 1.3.0 — use an API key or another provider with OpenCode.
MCP & Skills
The MCP server browser works with five of the seven agents — Claude Code, Codex, Grok Build, OpenCode, and Kimi Code. ServerCC reads and writes each agent’s own config format (Grok Build’s config.toml, OpenCode’s opencode.json, Kimi Code’s mcp.json) and preserves fields it doesn’t know about. Pi has no native MCP support, and OMP’s MCP config isn’t wired up yet. The Skills browser currently supports Claude Code and Codex.
What Stays Claude Code-Only
- Agent View & bg sessions — the supervisor-managed parallel dashboard has no equivalent in the other CLIs
- CLAUDE.md editing from the workspace page
- macOS Keychain unlock on session start
FAQ
Can I run OpenCode on an iPhone?
Yes. OpenCode runs on your own server; ServerCC connects over SSH from your iPhone or iPad and gives you a full terminal, session history with exact resume, an auto-approve toggle, and persistent tmux sessions for OpenCode.
Does ServerCC work with the community grok-cli?
ServerCC supports the official xAI Grok CLI (Grok Build). The community "superagent" grok-cli installs a binary with the same name — ServerCC verifies the version output and treats the community CLI as not installed.
Can OpenCode sign in with my Claude (Anthropic) subscription?
No. OpenCode removed Anthropic OAuth sign-in in version 1.3.0 at Anthropic’s request. Use an API key or another provider account with OpenCode — or run Claude Code itself, which ServerCC fully supports.
Why is my initial prompt ignored when starting Kimi Code?
Kimi Code’s interactive TUI has no initial-prompt flag, so ServerCC starts the session without pre-filling the prompt — type or paste it once the TUI is up. History, exact resume, and permission modes all work normally.
What is the difference between Pi and OMP?
Pi is Mario Zechner’s minimal coding agent: no permission prompts, installed via npm. OMP (oh-my-pi) is a community fork that ships as a single prebuilt binary and adds plans, background jobs, about sixty slash commands, a YOLO auto-approve switch, and reuse of existing Claude Code or Cursor credentials. ServerCC supports both since 1.8.1.
Why is there no permission mode setting for Pi?
Pi has no permission system — it always runs tools directly, so there is nothing to configure. For risky changes, run it in a worktree. OMP, its fork, does have a YOLO auto-approve switch.
See also: Quick Start to add your first server, Sessions for how history and resume work in detail, and Terminal for quick commands and multi-line input.