Direct Agent Launchers & Minimal-Impact Mode

Every supported AI coding tool has a first-class top-level subcommand in Agent Harbor:


ah <agent> [AH_OPTIONS...] [PROMPT] [-- NATIVE_AGENT_ARGS...]

When you run ah claude, ah codex, or ah agy interactively in your terminal without specifying a UI mode or workspace isolation strategy, Agent Harbor activates Minimal-Impact Mode.

What Minimal-Impact Mode Does Automatically

Native Interactive TUI (--agent-ui-style native): Your terminal attaches directly to the agent’s native interface via a transparent PTY proxy. Keyboard shortcuts, syntax highlighting, slash commands, and terminal colours behave identically to running the bare binary.
Local OS Sandbox (--sandbox-type local): File writes are confined to the current working directory (rw), /tmp, and the agent’s own state directory (such as ~/.claude or ~/.codex). The rest of your home directory—including ~/.ssh, ~/.aws, and browser profiles—remains inaccessible.
In-Place Working Copy (--working-copy in-place): Edits land directly in your current directory so your editor and Git status reflect changes immediately, while Agent Harbor records periodic background snapshots for diffing and rollback.
Live Anti-Halting Supervision: If the agent stalls mid-plan or asks an unnecessary confirmation question when instructions are already unambiguous, the local supervisor nudges it to verify its work and complete the task.
Remote Control & Cloud Sync (When Logged In): Once you have run ah login, every local ah <agent> session streams its live events.ahr log to the hub. You can inspect diffs, approve domain/sudo requests, or steer the session from agent-harbor.com, the Desktop App, ah tui, or your phone while your terminal stays attached.

Supported Agent Launchers

CLI Coding Agents

Command	Underlying Binary	Notes
`ah claude`	`claude`	Anthropic Claude Code CLI
`ah codex`	`codex`	OpenAI Codex CLI
`ah agy`	`gemini`	Google Gemini CLI
`ah grok`	`grok`	xAI Grok CLI
`ah copilot`	`copilot`	GitHub Copilot CLI
`ah amp`	`amp`	Sourcegraph Amp CLI
`ah cursor`	`cursor-agent`	Cursor headless terminal agent
`ah opencode`	`opencode`	OpenCode CLI
`ah qwen`	`qwen`	Alibaba Qwen Code CLI
`ah goose`	`goose`	Block Goose CLI
`ah agy`	`agy`	Google Antigravity CLI

GUI IDE Launchers

Agent Harbor also provides direct launchers for full desktop AI IDEs that run against an Agent Harbor workspace or Copy-on-Write branch:


# Launch the Cursor desktop IDE inside an isolated CoW branch
ah cursor-gui --branch feat/refactor-billing
 
# Launch the Google Antigravity desktop IDE
ah antigravity --branch exp/new-router

Passing Prompts & Native Agent Flags

Any positional argument that does not start with - is passed as the initial prompt to the agent. To pass tool-specific flags directly to the underlying agent binary, place them after -- (or pass unrecognised flags directly, which Agent Harbor forwards automatically):


# Start Claude Code with an initial prompt
ah claude "Fix the failing timezone tests in crates/core"
 
# Forward model and native flags to Claude Code
ah claude "Audit the authentication middleware" -- --model claude-opus-4-6 --verbose

Automatic Sandbox Bypass Mirroring

When you pass an agent’s native permission-bypass flag, Agent Harbor honours that intent and automatically disables its outer OS sandbox (--no-sandbox) unless you also explicitly pass --sandbox-type:

Agent	Native Bypass Flag	Implied Agent Harbor Behaviour
`claude`	`--dangerously-skip-permissions`	Implies `--no-sandbox`
`codex`	`--dangerously-bypass-approvals-and-sandbox` / `--yolo`	Implies `--no-sandbox`
`gemini` / `qwen`	`--yolo` / `--approval-mode=yolo`	Implies `--no-sandbox`
`amp`	`--dangerously-allow-all`	Implies `--no-sandbox`
`cursor`	`--force` / `-f`	Implies `--no-sandbox`
`copilot`	`--allow-all` / `--yolo`	Implies `--no-sandbox`

If you want the inner agent to skip interactive confirmation prompts while remaining strictly confined inside an Agent Harbor sandbox or CoW branch, pass both --sandbox-type local (or devcontainer / vm) and the native bypass flag: bash ah claude --sandbox-type local --branch exp/fast-iteration -- --dangerously-skip-permissions

Upgrading a Session with Opt-In Flags

You can layer isolation, branching, roles, extensions, and remote execution onto any ah <agent> command:

1. Run in a Copy-on-Write Branch (`--branch` or `--working-copy cow`)

Keep your main working tree untouched while the agent works in an instant CoW clone:


ah codex --branch feat/async-pool "Migrate database pool to deadpool-postgres"

2. Assign a Role and Portable Extensions (`--role`, `--with`)


ah claude \
  --role security-auditor \
  --with gh:org/sec-skills \
  --with npm:@modelcontextprotocol/server-postgres \
  "Review the OAuth callback handlers for open redirect vulnerabilities"

3. Background & Headless Execution (`--detach`, `--non-interactive`)

Launch an agent in the background and return its session ID immediately, or start an idle headless agent and send prompts to it programmatically:


# Launch detached with an initial prompt
SESSION_ID=$(ah claude --detach --branch fix/issue-104 "Fix issue #104 and run cargo test")
 
# Or start an empty headless session and send prompts later
SESSION_ID=$(ah codex --non-interactive --detach)
ah session prompt "$SESSION_ID" "Analyse the benchmark regression in parser.rs"
ah session view "$SESSION_ID" --follow

4. Dispatch to a Remote Executor (`--executor`, `--requires`)

Run the session on another machine in your fleet or auto-provision a cloud VM:


# Run on a named workstation
ah claude --executor home-pc "Run the CUDA kernel benchmarks and optimise occupancy"
 
# Run on any fleet host matching capability tags
ah codex --requires os=linux,gpu=nvidia "Verify TensorRT build"