AI agents get SSH-like reach, but with guardrails and an off switch

What it does

AgentDock is a standalone MCP runtime that acts as a bridge between AI clients like ChatGPT or Claude and actual computing environments. It exposes tools for file operations, shell commands, Git, browser automation, and long-running tasks across macOS, Linux, Windows, and containers. You connect multiple instances to a single conversation to coordinate work across your laptop, a LAN host, and a cloud VPS without switching contexts.

The interesting bit

The project treats permission boundaries as a first-class feature rather than an afterthought: atomic file writes, path restrictions, private-directory protection, environment variable filtering, and output redaction are all built in. It also persists task state with checkpoints, so an agent can resume after a network hiccup or a reboot instead of starting from scratch.

Key highlights

• Orchestrates multiple machines from one AI conversation via MCP Streamable HTTP.
• Isolates Skills and dynamic MCP servers in their own environments with separate variables and runtimes.
• Supports browser and macOS desktop automation, including authenticated sessions and screenshots.
• Offers recoverable tasks with staged checkpoints and evidence-based completion checks.
• Optionally pairs with NexusDock for long-term memory, runbooks, and multi-device state coordination.

Caveats

• Native ACP working directories follow the host or container security boundary, bypassing AgentDock’s filesystem allowlist.

Verdict

Worth a look if you want an AI agent to actually manage infrastructure, edit code, or run deployments on real hardware without giving it the keys to the kingdom. Skip it if you are looking for a chat interface or model inference—AgentDock does neither.