uvwt/agentdock

Secure MCP runtime for AI agents to operate local machines, servers, and containers with multi-device orchestration.

AgentDock MCP – What It Is

AgentDock is a runtime that lets large‑language‑model (LLM) chat agents (ChatGPT, Claude, Grok, etc.) safely act on real computers, servers, containers, or browsers. It does not provide a chat UI or run any model itself; instead it exposes a MCP (Multi‑Client Protocol) HTTP endpoint that AI clients can call to perform file operations, run shell commands, use Git, automate browsers, and invoke “Skills” (plug‑in tools). Multiple AgentDock instances can be linked so a single conversation can orchestrate work across many machines.

Core Capabilities (as described in the README)

Area	What AgentDock Provides
File & command access	Read/search UTF‑8 files, safe atomic writes, directory traversal limits, command execution with timeout, separate stdout/stderr, PTY support, long‑running sessions, output truncation and redaction.
Skills & dynamic MCP	Built‑in and community‑provided Skills (via `uvwt/agentdock-skills`), install/update/remove with digest‑based atomic swaps, isolated env vars, dynamic MCP server registration, schema inspection.
Native ACP	Optional adapter for OpenAI Codex, Anthropic Claude, Grok etc.; creates ACP sessions, runs prompts, handles permission requests, respects host security boundaries.
Browser & desktop automation	Launch/close browsers, navigate, click, type, wait, inspect page text/elements/network, persist login state, capture screenshots; also macOS desktop automation.
Recoverable tasks	Persistent task state, explicit goals/steps, checkpoints, blocker tracking, resume after interruption, final evidence‑based review, reusable workflow templates.
Cross‑device coordination	Connect multiple AgentDock nodes (local, LAN host, cloud VPS) via MCP; optional NexusDock integration for long‑term memory, runbooks, shared notes, artifact proxying.

Typical Use‑Case Flow

Install AgentDock on each machine (Docker, Linux, macOS, Windows, or container). The default MCP endpoint is http://127.0.0.1:8765/mcp.
Configure authentication – a Bearer token or OAuth is required for any public‑facing deployment.
Add the MCP URL & token to the LLM client’s tool configuration (e.g., in ChatGPT’s “Tools” settings).
From a chat, the LLM issues tool calls such as:
- read_file, write_file, run_command to edit code or run tests on the target host.
- git_pull, git_push to manage repositories.
- browser.navigate or desktop.click to interact with web pages or macOS apps.
AgentDock executes the request inside the host’s security sandbox, returns structured JSON with results, logs, and any required follow‑up prompts.
Multiple nodes can be addressed in the same conversation, enabling a single AI to coordinate a workflow that spans a developer’s laptop, a CI server, and a production VPS.

Installation & Getting Started (summary)

Docker – pull agentdockio/agentdock and run; see docs for port and token setup.
Linux – automated script or manual systemd service.
macOS / Windows – graphical installers.
After install, retrieve the MCP URL and auth token from the control panel or terminal and plug them into the AI client.

Who Might Use It?

Developers who want LLMs to edit code, run tests, or deploy directly on their machines without giving the model unrestricted shell access.
Ops / SRE teams looking to let an AI agent manage servers, containers, or reverse proxies under strict permission boundaries.
Product teams building custom AI‑driven assistants that need to perform real‑world actions (file edits, browser automation) safely.

Project Status & Community

Actively maintained (CI badge, regular releases, Docker pulls).
Open‑source under Apache 2.0.
Documentation hosted at uvwt.github.io/agentdock-docs.
Community chat via a QQ group; contributions accepted through GitHub Issues/PRs.

TL;DR

AgentDock is a secure runtime that bridges LLM chat agents and actual machines. It exposes a well‑defined HTTP‑based protocol for file manipulation, command execution, Git, browser/desktop automation, and extensible “Skills”, allowing a single AI conversation to orchestrate complex, multi‑machine workflows while keeping strict access controls.