Tidesman
A free native MCP server for running, understanding, and debugging Linux containers on macOS using Apple's container tool, with audit trail and configurable safety modes.
README
<p align="center"> <img src="assets/logo.svg" width="128" alt="Tidesman logo"> </p>
<h1 align="center">Tidesman</h1>
<p align="center"> A free native MCP server for running, understanding, and debugging Linux containers with Apple's <code>container</code> tool. </p>
<p align="center"> <img src="https://img.shields.io/github/v/release/JeronimoColon/tidesman-mcp?style=flat-square&label=release&color=1BAAC0" alt="Latest release"> <img src="https://img.shields.io/badge/dynamic/regex?url=https%3A%2F%2Fraw.githubusercontent.com%2FJeronimoColon%2Fhomebrew-tidesman%2Fmain%2FCasks%2Ftidesman.rb&search=version%20%22(%5B%5E%22%5D%2B)%22&replace=v%241&label=homebrew&style=flat-square&color=E08A2E" alt="Homebrew tap version"> <img src="https://img.shields.io/badge/macOS%2026-Apple%20Silicon-103E63?style=flat-square" alt="macOS 26 · Apple Silicon"> <img src="https://img.shields.io/badge/MCP-server-187E9E?style=flat-square" alt="MCP server"> <img src="https://img.shields.io/badge/license-EULA-475569?style=flat-square" alt="License: EULA"> </p>
<p align="center"> <a href="https://tidesman.dev">Website</a> · <a href="https://github.com/JeronimoColon/tidesman-mcp/releases">Downloads</a> </p>
It lets an AI assistant (Claude Desktop, Claude Code, or any MCP-compatible client) list, run, inspect, and clean up Linux containers on your Mac, safely and with a full audit trail.
This repository is the public download and documentation home for Tidesman. The source is maintained privately; the binaries published here are signed with an Apple Developer ID certificate and notarized by Apple.
Demo
<!-- A short screen recording is added here with the first release. --> <p align="center"><em>Demo coming with the first release.</em></p>
What it is
<p align="center"> <img src="assets/how-it-works.svg" width="860" alt="How Tidesman works: an AI client speaks MCP over stdio to Tidesman, which talks directly to Apple's container apiserver over XPC to run Linux containers in lightweight VMs."> </p>
MCP (Model Context Protocol) is an open standard that lets an AI assistant call external tools: small functions that do real work, such as listing or starting containers. Tidesman is the server side of that standard: it offers a set of container tools, and the AI is the client that calls them.
Apple's container is Apple's tool for running Linux containers on macOS. Normally you drive
it by typing commands in a terminal. Tidesman instead talks straight to the background service
that container relies on (its "apiserver"), using Apple's own Swift client library. Going
directly to that service is faster and sturdier than wrapping the command line.
Requirements
- An Apple Silicon Mac running macOS 26 (Apple's container runtime requires both).
- Apple's
containerinstalled and started:container system start.
Install
Tidesman is a signed, notarized binary. Pick whichever channel suits you.
Homebrew (recommended)
brew install JeronimoColon/tidesman/tidesman
This adds the tap and installs tidesman to /opt/homebrew/bin.
Installer package (.pkg)
Download tidesman-<version>.pkg from the
latest release, double-click
it, and follow the prompts. It installs tidesman to /usr/local/bin. The package is signed,
notarized, and stapled, so it verifies even offline.
Direct binary (zip)
Download tidesman-<version>-macos-arm64.zip, unzip it, and move tidesman onto your PATH.
Because the bare binary is not stapled, macOS checks it online the first time you run it.
Claude Desktop one-click (.mcpb)
Download tidesman-<version>.mcpb and open it with Claude Desktop (Settings → Extensions). The
bundle embeds the signed binary and registers the server for you in read-only mode.
Verify your download
Every release ships a SHA256SUMS file. From the folder holding your downloads:
shasum -c SHA256SUMS
Configure your MCP client
Tidesman runs as a local command your client launches over standard input/output. Point the
client at the binary and pass an access mode with --mode (omit it for read-only).
Claude Desktop, edit ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"tidesman": {
"command": "/opt/homebrew/bin/tidesman",
"args": ["--mode=safe"]
}
}
}
Claude Code:
claude mcp add --transport stdio --scope user tidesman \
-- /opt/homebrew/bin/tidesman --mode=safe
Any stdio MCP client (OpenAI Codex and others) works the same way: point it at the binary and pass the flags as arguments.
What it does
Twenty tools, each tagged by what it can do (Read, Write, or Destructive). Three report on or repair the engine itself, eleven act on containers, six on images:
| Tool | Capability | What it does |
|---|---|---|
system_ping |
Read | check the container service is reachable |
system_disk_usage |
Read | report disk use for images, containers, and volumes |
system_repair |
Write | re-download the engine's own infrastructure images |
container_list |
Read | list containers |
container_inspect |
Read | show a container's full details |
container_logs |
Read | fetch a container's recent output |
container_run |
Write | create and start a container from an image |
container_exec |
Write | run a command inside a running container |
container_start |
Write | start an existing stopped container |
container_restart |
Write | gracefully stop and then start a container |
container_stop |
Write | gracefully stop a running container |
container_kill |
Write | stop a container by sending it a signal (SIGKILL by default) |
container_delete |
Destructive | remove a container |
container_prune |
Destructive | remove every stopped container in one sweep |
image_list |
Read | list the images already downloaded |
image_inspect |
Read | show an image's full details |
image_pull |
Write | download an image from a registry |
image_tag |
Write | give an existing image another name |
image_delete |
Destructive | remove an image |
image_prune |
Destructive | remove untagged (or all unused) images in one sweep |
What the capability tags mean: a Read tool observes state and changes nothing. A Write tool changes state but never removes a resource; a stopped container survives a start or a restart and an engine repair only restores what the engine needs. Destructive is reserved for removing containers or images. Your MCP client may separately mark some write tools (exec, stop, kill, restart) as destructive, based on the MCP hints each tool carries: those tools can end work or change data inside a container, even though the container itself survives. In every mode, all twenty tools stay visible to your assistant; a tool the mode locks says so in its description and refuses until you raise the mode.
Access modes: safe by default
Tidesman runs at one of three authority levels, set with --mode=, and it starts in the
safest one. If no --mode is given, it is read-only. Every tool is always listed to your
assistant; the mode controls which of them may run.
| Mode | Allows | Tools callable |
|---|---|---|
read-only (default) |
Read | ping, disk usage, list, inspect, logs, and the two image reads (7) |
safe |
Read + Write | the above, plus run, exec, start, restart, stop, kill, image pull and tag, and the engine repair (16) |
full |
Read + Write + Destructive | all twenty, including the delete and prune tools |
A separate flag, --allow-host-mounts=/path/one[,/path/two], names the host folders
container_run may mount into a container; it is off by default because a host mount reaches
outside the container's isolation onto your real files. A mount is allowed only when its real
path (with symlinks resolved) sits under one of the folders you list.
Security and trust
- Signed and notarized. The binary is signed with an Apple Developer ID certificate and notarized by Apple, so Gatekeeper runs it without a warning.
- Safe by default. Unset mode is read-only; destructive operations require
full; host mounts require an explicit flag. - Audited. Every tool call is recorded to
~/Library/Logs/tidesman/audit.logand the macOS unified log: its name, its arguments (with secret-like values redacted), and its outcome. A destructive call also records exactly what it removed, and arguments a tool does not declare are recorded by name, so a mistaken call leaves a visible trace. - Honest tool annotations. Every tool declares MCP read-only and destructive hints that match what it can actually touch, so your client knows the stakes and can ask before anything risky runs.
- Guard your client config. Tidesman's tools cannot change the access mode; it is pinned by
the
--modeargument in your MCP client's configuration file. That file lives outside Tidesman's control, so protect it accordingly: a client that grants its assistant broad filesystem access would let the assistant edit its own mode. Every audit-log line records the mode the call ran under, so any change leaves a visible trail.
Privacy Policy
Tidesman collects nothing about you and sends nothing to us: no telemetry, no analytics, no crash reporting, no accounts. Everything it does happens on your Mac.
- The audit log (tool calls, arguments with secret-like values redacted, outcomes) is written
only to your Mac (
~/Library/Logs/tidesman/audit.logand the macOS unified log) and stays there unless you share it yourself. Delete it whenever you like. - Tidesman opens no network ports. Its only outbound connection is to a container registry (such as Docker Hub) when you or your assistant asks it to pull an image; that request shares the image name and your network address with the registry, like any container tool.
- Nothing is shared with, sold to, or retained by anyone else.
The full policy lives at tidesman.dev/privacy.html. Privacy questions: legal@tidesman.dev.
License
Tidesman is proprietary software, provided free of charge under the end-user license in EULA.txt. It bundles open-source components whose notices are preserved in THIRD-PARTY-LICENSES.
Links
- Website: https://tidesman.dev
- Report an issue: https://github.com/JeronimoColon/tidesman-mcp/issues
- Contact: hello@tidesman.dev
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。