aion-mcp
An MCP server that provides a set of tools for driving the AION AI assistant stack and managing a home fleet over SSH, including fleet control, Kali security tools, and memory management.
README
aion-mcp (mcpbuilder)
A Model Context Protocol server that gives an MCP client (Claude Desktop, Codex, etc.) a single set of tools for driving the AION assistant stack and the surrounding home fleet. It speaks MCP over stdio and is spawned on demand by the client — there is no long-running daemon to manage.
It is the client half of the aion-suite integration: aion-mcp calls
aion-core (Flask, 127.0.0.1:5000), which in turn talks to Ollama and Postgres.
MCP client (Claude Desktop / Codex)
│ stdio
▼
aion-mcp ──HTTP──▶ aion-core (Flask :5000) ──▶ Ollama (GPU) + Postgres
│
└──SSH──▶ fleet machines (draydev / ec2) + Kali container
Tool groups
Tools are organized into modules under src/, exposing ~45 tools total:
| Module | Tools | What it does |
|---|---|---|
aion-api-tools.ts |
aion_api_chat, aion_channels, aion_channel_history, aion_channel_presence, aion_activity, aion_memory_browse, aion_admin_*, aion_vast_*, aion_admin_network_run |
AION Core HTTP API: service chat, channels/presence, activity feed, fact-memory browse, admin, and Vast.ai GPU orchestration |
memory-tools.ts |
memory + conversation/profile tools | Read/write AION's SQLite-backed memory and conversation history |
chat-tools.ts |
chat/image/model tools | Direct chat, image generation, and model endpoints |
sonchat-tools.ts |
sonchat_* |
Bridge to the sonchat service |
portal-tools.ts |
portal_* |
Portal service endpoints |
fleet-tools.ts |
fleet_status, fleet_run, fleet_review |
Fan out agent runs across wsl / draydev / ec2 over SSH |
kali-tools.ts |
kali_* |
Run security tooling inside a Kali Docker container over SSH |
system-tools.ts |
system_status, … |
Local host/service status |
Build & run
npm install
npm run build # tsc → dist/
npm start # node dist/index.js (stdio server)
npm run dev # tsx src/index.ts (no build step)
npm run gateway # node dist/fleet-gateway.js (read-only fleet status over HTTP :5100)
Fleet gateway
fleet-gateway.ts is a tiny read-only HTTP service (localhost :5100) that exposes the same
fleet_status probe used by the MCP fleet_status tool, so a web page can read machine/agent
health even though the MCP server itself is stdio-only. It backs AION's /fleet topology page and chat control hook.
GET /health,GET /fleet/status— read-only, background-refreshed cache (fast page loads).POST /fleet/run,POST /fleet/review— guarded writes wrappingfleet_run/fleet_review.
Writes are gated: if FLEET_GATEWAY_TOKEN is set they require a matching x-fleet-token
header (the status probe stays open); set FLEET_GATEWAY_WRITE=off to disable writes
entirely. AION stages a confirmation before calling the write endpoints. Config:
FLEET_GATEWAY_PORT (5100), FLEET_GATEWAY_HOST (127.0.0.1), FLEET_GATEWAY_TTL_MS (120000),
FLEET_GATEWAY_TOKEN, FLEET_GATEWAY_WRITE.
The compiled dist/ is committed so the server runs via npx/node dist/index.js
without a build step on the client machine.
Configuration
All host-specific values and secrets come from the environment — nothing sensitive is
committed. .env and claude_desktop_config.json are git-ignored; there are no hardcoded
credential fallbacks (an unset secret disables the affected tool with a clear error rather
than falling back to a real password).
| Variable | Used by | Default | Notes |
|---|---|---|---|
AION_HOST / AION_PORT |
aion-api | 127.0.0.1 / 5000 |
AION Core Flask API |
AION_SERVICE_TOKEN |
aion_api_chat |
— | X-Aion-Service-Token for /api/service/chat; no user session needed. Required — chat is disabled if unset |
AION_SESSION_TOKEN |
session-scoped aion tools | — | Value of the aion_token cookie after logging in; required for channels/activity/memory/admin/vast tools |
KALI_HOST / KALI_USER / KALI_PASS / KALI_IMAGE |
kali | 192.168.0.200 / draygen / — / kali-custom:latest |
KALI_PASS required; kali tools return an error if unset |
FLEET_DRAYDEV_HOST / FLEET_DRAYDEV_USER / FLEET_DRAYDEV_PASS |
fleet | derived / draygen / — |
draydev SSH; if no password/SSH string is configured the machine is skipped with a clear error |
FLEET_EC2_SSH, FLEET_TIMEOUT_MS, FLEET_MAX_DEPTH |
fleet | — / 180000 / 2 |
ec2 SSH string; per-run timeout; recursion guard on delegated fleet runs |
Claude Desktop
Register the server in claude_desktop_config.json (kept local, not committed):
{
"mcpServers": {
"aion": {
"command": "node",
"args": ["/mnt/c/projects/mcpbuilder/dist/index.js"],
"env": {
"AION_SERVICE_TOKEN": "…",
"AION_SESSION_TOKEN": "…"
}
}
}
}
To get AION_SESSION_TOKEN: log in to AION at http://127.0.0.1:5000, open
DevTools → Application → Cookies, and copy the aion_token value.
Security notes
- No secrets in the repo:
.envandclaude_desktop_config.jsonare ignored and were scrubbed from history; committed source has no credential fallbacks. aion_memory_browseand all session-scoped tools requireAION_SESSION_TOKEN— they do not silently run unauthenticated.kali_*andfleet_*run commands on remote hosts over SSH and are intended for the author's own lab/home fleet only.
Status
Built and tested working end-to-end against a live AION stack — see
aion-suite/QA_MCP_AION_FLEET_REPORT.md
(45 tools registered, service + session endpoints passing, fleet fan-out verified).
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。
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。
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。
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。
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。
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。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
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。