cware-hil-mcp
Human-in-the-loop MCP hub for Claude Code agents, enabling agents to ask questions, request approval, or report status to a human via an Obsidian plugin.
README
cware-hil-mcp
Human-in-the-loop MCP hub for Claude Code agents, answered in Obsidian.
Agents call MCP tools to ask a human a question, request approval, or report status. This standalone hub holds the queue and exposes an MCP endpoint; the Obsidian plugin is the human's UI.
Claude Code agents ──HTTP /mcp──▶ hub (Node, :22360) ◀──ws /bridge── Obsidian plugin
└ SQLite queue (survives restarts)
Shared domain types + the bridge protocol live in cware-hil-lib (a git dependency).
Build
npm install
npm run build
Requires Node ≥ 22 (developed on Node 26; the hub uses the built-in node:sqlite,
so there is no native build step).
Run
node dist/cli.js start # or: cc-hitl start (once linked/installed)
node dist/cli.js status # health check
node dist/cli.js stop
node dist/cli.js token # print the bearer token
State lives in ~/.cc-hitl/ (hub.json holds the token; hub.db is the queue).
The hub binds to 127.0.0.1 by default; override with --host <addr> or
CC_HITL_HOST (e.g. 0.0.0.0 to accept connections from other machines).
This container serves no UI — only /mcp, /bridge, and /health (and a plain
pointer at /). The web UI is a separate deployable,
cware-hil-ui, that you point at this hub
for a setup page + live dashboard. The CLI also prints the connection snippets on
start (and via cc-hitl setup).
Docker
export CC_HITL_TOKEN=$(openssl rand -hex 24)
docker compose up -d --build
docker compose exec hub node dist/cli.js status
The container binds 0.0.0.0 and compose publishes it on 127.0.0.1:22360.
State persists in the hub-data volume (CC_HITL_HOME=/data). Without compose:
docker build -t cware-hil-mcp .
docker run -d --name cc-hitl -p 127.0.0.1:22360:22360 \
-e CC_HITL_TOKEN=$CC_HITL_TOKEN -v cc-hitl-data:/data cware-hil-mcp
Exposing beyond loopback: once published on a LAN/public interface, the bearer
token is the only thing protecting /mcp and /bridge. Use a strong
CC_HITL_TOKEN and put a TLS reverse proxy in front.
Prebuilt images
CI publishes images to GitHub Container Registry on every push to main:
docker pull ghcr.io/kiliansen/cware-hil-mcp:latest # newest build
docker pull ghcr.io/kiliansen/cware-hil-mcp:0.1 # latest 0.1.x
docker pull ghcr.io/kiliansen/cware-hil-mcp:0.1.1 # exact version
Versioning is driven by commit messages (Conventional Commits):
fix: bumps the patch, feat: the minor, and feat!: / a BREAKING CHANGE:
footer the major. A commit with no conventional prefix bumps the patch. Each push
tags the repo (vX.Y.Z) and pushes matching X.Y.Z, X.Y, X, and latest
image tags. See .github/workflows/docker.yml.
Connect Claude Code
cc-hitl start prints the exact command (or use the cware-hil-ui setup page). In your project:
claude mcp add --transport http --scope project hitl http://127.0.0.1:22360/mcp \
--header "Authorization: Bearer <token>"
Set a long per-server timeout (the printed snippet uses 24h) because the ask
tools block until a human answers.
Tools agents get
ask_user(prompt, title?, agent_label?)— open question, returns typed text. Blocks.ask_choice(prompt, choices[], multi?, …)— pick one/several (+ optional note). Blocks.request_approval(title, body, diff?)— approve/reject + comment. Blocks.notify(message, level?)— fire-and-forget notice. Non-blocking.update_status(label?, status?, current_task?, progress?, done?)— dashboard feed. Non-blocking.
Test
npm test # builds, then runs the integration + resilience suites
These spawn a real hub on a temp CC_HITL_HOME and drive each tool through an
MCP client + bridge WebSocket client.
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。