signet

signet

A self-hosted MCP server for the Pebble Index 01 ring that lets you press the ring to capture voice notes, search journals and the web, and execute tools on your own infrastructure rather than someone else's cloud.

Category
访问服务器

README

signet

A self-hosted MCP server for the Pebble Index 01 ring.

Press the ring, talk. The phone transcribes it and picks a tool. When it picks one of signet's, the request goes to your own server instead of someone else's cloud.

A signet is a ring you press to authorize things.

Status

Working and deployed. Four tools (capture, ask, schedule, do), SQLite with full text search, scoped tokens, web search, and a mobile admin portal. Runs in Docker behind a Cloudflare Tunnel, verified against the real protocol the Pebble app speaks.

Not built yet: Google Calendar, so schedule saves the request and says so rather than dropping it. Also no approval queue, so anything marked destructive refuses to run. The plan is in docs/03-implementation-plan.md.

How it works

Four tools is the whole surface the phone sees, because a ~1B on-device model has to pick one and more choices make it worse. Behind them sits an unbounded set of internal capabilities.

  • capture saves what you said. No model call, and it still works when everything else is off.
  • ask searches your journal, falls back to the web, and answers in one or two sentences.
  • schedule is a calendar stub for now.
  • do is the catch-all. Rules run first, then a model classifies what is left.

Answers come back in the MCP response itself, which is what renders in the app feed and reaches the watch. Push notifications are a fallback for work too slow to answer inline, not the normal path.

Nothing you say is dropped. No API key, spent budget, provider down, unrecognised request, or unbuilt feature all end with the text in your journal and a plain answer saying so.

What we learned about the Pebble app

This is probably the useful part if you own an Index 01. The app lets you add your own MCP server, but there are no docs, and a few things will bite you. All of this was read out of coredevices/mobileapp and then verified against a running server.

The app speaks MCP 2025-06-18. It ships MCP Kotlin SDK 0.8.3, which tops out there. Current SDKs default to 2026-07-28, where the whole initialize handshake no longer exists. Both eras can be served from one endpoint. The Python SDK still supports the old handshake, so no shim is needed.

Return all your tools in one page. The client hits a literal TODO("Handle pagination") and throws if you send a nextCursor.

Serve single JSON responses, not SSE. Cloudflare Tunnel buffers SSE until the stream closes, so an SSE response is one the ring never receives. The Python SDK's Streamable HTTP transport frames POST responses as SSE by default, which passes every local test and then hangs behind the tunnel. Turn it off with json_response=True.

Auth is one header, used verbatim. No OAuth, no dynamic registration. You type a single Authorization value into the app, so it has to include the word Bearer. Cloudflare Access service tokens will not work, because they need two headers the app cannot send.

instructions from initialize gets injected into the on-device model's context. Free system prompt channel. Use it.

Results can render as native items in the app feed. Set _meta.coreSchema to the integer 1 and return structuredContent with both output and semanticResult. Both keys are required: the app reads them with getValue() and throws if either is missing. The discriminator is type, and unknown fields are rejected rather than ignored, so send exactly the fields the variant declares.

You get no session context. No user id, no device id, no history. Every call arrives cold, so the server has to keep its own state.

Run it

uv sync
export SIGNET_TOKEN=$(python -c "import secrets; print(secrets.token_urlsafe(48))")
uv run python -m signet

Then check it, including the parts the ring cares about:

uv run pytest
uv run python scripts/legacy_handshake.py http://127.0.0.1:8300 "$SIGNET_TOKEN"

legacy_handshake.py is stdlib only, so you can run it against a deployed server from any box with Python.

To point the phone at it, add an MCP server in the Pebble app with your URL ending in /mcp, streamable set to true, and the auth header Bearer <your token>.

Deploy

Docker Compose, one container, no database yet. See DEPLOY.md.

Docs

File What's in it
docs/00-research.md How the ring works and the exact MCP dialect its app speaks
docs/01-design-options.md Four architectures considered, and why one won
docs/02-architecture.md The design: inlets, router, capabilities, outlets
docs/03-implementation-plan.md Build order, with what has been verified so far

License

MIT

推荐服务器

Baidu Map

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。

官方
精选
JavaScript
Playwright MCP Server

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。

官方
精选
TypeScript
Magic Component Platform (MCP)

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。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

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。

官方
精选
本地
TypeScript
VeyraX

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。

官方
精选
本地
graphlit-mcp-server

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。

官方
精选
TypeScript
Kagi MCP Server

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。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

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。

官方
精选