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.
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.
capturesaves what you said. No model call, and it still works when everything else is off.asksearches your journal, falls back to the web, and answers in one or two sentences.scheduleis a calendar stub for now.dois 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
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.