How do I use a local LLM with Apple Health?
Syncky Health sends selected Apple Health data to software you run. A local model can work with that data through a client that starts Syncky's local stdio MCP server.
What do I get with a local model?
The iPhone seals Health messages for your agent → hosted Syncky Relay queues them → a separately running receiver decrypts them on your machine → local stdio MCP exposes tool output to your client → your selected model receives that tool output.
Relay sees delivery metadata and no health values, and emits plaintext hub notices. Sealed envelopes stay queued until acknowledgment or expiry, at most 7 days. Your agent, client and model can see plaintext. The Syncky agent never calls a model: it is model-agnostic at the MCP boundary. Tool-calling quality depends on the model and client you choose.
This route still uses Syncky Relay and is not an offline iPhone connection. A local MCP process does not select local inference. Your client can use a hosted model or other services; check its model, endpoint and other settings before sharing data. Selecting a local model does not establish privacy for all client traffic.
Agent delivery uses the Sync + Relay tier and background server delivery uses Sync, both unlocked free during launch. See launch access and pricing.
Which local clients can start the MCP server?
Jan
Checked on 12 November 2026.
Jan's MCP server documentation lists STDIO as a local process, with a command, arguments and environment variables. Its MCP guide describes enabling tool calling for local models and warns that model capabilities vary. Its model guide describes running local GGUF models through Llama.cpp on your own hardware. Choose a local model with tool-calling support and enable that capability as documented.
These sources establish documented capabilities. They do not establish a tested Syncky/Jan/model combination, reliable tool execution or privacy of the whole client.
How do I set it up?
These steps use commands available with Syncky Health app 0.1.3 or newer and npm agent 0.1.4. The agent reference has further background on homes, pairing and tools.
Install Syncky Health on your iPhone and Node.js 22.15 or newer on the computer that will run the receiver and MCP server. Install the client described above and a local model that supports tool calling.
Initialize only a fresh home. If you already have an initialized agent home, use it without initializing again. Run:
npx -y @syncky/agent init --name "My agent"Start pairing:
npx -y @syncky/agent pairThen open the link or scan the QR on the iPhone. Pick the matching code in Syncky Health, finish confirmation, choose the Health types to share and grant their Health permissions. Keep the pairing link, QR and code private. Scanning alone does not complete confirmation.
In a separate terminal, keep one receiver per home running:
npx -y @syncky/agent receive >/dev/nullFrom agent 0.1.4,
receiveprints no records to stdout by default; earlier agent releases print private records there. The redirect discards stdout either way. Keep setup output and agent files out of shared logs.In the client's MCP server settings, select local STDIO with command
npxand arguments-y,@syncky/agent,mcp. This launches:npx -y @syncky/agent mcpUse the same agent home for initialization, pairing, receiving and MCP. The default home belongs to your operating-system user; for a custom home, give every process the same
SYNCKY_AGENT_HOME.mcpstarts no receiver.Choose the intended local model in the client, enable its documented tool-calling capability and check that inference uses your local model or endpoint. Configuring the MCP process does not choose the inference endpoint.
How do I check the connection?
First ask the client for the tool inventory and only syncky_status, without requesting sample values. The inventory should show syncky_status, syncky_inbox, syncky_skill, syncky_send and syncky_ack. Status reports local pairing and transport metadata; it does not fetch new phone data.
Status, inbox and skill read local state; syncky_send and syncky_ack change state. A status-only request does not disable those tools. Review the operator security rules before allowing actions. Send only what the owner intends. This guide does not establish a per-tool disabling control for the named client.
Check pairing and receiver progress separately from whether the model successfully calls tools. A model failing to call a tool does not prove delivery has stopped.
Can I store data in my own database without MCP?
Yes. Syncky Health can send directly to a reachable server or HTTP receiver that you run. Your importer stores supported data in a database you choose on your computer. The receiver receives plaintext; make it reachable from the iPhone and use HTTPS. This direct route does not require MCP or Syncky Relay, and no agent pairing is needed.
A database alone is not an HTTP receiver. Syncky agent 0.1.4 has no Health database; you supply the receiver, importer and storage for the direct route. Database storage does not select or invoke a model.
Follow Apple Health webhook setup and the server contract for formats, authentication, signing and deduplication. For a snapshot instead of ongoing delivery, see exporting Apple Health data.
Why are tools or data missing?
- The client cannot start MCP: check that Node.js and
npxare on the client's executable path, and that the command and arguments match the setup above. A desktop app can have a different PATH from your terminal. - Tools exist but data is stale: check that the separate receiver is running in the same home, pairing is confirmed and the phone's Health selection and permissions include the intended types. iOS schedules background work; Health reads require an unlocked iPhone. Data can be hours old. Missing data does not mean zero.
- The model ignores tools: check the model's tool-calling support and enabled capabilities. This is separate from stdio startup and phone delivery.
- Inference goes elsewhere: check that you selected the local model or endpoint rather than a hosted provider, including any separate model used by the client for tool routing.
- Tool output fills the context:
syncky_inboxreturns raw JSON, up to 1000 records per call. Request a smalllimit, filter bytype, and pass the returnednext_after_seqas the nextafter_seq. Tool output and context size depend on the data requested. Never copy private inboxes into diagnostics or shared logs.
What else should I know?
Does Syncky call a model?
No. The agent transports and decrypts messages and exposes tools. Your client selects and calls the model. Start with agent options, Apple Health MCP options or the setup guides.
Does local MCP keep inference local?
Only selecting local inference in your client determines where the model runs. Local stdio describes the connection between your client and the MCP process. The client and selected model receive any plaintext tool output you request. Read the Privacy Policy and check your client's settings.
Can I run without Relay or MCP?
Use the direct-server route above to receive data without Relay, MCP or agent pairing, then maintain your own importer and database. The agent route described here still uses hosted Relay.
Can a local model write to Health?
Yes, for supported and allowed types with Health permissions on the iPhone. Use only owner-intended agent actions. Writes apply at the next sync when the app runs and the iPhone is unlocked; a queued write is not an applied write. Entries record their source. See writing to Apple Health for permissions, results and source marking.