Remote MCP support?

[PfisterFactor]

Hey there - cool project.

For my use case I would love to run this on my homelab and then expose it as a remote MCP so I can use it from my phone. Does this project support that?

[impactjo]

Thanks, glad the idea resonates. Phone access is a scenario I find really compelling too.

What I use myself is more local first than a remote MCP setup would be. Claude Code runs on my homelab with the stdio MCP server, and I reach it from my phone. I have experimented with Claude Code Remote Control for this. Claude Code Channels (Telegram, Discord) is another variant for the same idea. In all of these Claude Code itself has to be running somewhere, the HM server stays stdio, and nothing needs to be exposed publicly. OpenClaw might become another alternative once they provide proper MCP support.

Remote MCP, meaning connecting from claude.ai or the Claude mobile app to an HTTP endpoint you host yourself, would reach the same scenario without a running Claude Code session in the middle. Home Memory only ships stdio today, so HTTP transport, auth (typically OAuth), TLS and session handling would all need to be added. No concrete plans on that right now, since the path above already covers the use case for my own setup.

Curious what you are picturing. Would the first path actually cover your scenario, or would it miss something? And if you lean toward remote MCP: is that a setup you already run, or still more speculative? Which phone client do you have in mind (Claude mobile app, claude.ai web, ChatGPT with MCP connectors, something else)? With any of those the tool call originates from the vendor's backend, so the MCP endpoint would have to be reachable from the public internet (TLS + OAuth, Cloudflare Tunnel, or similar).

[PfisterFactor]

I primarily use the Claude mobile app. This is a speculative setup, but I imagine it would be easier to connect as well as give the flexibility to connect it to other services so that they could ingest or query information about my home (home assistant comes to mind).

Claude Code remote control is an angle I didn't think about - and would probably solve my immediate usecase (accessing via phone). Though I would have to have Claude code running at all times on my server so I can remote into it via the app or Termux whenever I am working on the house and think of something to add to home-memory.

HTTP transport would also allow multiple services to connect to the same home-memory - as of now with just stdio you would have to use an external service in order to sync the database file across machines.

I can probably use something like this repo to implement it if I end up needing it - but I do think that this will probably be a very common ask of your project and makes sense to support it.

I'm gonna try getting this setup and access it via Claude Code remote control - that will probably satisfy my usecase for me for now!

[impactjo]

Appreciate the input, genuinely useful angles. Hope Remote Control works out well for your setup.

Thinking about it further, I'd probably split this into smaller steps than "stdio vs remote MCP":

1. Firebird Server support as an option alongside embedded. HM ships with Firebird Embedded today, which can run into file-lock issues under concurrent multi-session access. Pointing HM at a Firebird Server instance (DB on a NAS or file server) would solve the shared-DB and concurrent-access case on its own. Each user keeps their own stdio HM process, they just share the backend. More setup effort for the user, but it's the clean path to concurrent access.

2. HTTP/SSE transport on top of that, mainly for hosts that don't do stdio. HA is the concrete case. I'd probably front it with a static API key for LAN/VPN use. Big chat clients (claude.ai, mobile app, ChatGPT Connectors) expect OAuth, so those would still need step 3. Supergateway is a workable interim today if you want HTTP before HM ships it natively.

3. Full internet-facing with OAuth, TLS, and a proper deployment story. Considerably bigger, and I'd want to see strong demand before considering it further.

Nothing committed on timing. HM is still early in terms of real user feedback and I want to keep the scope honest.