How do you let Claude/ChatGPT reach self-hosted apps without exposing them? (Synology, Tailscale)
retained text
Hi all. Looking for advice from people who've actually done this. **My setup** * Synology NAS - apps in Container Manager * A few personal apps - self-developed app for health management, Firefly III, SparkyFitness, and a couple more under development * Everything is reachable only over Tailscale, behind a small Tailscale "front door" container. Nothing is port-forwarded, DSM firewall is deny-by-default * I travel a lot (work and personal), so it all has to be accessible and manageable remotely **What I want** Ask questions about my own data from a normal Claude or ChatGPT chat on my phone, and eventually make changes, but only after I confirm them. The reason for this is because I have set it up in that way that I can use standard chat windows, which let me take advantage of AI capabilities including submitting receipts via chat windows (in addition to an inbox function), taking pictures of food items and having them uploaded to SparkyFitness, running skills and scripts that analyse, categorise, and file and store financial transactions and data, and some limited integration between SparkyFitness and my personal health management app. It also allowed me to keep everything behind tailscale without exposing anything to the internet. **The problem** Until last week, I was able to use standard chats to reach the apps on my NAS via Claude desktop app on an always-on home laptop. That link recently became unreliable for reasons I still don't understand (but still testing). It's strange because new chats can't connect, however the old ones still can. Because the connectors in Claude AI and ChatGPT are called from the vendor's cloud they can't join my tailnet. So it seems the only lasting option is a public HTTPS endpoint with OAuth in front of MCP servers for each app. I am leery about exposing services to the internet, but I can't see another way to solve this. That would let me ditch the finicky chat solution I have set up now - provided the reward outweighs the risk, and depending on what mitigations are available. **What I'm considering** * One self-hosted MCP gateway on the NAS, exposed via Tailscale Funnel (so TLS ends on my box), OAuth sign-in with 2FA, apps themselves stay tailnet-only * Gateway off by default, switched on when I need it Candidates so far: * **Pomerium** (their documentation says their MCP support is experimental) * **MCP Home** \- researched information has been promising, but seems like a relatively new project * **Build my own gateway** (last resort as that is a lot of faff, but prepared to invest the time to do it if needed) Ruled out: * **Cloudflare Tunnel** \- I don't want a third party terminating TLS and seeing my data * **Pangolin** \- I don't want a VPS to patch from the road * **ContextForge** \- too heavy for my modest Synology NAS **Questions** 1. Has anyone run Pomerium, MCP Home or another gateway in front of MCP servers for Claude or ChatGPT? How did it go? 2. Is Funnel a sensible way to expose just one gateway, or is there a better pattern? 3. Anything you'd lock down that I've missed? (I'm planning per-app scoped keys, read-only first, rate limits, logging and alerts.) 4. Or is there a simpler approach I'm overthinking? Thanks and grateful for any feedback.
document fetched 2026-11-03 00:30 · canonical https://reddit.com/r/selfhosted/comments/1xly5s7/how_do_you_let_claudechatgpt_reach_selfhosted