As I still run my Indigo server on Intel silicon I needed a solution for the fact that Mike’s excellent MCP server no longer runs on Intel.
So I’ve spent the last 2 days, and evenings build a different take on an indigo mcp server that initially is a lite version that runs anywhere and is all client side, no embeddings so no cost to use. It uses Claude to handle the semantic search and look up to various indigo device types.
I’d love some testers to ensure it works more widely and to test out whether, and where, embeddings are useful, whether that is semantic search Claude can’t do, use of descriptions or something else, and so whether that should be an extension. If it is, it would be with support for intel silicon.
MCP Lite
Re: MCP Lite
Two ways to give Claude access to Indigo — and why I built a lite alternative
A lot of you are using Mike Lamoureux's excellent indigo-mcp-server to talk to Claude about your Indigo system. It's a great plugin, and I want to be clear up front: this isn't a replacement, it's an alternative for people who can't run his.
How Mike's plugin understands your devices
When you ask Claude something like "find anything that detects moisture", Mike's plugin needs to figure out that your device named Basement Leak Sensor is a match — even though the word "moisture" doesn't appear anywhere in its name.
It does this by sending all your device names and descriptions to OpenAI, which converts each one into a long list of numbers (called an "embedding") that captures the meaning of the words. Your query gets the same treatment, and the plugin finds the device whose meaning is closest mathematically. It's clever and works really well — that's why Mike chose this approach.
The downside: it needs an OpenAI API key (small ongoing cost), it relies on a library called LanceDB which has gotten harder to install on Intel Macs, and it stopped working entirely on Intel Macs running Indigo 2025.2 because LanceDB no longer publishes the right packages for that combination.
How my "lite" plugin does the same thing
Indigo MCP Lite skips the OpenAI step. Instead, it relies on the fact that Claude is already on the other end of the conversation, and Claude already knows that "moisture" means leak/water/damp/wet. So when you ask "find anything that detects moisture", Claude rewrites it to look for any of those words before sending the query to your Indigo system. The plugin then does a much simpler text search — the same kind your Mac uses when you Spotlight-search for a file.
The result is roughly the same: you get your basement leak sensor. But the plugin does almost nothing clever itself — it just hands Claude's reformulated query to a fast keyword search and ranks the results sensibly.
Which should you use?
- If you're on Apple Silicon and Mike's plugin works for you — keep using it. The semantic search is a real edge, and you've already paid the install cost. No reason to switch.
- If you're on Intel Mac and Indigo 2025.2 broke Mike's plugin — try lite. It's why I built it.
- If you want the simplest possible setup — try lite. No external accounts, no API keys, no monthly cost.
- If you want both — you can run them side-by-side under different names. Mike's for vibe queries, lite for everything else.
The two plugins share most of their tool surface (turn lights on/off, read sensors, run action groups, etc.) so Claude won't really know the difference for routine commands. The split only shows up in fuzzy natural-language searches.
And both work alongside the Home Intelligence MCP server that does some complimentary and additional functionality. I didnt bundle lite into that one so that it can run alongside either of the core MCP servers. You can use any combination in Claude Code or Claude CoWork or any other AI interface of your choice.
Any questions, please ask.
A lot of you are using Mike Lamoureux's excellent indigo-mcp-server to talk to Claude about your Indigo system. It's a great plugin, and I want to be clear up front: this isn't a replacement, it's an alternative for people who can't run his.
How Mike's plugin understands your devices
When you ask Claude something like "find anything that detects moisture", Mike's plugin needs to figure out that your device named Basement Leak Sensor is a match — even though the word "moisture" doesn't appear anywhere in its name.
It does this by sending all your device names and descriptions to OpenAI, which converts each one into a long list of numbers (called an "embedding") that captures the meaning of the words. Your query gets the same treatment, and the plugin finds the device whose meaning is closest mathematically. It's clever and works really well — that's why Mike chose this approach.
The downside: it needs an OpenAI API key (small ongoing cost), it relies on a library called LanceDB which has gotten harder to install on Intel Macs, and it stopped working entirely on Intel Macs running Indigo 2025.2 because LanceDB no longer publishes the right packages for that combination.
How my "lite" plugin does the same thing
Indigo MCP Lite skips the OpenAI step. Instead, it relies on the fact that Claude is already on the other end of the conversation, and Claude already knows that "moisture" means leak/water/damp/wet. So when you ask "find anything that detects moisture", Claude rewrites it to look for any of those words before sending the query to your Indigo system. The plugin then does a much simpler text search — the same kind your Mac uses when you Spotlight-search for a file.
The result is roughly the same: you get your basement leak sensor. But the plugin does almost nothing clever itself — it just hands Claude's reformulated query to a fast keyword search and ranks the results sensibly.
Which should you use?
- If you're on Apple Silicon and Mike's plugin works for you — keep using it. The semantic search is a real edge, and you've already paid the install cost. No reason to switch.
- If you're on Intel Mac and Indigo 2025.2 broke Mike's plugin — try lite. It's why I built it.
- If you want the simplest possible setup — try lite. No external accounts, no API keys, no monthly cost.
- If you want both — you can run them side-by-side under different names. Mike's for vibe queries, lite for everything else.
The two plugins share most of their tool surface (turn lights on/off, read sensors, run action groups, etc.) so Claude won't really know the difference for routine commands. The split only shows up in fuzzy natural-language searches.
And both work alongside the Home Intelligence MCP server that does some complimentary and additional functionality. I didnt bundle lite into that one so that it can run alongside either of the core MCP servers. You can use any combination in Claude Code or Claude CoWork or any other AI interface of your choice.
Any questions, please ask.
-
durosity
- Posts: 4810
- Joined: Thu May 10, 2012 3:21 pm
- Location: Newcastle Upon Tyne, Ye Ol' England.
Re: MCP Lite
Is this still something you're working on? It sounds more suitable to my needs.
Computer says no.
Re: MCP Lite
Yes. Using this actively but realised not put into the plugin store. Will do otherwise it’s in my GitHub repo.
Re: MCP Lite
This is now in the plugin store. https://www.indigodomo.com/pluginstore/329/
-
durosity
- Posts: 4810
- Joined: Thu May 10, 2012 3:21 pm
- Location: Newcastle Upon Tyne, Ye Ol' England.
Re: MCP Lite
Page not found!siclark wrote: ↑Wed May 20, 2026 9:47 am This is now in the plugin store. https://www.indigodomo.com/pluginstore/329/
Computer says no.
Re: MCP Lite
Waiting on Jay to approve.
-
jay (support)
- Site Admin
- Posts: 19232
- Joined: Wed Mar 19, 2008 11:52 am
- Location: Austin, Texas
- Contact:
-
durosity
- Posts: 4810
- Joined: Thu May 10, 2012 3:21 pm
- Location: Newcastle Upon Tyne, Ye Ol' England.
MCP Lite
Major rewrite of the mcp lite plugin with the latest documentation update.
Also added semantic search like capability without need for using Open AI (whether through choice or force due to intel silicon)
As the name suggests this is focused on core indigo capability without extra libraries or heavy code on non mcp capability.
Use with Claude desktop and via cowork on your phone, in Claude code or any other favourite mcp tool.
Claude’s write up below.
Indigo MCP Lite v2026.7.0 — your whole house, now readable and controllable by AI assistants
A big update to Indigo MCP Lite, the plugin that lets AI assistants like Claude talk to your Indigo server.
Nothing extra to install — genuinely. No pip packages, no Python dependencies, no vector database, no OpenAI account, no API keys, no embeddings library. Everything runs on what Indigo already ships, which means it installs on any Mac Indigo supports — including older Intel machines where heavier AI plugins can no longer build their dependencies. Download, double-click, enable, done.
Since the last announcement the tool count has doubled (33 → 68), driven by a full audit against the current Indigo 2025.2 documentation to find everything an assistant couldn’t yet see or do. Highlights:
Search that speaks human — without a cloud AI in the loop. This is the bit I’m most pleased with: you get semantic-style search with zero external services. A curated synonym layer means “telly” finds your TV devices, “lounge” finds things filed under Living Room, “radiator” finds your TRVs, and “socket” finds plugs and outlets — the kind of matching other plugins have used OpenAI embeddings for, done locally in milliseconds with no key, no per-query cost, and nothing leaving your Mac. Colour commands got the same treatment: ~950 everyday names understood, “duck egg blue” and “burnt orange” included.
It can finally read what you can see. Asking about a device now returns everything Indigo knows — temperatures, setpoints, sensor readings, battery levels, energy usage, last-seen times. “What’s the study thermostat set to?” and “which sensors are below 20% battery?” just work.
Your automations are visible. Triggers and schedules used to be invisible to an assistant. Now it can list them, see what’s enabled and when a schedule next fires, turn them on or off, and run one on demand — “disarm the morning schedule while we’re on holiday” is a sentence, not a settings hunt.
Device history, charted from your own data. If you run the SQL Logger, the assistant can pull time-series history for any device — “show me the server rack’s power draw over the last week” comes back as real data from your own database.
More of the house is controllable. Toggle (including “on for 5 minutes”), locks, sprinkler zones and watering schedules, fan speeds, relative dimming (“a bit brighter”), and utility commands like pinging a Z-Wave module to check it’s reachable or beeping a device to find it.
Z-Wave health. It can inspect your Z-Wave network per node — neighbour tables, signal paths, battery states — genuinely useful for spotting weakly-meshed devices before they start misbehaving.
Built to be trusted. Every action an assistant takes is checked before it runs: it’s told to verify what depends on something before deleting it, lock commands report honestly that they’ve been sent rather than pretending the door moved, and mistyped requests get a helpful correction instead of silently doing the wrong thing.
Update on the plugin store — and if you try it with Claude, Cursor, or any MCP-capable assistant, I’d love to hear what you ask your house to do.
GitHub: https://github.com/simons-plugins/indigo-mcp-lite
Also added semantic search like capability without need for using Open AI (whether through choice or force due to intel silicon)
As the name suggests this is focused on core indigo capability without extra libraries or heavy code on non mcp capability.
Use with Claude desktop and via cowork on your phone, in Claude code or any other favourite mcp tool.
Claude’s write up below.
Indigo MCP Lite v2026.7.0 — your whole house, now readable and controllable by AI assistants
A big update to Indigo MCP Lite, the plugin that lets AI assistants like Claude talk to your Indigo server.
Nothing extra to install — genuinely. No pip packages, no Python dependencies, no vector database, no OpenAI account, no API keys, no embeddings library. Everything runs on what Indigo already ships, which means it installs on any Mac Indigo supports — including older Intel machines where heavier AI plugins can no longer build their dependencies. Download, double-click, enable, done.
Since the last announcement the tool count has doubled (33 → 68), driven by a full audit against the current Indigo 2025.2 documentation to find everything an assistant couldn’t yet see or do. Highlights:
Search that speaks human — without a cloud AI in the loop. This is the bit I’m most pleased with: you get semantic-style search with zero external services. A curated synonym layer means “telly” finds your TV devices, “lounge” finds things filed under Living Room, “radiator” finds your TRVs, and “socket” finds plugs and outlets — the kind of matching other plugins have used OpenAI embeddings for, done locally in milliseconds with no key, no per-query cost, and nothing leaving your Mac. Colour commands got the same treatment: ~950 everyday names understood, “duck egg blue” and “burnt orange” included.
It can finally read what you can see. Asking about a device now returns everything Indigo knows — temperatures, setpoints, sensor readings, battery levels, energy usage, last-seen times. “What’s the study thermostat set to?” and “which sensors are below 20% battery?” just work.
Your automations are visible. Triggers and schedules used to be invisible to an assistant. Now it can list them, see what’s enabled and when a schedule next fires, turn them on or off, and run one on demand — “disarm the morning schedule while we’re on holiday” is a sentence, not a settings hunt.
Device history, charted from your own data. If you run the SQL Logger, the assistant can pull time-series history for any device — “show me the server rack’s power draw over the last week” comes back as real data from your own database.
More of the house is controllable. Toggle (including “on for 5 minutes”), locks, sprinkler zones and watering schedules, fan speeds, relative dimming (“a bit brighter”), and utility commands like pinging a Z-Wave module to check it’s reachable or beeping a device to find it.
Z-Wave health. It can inspect your Z-Wave network per node — neighbour tables, signal paths, battery states — genuinely useful for spotting weakly-meshed devices before they start misbehaving.
Built to be trusted. Every action an assistant takes is checked before it runs: it’s told to verify what depends on something before deleting it, lock commands report honestly that they’ve been sent rather than pretending the door moved, and mistyped requests get a helpful correction instead of silently doing the wrong thing.
Update on the plugin store — and if you try it with Claude, Cursor, or any MCP-capable assistant, I’d love to hear what you ask your house to do.
GitHub: https://github.com/simons-plugins/indigo-mcp-lite
Re: MCP Lite
Fable has been busy.
Whilst still keeping this generic to indigo and not specific home use I realised that i had created framework for users to share the device specifics for plugins with my device catalogue plugin.
So now the mcp lite server knows about plugin capabilities from that catalogue. If you’ve not contributed or plugins and missing please install and contribute.
Claude’s enthusiastic sales pitch is below.
Indigo MCP Lite v2026.9.0 — now capability-aware, powered by the community Device Catalog
A focused update to Indigo MCP Lite, the zero-dependency plugin that lets AI assistants like Claude see and control your Indigo server — this one’s about the plugin learning what your devices can actually do.
The problem it solves: Indigo’s scripting API tells an assistant a device’s type and states, but not its capabilities. Ask Claude to set a plain on/off bulb to “warm white” and the best it could do before was fire the command and relay a cryptic failure. It was guessing.
Now it knows. MCP Lite ships with a built-in snapshot of the community Device Catalog — the same capability database that powers device rendering in the Domio iOS app. Two things happen with it:
• Device details include real capabilities. Ask about a bulb and the answer now includes whether it supports colour, white temperature, dual white channels, and so on — per plugin, per device type, as catalogued by people who own the hardware.
• Helpful refusals instead of cryptic failures. Ask for RGB on a device the catalogue says can’t do it, and the assistant gets back “this device doesn’t support RGB — it supports on/off” — so it corrects course or tells you plainly, instead of pretending it worked. And if a device isn’t in the catalogue at all, nothing is blocked: commands pass through exactly as before. The catalogue only ever adds knowledge, never takes away control.
Still nothing extra to install. The catalogue rides along inside the plugin as plain data — no network calls, no accounts, no dependencies. It refreshes automatically with plugin updates: when the community catalogue gains new profiles, a new MCP Lite release follows with the updated snapshot baked in.
And it helps grow the catalogue itself. A new list_uncataloged_devices tool shows which of your devices have no catalogue profile yet — effectively a personalised contribution list. On my own server it found 30 devices the moment it went live (all from one plugin — profiles on the way). If it surfaces devices you own, a profile contribution takes minutes and improves things for every catalogue consumer: AI assistants, Domio, and whatever comes next.
Grab it from the Releases page (updates via the Plugin Store). If you run it, I’d genuinely love to hear what list_uncataloged_devices finds on your server — every profile contributed makes the whole ecosystem a bit smarter.
Whilst still keeping this generic to indigo and not specific home use I realised that i had created framework for users to share the device specifics for plugins with my device catalogue plugin.
So now the mcp lite server knows about plugin capabilities from that catalogue. If you’ve not contributed or plugins and missing please install and contribute.
Claude’s enthusiastic sales pitch is below.
Indigo MCP Lite v2026.9.0 — now capability-aware, powered by the community Device Catalog
A focused update to Indigo MCP Lite, the zero-dependency plugin that lets AI assistants like Claude see and control your Indigo server — this one’s about the plugin learning what your devices can actually do.
The problem it solves: Indigo’s scripting API tells an assistant a device’s type and states, but not its capabilities. Ask Claude to set a plain on/off bulb to “warm white” and the best it could do before was fire the command and relay a cryptic failure. It was guessing.
Now it knows. MCP Lite ships with a built-in snapshot of the community Device Catalog — the same capability database that powers device rendering in the Domio iOS app. Two things happen with it:
• Device details include real capabilities. Ask about a bulb and the answer now includes whether it supports colour, white temperature, dual white channels, and so on — per plugin, per device type, as catalogued by people who own the hardware.
• Helpful refusals instead of cryptic failures. Ask for RGB on a device the catalogue says can’t do it, and the assistant gets back “this device doesn’t support RGB — it supports on/off” — so it corrects course or tells you plainly, instead of pretending it worked. And if a device isn’t in the catalogue at all, nothing is blocked: commands pass through exactly as before. The catalogue only ever adds knowledge, never takes away control.
Still nothing extra to install. The catalogue rides along inside the plugin as plain data — no network calls, no accounts, no dependencies. It refreshes automatically with plugin updates: when the community catalogue gains new profiles, a new MCP Lite release follows with the updated snapshot baked in.
And it helps grow the catalogue itself. A new list_uncataloged_devices tool shows which of your devices have no catalogue profile yet — effectively a personalised contribution list. On my own server it found 30 devices the moment it went live (all from one plugin — profiles on the way). If it surfaces devices you own, a profile contribution takes minutes and improves things for every catalogue consumer: AI assistants, Domio, and whatever comes next.
Grab it from the Releases page (updates via the Plugin Store). If you run it, I’d genuinely love to hear what list_uncataloged_devices finds on your server — every profile contributed makes the whole ecosystem a bit smarter.