Larnitech

Larnitech MCP

Version 1.4.0 Beta · Changelog · MIT licensed

An MCP server that lets an AI agent read and control a Larnitech smart-home installation over the API2 protocol — lights, climate, blinds, sensors, meters.

It ships with a documented device-type reference built from live testing against real controllers, so the agent looks up how a widget actually behaves instead of guessing from key names. That matters more than it sounds: on this platform writing state: "closed" to a gate is acknowledged with success: true and then silently ignored, and several climate types drop half of any two-key write. Those traps are documented, checked before every write, and listed in this README's own safety section.

---

What it does

Reading — open, no configuration beyond the API key.

Consecutive calls share one authorized connection: the socket stays open for four minutes after a request, so a burst of reads costs one connect instead of one per call. It is retired before the controller's own five-minute idle timeout, and one the controller has already dropped is replaced without the call failing.

Understanding what came back — statuses are type-specific and occasionally not key/value at all.

Responses flag statuses that need care: an opaque hex blob, a malfunction fault code in place of a normal reading, an all-null payload from a meter that missed its poll cycle — which means *no data*, not zeros — or a bare state: "undefined", which is how a device that is physically offline still answers.

Watching — non-blocking, for "press the switch and tell me what moved".

A watch keeps its own connection alive, so it survives the controller's 5-minute idle timeout and can stay open across a long conversation.

Writing — off by default; one call, checked before and after.

Learning — findings survive the session.

Saving data — a snapshot is worth more than scrollback.

Files land in data/ / _ _ .txt, one folder per controller — or in a folder of your own, if that installation already has a project directory holding its drawings, configs and controller backups:

python -m larnitech_mcp data-dir "Home" "/projects/home/backups"

Snapshots then sit with the rest of that site's files instead of in a separate store. list_snapshots finds both kinds.

Each is a JSON envelope — object, saved_at, comment and the payload under data — so provenance travels with the file and read_snapshot hands back parsed structure rather than text to re-parse. Numbers stay numbers and null stays null, which matters here: telling "no data" apart from zero is the difference between a quiet meter and a reading of nothing.

Ask for "save the current state" and the agent files it there; ask "has this changed since last week" and it has something to compare against.

Reporting — report_bug turns something you hit into a ready-to-file issue for this repository. It composes the report, strips identifiers (API keys, serials, hostnames, site names) and returns a prefilled link. Nothing is posted: you open the link, read exactly what would be published, and submit it yourself under your own account. No token is needed by anyone.

---

Requirements

• Python 3.11+
• A Larnitech controller you administer, reachable via the Larnitech cloud or on your LAN

Install

pip install larnitech-mcp

A virtual environment is worth using, since you'll point Claude Code at that interpreter's absolute path:

python -m venv ~/.venvs/larnitech

~/.venvs/larnitech/bin/pip install larnitech-mcp # Windows: Scripts\pip.exe

Your API keys, preferences, saved snapshots, and any device notes the agent records live in ~/.larnitech-mcp/, outside the package, so upgrading never touches them.

To work on the server itself, install from a clone instead:

git clone https://github.com/mpopovych-thinkhome/larnitech-mcp.git

cd larnitech-mcp pip install -e .

Where to get your API key

Both values come from LT_Setup, Larnitech's own configuration app, on the installation you administer:

The key grants full read and write access to the installation. Treat it like a password.

Connect a controller

Register it, then store the key:

python -m larnitech_mcp add "Home" cloud --serial YOUR_SERIAL

python -m larnitech_mcp auth "Home"

auth prompts with hidden input and writes the key to ~/.larnitech-mcp/project_keys.json. The key never passes through the chat transcript this way — prefer it over the object_set_key tool, which works but leaves the key in the session log.

For a controller on your LAN instead of via the cloud:

python -m larnitech_mcp add "Home" local --host 192.168.1.50

Check it works:

python -m larnitech_mcp test "Home"      # connect, authorize, count devices

python -m larnitech_mcp devices "Home" # full snapshot, counts per type

Add to Claude Code

Add the server to the top-level mcpServers object in ~/.claude.json, using the absolute path to the venv's Python:

{

"mcpServers": { "larnitech": { "command": "/home/you/.venvs/larnitech/bin/python", "args": ["-m", "larnitech_mcp", "serve"] } } }

On Windows the command is the .exe, with escaped backslashes:

"command": "C:\\Users\\you\\.venvs\\larnitech\\Scripts\\python.exe"

It must be the top-level mcpServers key in ~/.claude.json itself — a .mcp.json placed inside the ~/.claude/ folder is never read. To scope it to one project instead, put the same block in a .mcp.json at that project's root. Restart Claude Code fully afterwards; closing the window is not enough.

Other MCP clients work the same way — the server runs on stdio via python -m larnitech_mcp serve.

---

Safety model

Reading is open. Writing is off until you turn it on, per controller, from a terminal:

python -m larnitech_mcp allow-write "Home" on

There is deliberately no tool for this — an agent cannot grant itself write access, only tell you the command.

Every write is checked before it is sent. set_device validates the payload against the documented vocabulary and ranges for that type — an on/off key gets only on/off, an AC setpoint stays inside the range the unit itself reports, a read-only sensor takes nothing — and rejects a malformed or out-of-range command without touching the controller.

And checked after it lands. An ack only means the command was accepted; some types ack a write and then ignore it. So set_device waits for the device to go quiet and reads it back: mismatched lists requested keys that did not take, unrequested_changes anything the controller altered on its own.

Writes can be sequences, because some types cannot be driven with one frame:

[{"status": {"mode": "heat"}, "delay_after": 1.0},

{"status": {"state": "on"}}]

Sending mode and state together loses the state: the controller re-evaluates the channel after a mode change and overrides whatever arrived behind it. The same applies to vent (state + fan) and to clearing an automation before switching a channel off. set_device rejects the combined forms and tells you the sequence to use instead.

No key is ever returned by a tool, and keys are masked out of error messages.

Device documentation

larnitech_mcp/docs/device-types/ holds one file per device type plus an index, covering the API2 status keys, XML attributes, script-side byte layout, and every quirk confirmed by live testing. bugs.md alongside it is a numbered registry of confirmed vendor bugs that the type files reference. Read them through get_docs rather than by path — that also picks up anything you've added locally.

Notes the agent records with add_docs_note go to ~/.larnitech-mcp/docs/, not into the installed package, so they survive upgrades. Your copy wins on read; everything you haven't edited still comes from the shipped set.

This is a working knowledge base, not a spec: entries say plainly when something is confirmed live, observed but unexplained, or still unknown. Corrections and additions are welcome — that is the most valuable kind of contribution here.

Command reference

Status

Beta. Reading, watching, and writing all work and have been exercised against live hardware, but this has been tested against a limited set of installations. Device types documented as unconfirmed genuinely are — see get_docs output and the per-type files.

Modbus on a com-port

modbus_request sends one frame — slave, function 03/04/06/16, register, count or values — onto the bus behind a com-port and returns the decoded answer. Reads work on any object; writes need allow_write. Modbus TCP ports take the same frame as RTU, CRC included (assumed; verified on RTU only).

The port may be in use by the controller's own scripts, so other requests can be on the bus at the same time, and an answer can belong to someone else's request rather than yours — observed live. One request is the default; when a value really matters, send the same request twice about a second apart and trust it only if both answers agree. A device that does not answer within 2 s is reported as a timeout, never as the port's last known frame. Details: com-port.

Not affiliated with or endorsed by Larnitech.

License

MIT

Contact

Mykhailo Popovych

• Telegram: [t.me/M_Popovych_ThinkHome](https://t.me/M_Popov