tolpo

[tolpo]

Local-first control of TP-Link Tapo smart-home devices for AI agents.

[CI] [License: MIT] [Python 3.11+]

tolpo turns the Tapo devices on your network into a single interface that an AI agent, or your own code, can read and control. It talks to each device directly over the LAN through the open source python-kasa library, so no request ever leaves your home and no vendor cloud sits in the path. The only credential it needs is the Tapo account the devices already belong to.

It ships two interfaces over the same core:

• An MCP server that assistants such as Claude use as tools (turn off the office lights, is any door open?).
• An HTTP API with an OpenAPI description at /docs, for scripts and dashboards.

The aim is that everything the Tapo app can do on your network can also be done by the agent: lights, plugs, hubs and their sensors, cameras, doorbells, robot vacuums, schedules, firmware, rooms, pairing, and even what the devices say to the internet. That is a lot of power, so every action has a level, and only the first levels are on until you raise them. See What an agent can do and the Security model.

Features

• Lights, plugs, the H100/H200 hub and its sensors, cameras and the doorbell, each with live state and a list of what can be changed.
• The parts of the Tapo app are split into areas, such as lights, plugs, cameras, doorbells and the network. tolpo remembers which kinds of device it found in earlier runs, and an area for a kind you have never had stays off.
• Every action has a level from 0 to 4. Reading and reversible control (power, brightness, color) are on. Settings, security changes and things that are hard to undo, such as pairing or a factory reset, stay off until you raise the level, and the last two kinds also need a confirmation phrase on every call. Wi-Fi names, setup codes, passwords and tokens are never returned.
• A camera you name in the config gives an agent a still picture, a short run of pictures or a saved clip, through the same video channel the Tapo app uses. No extra camera account is needed.
• Discovery combines a broadcast with a direct per-address sweep, remembers devices between runs and retries the ones that are slow to wake. When a device cannot be reached, the answer says which one and why.
• State is cached briefly and device names resolve at once, so an agent is not blocked on a slow hub.
• A full test suite runs against simulated devices, so it needs no hardware.

Supported devices

tolpo works with whatever python-kasa can talk to, so the list is longer than the hardware we own. The table keeps two things apart: what we have run against real devices, and what python-kasa documents as supported but we have not tried.

The C610 and the C465 are not on python-kasa's list yet, but both connected fine. The C465 needs the small certificate fix in compat.py. The full upstream list is in python-kasa's SUPPORTED.md.

A model that is not listed may still work, because tolpo hands the device to python-kasa and shows whatever it reports. Run tolpo scan to find out. If a device works or does not, tell us through a device support issue so the table can improve.

Requirements

• Python 3.11 or newer
• One or more TP-Link Tapo devices on the same network, already set up in the Tapo app
• Your Tapo account email and password
• ffmpeg, only if you want pictures from cameras
• uv (or plain pip) to install

Installation

With npx

You need Node and uv. Nothing is installed globally:

npx -y tolpo scan      # find your devicesnpx -y tolpo mcp       # run as an MCP server

tolpo is a Python program. The npm package carries the Python code and runs it with uv, so the first run downloads about forty Python packages and takes longer than later runs. If an AI client gives up waiting, run npx -y tolpo --help once and start the client again.

With Docker

For a machine that stays on, such as a Raspberry Pi or a Linux home server. The image is public, so pulling it needs no login, and it is built for amd64 and arm64.

Keep the credentials in a file that only you can read, and pass it with --env-file:

(umask 077; printf '[email protected]\nTAPO_PASSWORD=your-password\n' > ~/.tolpo.env)
docker run --rm -i --network host --env-file ~/.tolpo.env ghcr.io/signerless/tolpo

That runs the MCP server over stdio. For the HTTP API, add a token, because it is no longer bound to loopback:

echo "TOLPO_TOKEN=$(openssl rand -hex 24)" >> ~/.tolpo.env
docker run -d --name tolpo --restart unless-stopped --network host --env-file ~/.tolpo.env \  ghcr.io/signerless/tolpo serve --host 0.0.0.0

--network host lets the container find devices by UDP broadcast, and it only works on Linux. Docker Desktop on macOS and Windows runs containers in a virtual machine that cannot see your devices that way, so use the npx install there.

CI builds the image for both architectures and checks the MCP handshake inside it. Finding real devices from inside a container has not been tried yet. The image has no ffmpeg, so camera pictures do not work in it yet.

From source

git clone https://github.com/signerless/tolpo && cd tolpouv venv --python 3.13 .venvuv pip install --python .venv/bin/python -e ".[dev]"

Configuration

Credentials can come from environment variables or a config file. The server reads the file first, then lets the environment override it.

export TAPO_EMAIL='[email protected]'export TAPO_PASSWORD='your-password'

Or ~/.config/tolpo/config.toml (run chmod 600 on it, since it holds a password):

[tapo]email = "[email protected]"password = "your-password"# or point at an existing JSON file instead of copying the secret:# credentials_file = "~/.tapo_config.json"
[server]host = "127.0.0.1"          # loopback only; a token is required to bind elsewhereport = 8765# token = "a-long-random-string"# cache_ttl_seconds = 10# areas = ["lights", "plugs", "hub"]   # which parts offer tools; empty means every part for the devices you own
[discovery]# subnets = ["192.168.1.0/24"]   # default: the /24 of your default-route interface# hosts   = ["192.168.10.20"]    # devices that ignore broadcast or live on another VLAN# sweep   = true                 # also probe every address in the subnet (networks up to /22)interval_seconds = 600
[policy]# The highest level of action an agent may use, from 0 to 4. See the Security model.tier = 1# Single device methods an agent may use above `tier`, for example ["setLedStatus"].allow_methods = []# Features an agent may write beyond power, brightness and color.# Plain ids ("pan_left") or kind-qualified ones ("camera.state").allow_write = []
[cameras]# Pictures are off for every camera until you name it here, as it is called in the Tapo app.snapshots = []                    # for example ["Hallway Cam"]# Optional. A camera account made in the Tapo app switches pictures to the standard RTSP stream.# username = "camera-account-name"# password = "camera-account-password"# ffmpeg = "/opt/homebrew/bin/ffmpeg"   # only if ffmpeg is not on your PATH# clip_dir = "/Users/you/tolpo-clips"   # the one folder where camera clips may be saved; no folder, no clips

Environment variables override the file: TOLPO_CONFIG (the file itself), TOLPO_CREDENTIALS_FILE, TOLPO_TIER, TOLPO_AREAS (names separated by commas), TOLPO_SNAPSHOTS, TOLPO_CLIP_DIR, TOLPO_FFMPEG, TOLPO_CAMERA_USERNAME and TOLPO_CAMERA_PASSWORD. A tier that is not a whole number from 0 to 4 stops the server from starting.

Some parts keep their own files, so that no secret ends up in the main one. Automation keeps its rules and alert outbox, and the local rooms and groups are kept in ~/.local/state/tolpo (change it with TOLPO_STATE_DIR). The network area reads ~/.config/tolpo/network.toml (change it with TOLPO_NETWORK_CONFIG), which names the environment variable that holds each router password and never holds the password itself.

Verify your setup with a read-only scan: npx -y tolpo scan, or .venv/bin/tolpo scan from a source checkout.

Usage with Claude (MCP)

Put your credentials in ~/.config/tolpo/config.toml (see Configuration, and chmod 600 it). The server reads that file by itself, so nothing secret ends up in your AI client's settings. Then register the server once:

claude mcp add tolpo -- npx -y tolpo mcp

From a source checkout, point at your own copy instead:

claude mcp add tolpo -- $PWD/.venv/bin/tolpo mcp

If you would rather keep the credentials out of a file, pass them with -e TAPO_EMAIL=... -e TAPO_PASSWORD=..., but your client will then store them in its own settings in plain text.

After the next restart you can ask things like "dim office-1 to 30%", "is any door open?" or "what is the temperature upstairs?".

Read tools are marked read-only and set_feature is marked destructive, so MCP clients can ask for confirmation where it matters. Responses are compact: all the devices of a typical home fit in under a thousand tokens.

What an agent can do

The Tapo app is split into areas. Each area is a group of MCP tools with its own page that lists every tool. tolpo remembers the kinds of device it found in earlier runs, and an area for a kind you have never had is off, so a house without a vacuum does not load vacuum tools. On a first run, or when a model is not recognized, every area is on. Pick the areas yourself with areas under [server], or TOLPO_AREAS. The tool tolpo_areas tells an agent what is on.

On top of these, home_overview, list_devices, get_device, set_power, set_light, set_feature, camera_snapshot and rescan work for every device.

All the areas together are 242 tools. Their names, descriptions and parameter lists come to about 40,000 tokens. The standard setup switches on only the areas that match the devices found at home. A client that loads every tool at once can use areas to keep only what it needs.

Most of this was built from analysis of the Tapo app and tested against simulated devices. Each area page says which parts were tried on real devices and which were not. Where the exact form of a request was a guess, the page marks it UNVERIFIED, and the tool refuses when the device does not report support for it.

Usage over HTTP

npx -y tolpo serve          # or .venv/bin/tolpo serve from a source checkout# http://127.0.0.1:8765, interactive docs at /docs

# Readcurl -s localhost:8765/devices                 # everything, with state and capabilitiescurl -s localhost:8765/devices/front-door       # one device, every featurecurl -s "localhost:8765/devices?kind=sensor"    # doors, windows, water, temperature
# Writecurl -s localhost:8765/devices/hallway-cam/snapshot -o picture.jpg    # a camera you enabledcurl -s -X POST localhost:8765/devices/desk-lamp/power \  -H 'content-type: application/json' -d '{"on": true}'curl -s -X POST localhost:8765/devices/office-1/light \  -H 'content-type: application/json' -d '{"brightness": 40, "color_temp": 3000}'curl -s -X POST localhost:8765/batch/light \  -H 'content-type: application/json' -d '{"targets": ["office*"], "brightness": 60}'

A device is addressed by its short id (office-1), by its name in the Tapo app (case-insensitive), or, in the batch endpoints, by a glob such as office*. If a name matches several devices the server returns 409 with the candidates instead of guessing.

The house in 3D

tolpo serve also serves a 3D view of the house at http://127.0.0.1:8765/home. It draws the rooms, puts every device in its room and shows its state: lights glow in their real color, a door or window drawn on the wall opens when its sensor reports open, a water leak turns red, the temperature floats over its sensor. Click a device to switch it, dim it or change its color. Click a camera or the doorbell to watch it live. Rooms without a device are left out.

The floor plan is yours and stays on your machine, in ~/.config/tolpo/home.json (server.home_layout or TOLPO_HOME_LAYOUT to move it). The easiest way to make one is to ask your agent to "build my home plan": with the home_plan_draft, home_plan_get and home_plan_set tools it drafts a plan from your devices, asks you about your rooms and saves it. You can also press Edit layout in the view, drag devices where they belong, and save. The format is in docs/areas/home_plan.md:

{"version": 1, "rooms": [{"id": "office", "name": "Office", "x": -7, "z": -5, "w": 5, "d": 5},           {"id": "garden", "name": "Garden", "x": -9, "z": -12, "w": 21, "d": 7, "outside": true}], "devices": {"office-1": {"room": "office", "x": -5.75, "z": -3.75},             "garden-cam": {"room": "garden", "x": 1.5, "z": -5.6, "yaw": 0}}}

Live video comes from GET /devices/{id}/live, a stream of JPEG pictures (about 10 per second) from a camera you listed in cameras.snapshots. One stream per camera and two cameras at once. A view ends after 15 minutes, or after 2 minutes on a battery camera, because live video keeps it awake. The page loads three.js from one pinned jsDelivr path and may talk to this server only. It asks battery cameras for their state every five minutes, not every few seconds, so it does not wake them.

Door and window history

A Tapo hub keeps no history of its own for sensors connected to TP-Link's cloud: asked the way the app asks, an H200 answers with an empty list for a door sensor, and the app reads that history from the cloud instead. So tolpo records its own:

tolpo record garage

It reads the sensor every 5 seconds (--every to change it) and saves each opening and closing, with its time, in ~/.local/share/tolpo/history.jsonl (server.history_file or TOLPO_HISTORY_FILE), a file only you can read. Leave out the name to record every door and window sensor. It records only while it runs, so run it as a background service for a history without holes; times when the computer slept or the sensor was out of reach are written down as gaps. The recorded_history tool then tells an agent how often a door opened, for how long, per day and per hour, and lists the times nothing was recorded instead of guessing about them.

Camera pictures

An agent can ask a camera for one still picture. Every camera is off until you name it in the config, with the name it has in the Tapo app:

[cameras]snapshots = ["Hallway Cam"]

You also need ffmpeg on the machine that runs tolpo. Check the setup yourself before an agent uses it. This saves one picture to a file that only you can read:

npx -y tolpo snapshot "Hallway Cam" -o hallway.jpg

To watch a camera live on your own screen, use npx -y tolpo watch "Hallway Cam". It opens ffplay, which comes with ffmpeg, and closes when you close the window. --seconds 30 stops it sooner. Nothing is recorded and nothing goes to an agent.

tolpo takes one frame from the camera's live stream with ffmpeg and returns it as a JPEG, at most 1280 pixels wide. --quality high asks for the main stream instead of the small one. There are two ways to reach the stream:

• The video channel of the Tapo app, on port 8800 of the camera. It uses the Tapo account tolpo already has, so there is nothing more to set up.
• RTSP, if you create a camera account in the Tapo app and add it under [cameras] as username and password. When a camera account is configured, tolpo uses RTSP.

What to know before you turn it on:

• A picture goes to the AI model your agent runs on. Enable only the cameras you are happy for it to see.
• A camera in privacy mode gives no picture. An agent cannot switch privacy mode off unless you allow camera.state in allow_write.
• Every picture is logged with the camera and the time, never with the picture itself.
• If a camera rejects the login, tolpo leaves it alone for five minutes, because repeated failures can lock a camera out.
• A picture is one frame from now, not a recording. Battery cameras can be asleep and may not answer.
• With RTSP, ffmpeg receives the camera account on its command line, so other users on the same machine can see it for a moment.
• A doorbell behind the hub has no address of its own. Its pictures and live video come through the hub, like the app's, and the first picture takes about 5 s while a battery doorbell wakes up. Its recordings are not reachable yet.

The camera_media area builds on this. It can read the memory card and its recording index, return the thumbnail of a past event, take a short run of pictures from the live stream, and save a recording to the one folder you set as clip_dir. Thumbnails, runs of pictures and clips follow the same rules as pictures: only cameras you named, never in privacy mode, and the tool that saves a clip returns the path and never the video. Reading the card status, the record plan and the list of recordings and events works for any camera. Formatting a card is a level 4 action.

Security model

This server controls physical things, so it refuses by default and widens only on request.

Every action on a device has a level. You choose the highest level an agent may use with tier under [policy] (1 by default), and you can allow single device methods above it with allow_methods.

Levels 3 and 4 also need a confirmation phrase on every call, such as setLedStatus hallway-cam. The refusal says what the phrase is, so the phrase is a pause for the agent to ask the person, not proof that the person agreed. The phrase names the action and the device (a few name a change or a plan instead), so one confirmation cannot be reused for another device or another action. A refusal says what the action is, and never how to widen the permission.

Cameras are read-only through set_feature: the state switch on a camera is its privacy shutter, not a power relay, so it is not treated as one. Enable it deliberately with allow_write = ["camera.state"], or use the privacy tool at level 3. A feature named in allow_write is allowed at any level and needs no phrase, so name only what you mean to give away.

Rules that run without anyone asking (automation) are held to the same levels at the moment they run, not only when they are made. Setting them up needs level 2. An agent can stop all rules at any time, but starting them again needs level 3 and a phrase. You can do it yourself with tolpo automation resume, which a server that is already running picks up when it restarts.

The server listens on 127.0.0.1 only. It refuses to start on any other address without a bearer token, and without a token it also rejects requests whose Host header is not local, which blocks DNS-rebinding attacks from a web page.

Device names are free text typed by whoever set the device up. The server flattens and shortens them before an agent sees them, and the instructions it sends tell the agent to treat names and readings as data, not commands. A shared Tapo account still means anyone in it can rename a device, so keep the account to people you trust.

How discovery handles a busy network

Homes often have many devices, some battery-powered and slow to answer. Discovery is designed for that:

• A UDP broadcast, followed by a direct query to each address in the subnet (discovery.sweep), because some battery cameras ignore broadcasts yet always answer a direct question.
• Devices seen before are remembered in ~/.cache/tolpo/hosts.json (address, name and model, no secrets) and queried directly at startup. A device missing from one scan is never forgotten; entries unseen for 30 days are dropped.
• Devices that do not answer are retried after 5, 15 and 45 seconds, and again on every rescan.

Startup is staged so the server is useful immediately: the first scan runs in the background, and each device is connected the moment it answers. A device you name exactly is usable as soon as it is up, without waiting for the hub, which can take about ten seconds from cold. Operations that span the whole house wait for the scan, but never longer than fifteen seconds; anything still connecting finishes in the background.

When something cannot be reached, the answers say so. home_overview and list_devices carry complete: false and a not_connected list with the name, model and reason for each unreachable device, so an agent does not mistake an offline door sensor for a closed door.

For AI agents

Call tolpo_areas first. It lists the parts of tolpo that are on, what each one offers, and the level the owner allows. If you are an AI agent asked to set tolpo up or to use it, start with llms.txt. If you are asked to change the code, read AGENTS.md first.

Development

.venv/bin/python -m pytest        # runs against simulated devices, no hardware required

Project layout:

src/tolpo/  registry.py    device discovery, state cache, control operations  policy.py      what may be read and written (deny-by-default)  views.py       shared presentation for the HTTP and MCP layers  api.py         FastAPI HTTP server  mcp_server.py  MCP server (stdio)  config.py      config file and environment loading  compat.py      small fixes for devices python-kasa cannot reach yet  snapshot.py    one picture from a camera, taken with ffmpeg  media.py       the video channel of the Tapo app (port 8800)  tsvideo.py     finds the video track in a camera's stream  calls.py       the one way an area sends a method to a device, checked against its level  provision.py   Wi-Fi setup of a new device from this machine (SoftAP), behind a plan  areas/         one file per area of the Tapo app; adding an area adds a file  automation/    the local rule engine  network/       providers for routers, DNS servers and firewalls  cli.py         serve / mcp / scan / snapshot / watch / automation entry pointstests/           full suite against simulated devicesdocs/protocols.md  how Tapo devices talk on the local networkdocs/areas/      one page per area, with every tool and what was only tried with simulated devicesassets/          logo, device icons and social preview

See CONTRIBUTING.md for the ground rules, docs/protocols.md for the protocol notes and SECURITY.md to report a vulnerability.

Background

The protocols behind Tapo devices are not documented by the vendor. Part of this project is independent analysis of the official Tapo app and of real devices, done so that software can interoperate with hardware its owners already have. The result is written down in docs/protocols.md: which transport each model uses, how discovery works, what the newer TPAP protocol looks like, and what we could not establish. The notes are in our own words. They contain no vendor code, keys or credentials, and tolpo itself only ever signs in with your own Tapo account.

Compatibility and status

This is version 0.2.1. Lights, plugs, hubs and sensors are supported over the KLAP and SSL-AES transports through python-kasa, plus a small patch (compat.py) for cameras with an ECDSA certificate, such as the C465, that python-kasa cannot reach on its own.

The areas beyond the first release were built from analysis of the Tapo app and tested with simulated devices. Reading was tried on a real home with four bulbs, a hub, eight sensors, a doorbell, three cameras and a plug. A few reversible writes were tried there too: the bulbs (power, brightness, color, power-on defaults, saved presets and disabled rules), the plug (power, countdown, LED, power-on default and disabled rules), a camera light, the hub light, a microphone volume, a detection switch and an automation rule, each set back afterwards. Most other writes have not been tried on real hardware, and the area pages say which. Robot vacuums, door locks, radiator valves and the firewall providers were never tried on the real thing.

Known limits: cameras give still pictures, short runs of pictures and saved clips, but no continuous video to an agent, and a doorbell behind the hub gives pictures but not its recordings; the cloud-only parts of the app (account, homes, sharing, push notifications, Tapo Care, voice assistants) are not reachable, and the automation area replaces some of them locally; the MCP server runs over stdio, one process per client, with no shared HTTP transport yet; the HTTP API has the original routes plus a few for the camera memory card (/devices/{id}/storage, /recordings/...) and the door lock state, and the other areas are MCP tools only; and devices on a second subnet have not been tested.

License

MIT, see LICENSE.

This project is not affiliated with or endorsed by TP-Link. Tapo is a trademark of TP-Link. All device communication goes through the open source python-kasa library.