Skip to main content

readerboard

CI Release PyPI GHCR Docker Hub License: MIT

An HTTP service that drives a BetaBrite Classic sign, either through a serial cable or through an Ethernet to RS-232 adapter.

Several sources can share the sign at once. Each registers a named slot, and the sign rotates through the registered slots by itself. A variable is a value that messages call by name, such as a temperature, and changing it does not blank the sign or restart the message showing it. An alert takes the whole display over until it is released, after which the rotation resumes.

What it does

  • Many messages, one sign. Home Assistant can own temperature while a doorbell automation owns doorbell, without either knowing about the other.
  • The sign does the rotating. Each message lives in its own sign file and the sign cycles them on its own, so rotation costs no serial traffic at all.
  • Live values without a blink. A message such as Outside <var:temp><degree>F calls the variable temp, and a new value is one small write that the sign shows the next time it draws the message, with no blank and no restart. One variable can appear in any number of messages, and a value that stops arriving can be made to go stale.
  • Alerts. Take the display over, optionally with a deadline, then hand it back.
  • It keeps the sign's clock right, at startup, hourly, and whenever the link comes back. That last trigger is the one that matters: a sign returning from a power cut does so at no particular minute. The sign is set one minute fast on purpose: the protocol has no seconds field, so a sign told the current minute reads behind for the rest of it and never ahead, and a minute of lead puts the error on the side that reads as a clock being a touch fast rather than most of a minute slow.
  • It does not redraw the sign for nothing. A write of bytes the sign already holds is suppressed, so a source re-sending an unchanged temperature does not make the display flicker.
  • It survives restarts and outages. The registered messages are persisted and pushed to the sign again whenever the link returns, so a restart or a power cut leaves the rotation intact. A write that arrives while the sign is unreachable is refused with a 503 rather than silently held, so the caller learns it did not land.
  • Errors are errors. A dead serial link is a 503 and a message the sign cannot render is a 400, each with the reason in the body. Nothing here reports a failure under a 200.

Requirements

  • Python 3.11 or newer.
  • A BetaBrite Classic, reachable either at a serial device such as /dev/ttyUSB0 or over the network through an Ethernet to RS-232 adapter at socket://host:port.
  • Either a machine running systemd, for scripts/install.sh, or a container runtime, for the published image. The service itself runs anywhere Python does; only the installer is Linux specific.

Try it without a sign

loop:// is pyserial's loopback, so the service will start and serve its API with nothing attached. From a checkout:

pip install -e ".[dev]"
READERBOARD_SERIAL_URL=loop:// READERBOARD_API_KEY=dev-key \
    READERBOARD_STATE_PATH=./state.json \
    python -m readerboard

Or without one:

docker run --rm -p 5001:5001 \
    -e READERBOARD_SERIAL_URL=loop:// -e READERBOARD_API_KEY=dev-key \
    ghcr.io/mjaksn/readerboard:latest

Then open http://127.0.0.1:5001/docs.

loop:// swallows everything written to it, so the service runs but there is nothing to see. To watch what it would have sent, run it against the sign simulator in tools/signsim/ instead:

pip install --require-hashes -r tools/signsim/requirements.lock
python scripts/run_with_simulator.py

That starts the simulator and the service together, already pointed at each other, and stops both on Ctrl+C. The simulator decodes each transmission, says what every byte of it means, and shows what the sign would be holding as a result. tools/signsim/README.md has the details.

Running it against a real sign

From a checkout, with the sign on a cable or on an Ethernet to RS-232 adapter:

pip install -e ".[dev]"
pip install --require-hashes -r tools/apiclient/requirements.lock
python scripts/run_against_a_sign.py --serial-url socket://192.168.2.51:4001

That starts the service and the client together, with no simulator. The service comes up on http://127.0.0.1:5001 with /docs beside it, the client comes up pointed at that address, and the API key to paste into the client is printed in the same window. --no-client leaves the client out. Ctrl+C stops everything, and closing the client leaves the service running.

Both editors carry it as a launch configuration named "readerboard against the real sign and the client". The sign's address is an argument in those, not a setting in a file, so changing which sign is driven means editing the Parameters field in PyCharm's run configuration dialog, or args in .vscode/launch.json. They also pass --api-port 5002, so a second checkout of this repository on the same machine can run beside them; the launcher checks that port before it starts anything rather than letting the service bind, fail and stop after the client has been pointed at whatever else answered.

Writing the address

It is a pyserial URL, and there is no slash between the host and the port. socket://192.168.2.51/:4001 looks close enough to right and is not: pyserial answers it with a bare TypeError from deep inside a connection attempt, naming neither the setting nor the value. The launcher checks the address before it opens anything and says which part is wrong. The four forms are:

socket://192.168.2.51:4001   an Ethernet to RS-232 adapter passing raw TCP
rfc2217://192.168.2.51:23    an adapter speaking the telnet serial protocol
COM3                         a cable on Windows
/dev/ttyUSB0                 a cable on Linux

Most adapters pass raw TCP, so try socket:// first. If the link opens but the sign shows nothing or shows rubbish, and the adapter answers on port 23, it is probably negotiating telnet rather than passing bytes through, and rfc2217:// is the form that speaks that.

The API key, and config.local.toml

The key is not an argument. A launch configuration is a tracked file and a command line is a shell history, and anyone holding the key can write to the sign. It lives in config.local.toml at the root of the checkout, which .gitignore covers and which the launcher writes with a generated key the first time it runs. Given no --serial-url, the address is read from there too.

The first run erases the sign

Writing a memory configuration erases every message on the sign, and the service writes one whenever it has no record of the configuration already applied. The first run against a sign this machine has never driven therefore erases it, which is also the only way to allocate the files it then writes into. Every run after that reads the record and leaves the sign alone.

That record is .local-sign-state.json, and it belongs to this launcher alone. scripts/run_with_simulator.py deletes its own .local-state.json on every launch, because the simulator starts empty every time and the service has to reconfigure it. If the two shared one file, a simulator session would throw the sign's record away and the next run against the sign would erase it.

Installing it properly

Two ways, which do the same job. Pick whichever suits the machine.

With systemd

sudo scripts/install.sh --serial-url socket://192.168.2.51:4001

This creates a readerboard system user, builds a virtual environment in /opt/readerboard, writes /etc/readerboard/config.toml with a freshly generated API key, and enables the readerboard service. It prints the key once, and it is safe to run again after pulling a new version: your config file and key are left alone.

sudo scripts/uninstall.sh removes the service and the program but keeps your config and your registered messages, so reinstalling puts the sign back as it was. Add --purge to remove those too.

With Docker

The image is published to both registries on every release, for linux/amd64, linux/arm64 and linux/arm/v7, so a Pi pulls the same tag an x86 server does.

docker run -d --name readerboard --restart unless-stopped -p 5001:5001 \
    -e READERBOARD_SERIAL_URL=socket://192.168.2.51:4001 \
    -e READERBOARD_API_KEY=YOUR-KEY \
    -v readerboard-state:/var/lib/readerboard \
    ghcr.io/mjaksn/readerboard:latest

packaging/docker-compose.yml is the same thing as a Compose file, with the settings worth knowing about written out beside it.

The volume is what matters here. The registered messages are persisted to /var/lib/readerboard, and without it the sign comes back empty after a restart rather than putting back what was on it.

Every setting is available as an environment variable, so no config file is needed. Mount one at /etc/readerboard/config.toml if you would rather have it, in the format packaging/config.example.toml documents; the environment still wins over the file.

For a sign on a cable rather than on the network, the container needs the device passed in and needs to be in the group that owns it. The group has to be given as a number, because the container has no /etc/group entry for the host's dialout:

stat -c '%G %g' /dev/ttyUSB0        # 20 on Debian and Raspberry Pi OS, 18 on Fedora
docker run ... --device /dev/ttyUSB0 --group-add 20 \
    -e READERBOARD_SERIAL_URL=/dev/ttyUSB0 ...

Using it

Every write needs an X-API-Key header, and so does GET /sign/information, which asks the sign a question rather than reading the service's own record. The service's other reads and GET /health do not. In the Swagger UI at /docs, the Authorize button puts it in once for the whole page.

Register a message:

curl -X PUT http://localhost:5001/messages/temperature \
     -H 'X-API-Key: YOUR-KEY' -H 'Content-Type: application/json' \
     -d '{"message": "<green>18.4<degree> <red><time>", "display_mode": "HOLD"}'

Register a second one and the sign rotates between them:

curl -X PUT http://localhost:5001/messages/doorbell \
     -H 'X-API-Key: YOUR-KEY' -H 'Content-Type: application/json' \
     -d '{"message": "<amber>Someone at the door", "ttl_seconds": 300}'

Show a live value. Create the variable first, then a message that calls it:

curl -X PUT http://localhost:5001/variables/temp \
     -H 'X-API-Key: YOUR-KEY' -H 'Content-Type: application/json' \
     -d '{"value": "72", "ttl_seconds": 1800, "stale_value": "--"}'

curl -X PUT http://localhost:5001/messages/weather \
     -H 'X-API-Key: YOUR-KEY' -H 'Content-Type: application/json' \
     -d '{"message": "Outside <var:temp><degree>F", "display_mode": "ROTATE"}'

From then on only the variable needs writing, and the message picks each new value up on its next pass across the sign.

Take the sign over for thirty seconds:

curl -X POST http://localhost:5001/alerts \
     -H 'X-API-Key: YOUR-KEY' -H 'Content-Type: application/json' \
     -d '{"message": "<red><flash_on>SMOKE ALARM", "ttl_seconds": 30}'

Make a noise, which is worth pairing with an alert if the sign is somewhere nobody is watching it:

curl -X POST http://localhost:5001/sign/command \
     -H 'X-API-Key: YOUR-KEY' -H 'Content-Type: application/json' \
     -d '{"command": "SOUND", "parameter": "BEEPS"}'

BEEPS is three short beeps and TONE is one continuous tone of about two seconds. Those are the only two sounds there are: the sign has a fixed-pitch buzzer, so there is no pitch or volume to choose.

Silence it with {"command": "SPEAKER", "parameter": "OFF"}, and turn it back on with ON. That is a real mute: SOUND is still accepted and makes no noise. The setting lives on the sign and survives a restart, so it is also the first thing to check if SOUND ever seems to do nothing.

The full API is at /docs. Every markup token, value token, display mode and control command is listed by the /enumerations reads there, which answer at request time rather than being frozen into the description.

Writing messages

A message is plain text plus tokens written as <name>: <green>18.4<degree> is a colour change, a number, and a degree symbol. GET /enumerations/markup-tokens lists them all.

Text is encoded against the sign's own character table rather than as UTF-8, so café displays correctly. A character the sign cannot render is rejected with a 400, as is an unknown token: a write is told what the sign would have made of it rather than being shown something it did not ask for.

Live values: variables

A variable lives in a small file of its own on the sign, and a message or an alert calls it with <var:name>. Writing a new value rewrites that file and nothing else, which the sign takes without blanking, so a message showing a temperature or a count can change every minute and never restart. In a scrolling mode the new value appears on the message's next pass. An alert carrying a live wind speed works the same way, and keeps the sign while the number changes.

A few things are worth knowing, all of them measured on the sign:

  • Create the variable before a message calls it. A message or an alert naming a variable that does not exist is refused with a 400, and a variable that a message or the alert still calls cannot be deleted: that is a 409 naming what calls it.
  • Formatting in a value carries on after it. A value of <red>DOWN turns the rest of the message red as well, and so does a character set or a speed. If the text after the call matters, set it again in the message after the call.
  • A value takes the message markup, less two tokens. <week_day> draws as a literal 9 from inside a variable, and a variable cannot call another. GET /enumerations/value-tokens lists what is allowed.
  • Keep a changing number from pushing the text around it. With the sign's usual proportional spacing, 11 and 88 are different widths and the text after them moves. Put <fixed_width> in the message before the call and send values of the same length. Fixed width also left-justifies the line.
  • A value that stops arriving can go stale. Give a ttl_seconds and a stale_value such as --, and once the time passes without a new value the sign shows the stale value instead. The variable itself stays, since messages call it.
  • A value that does not fit is refused. Each holds variable_capacity bytes after rendering, 32 by default. The sign does not cut an overlong value short, it empties it, so the service refuses one rather than send it.

A Home Assistant rest_command that keeps a sensor on the sign:

rest_command:
  sign_temperature:
    url: http://readerboard.local:5001/variables/temp
    method: put
    headers:
      X-API-Key: !secret readerboard_key
    content_type: application/json
    payload: '{"value": "{{ value }}", "ttl_seconds": 1800, "stale_value": "--"}'

Call it from an automation triggered by the sensor, with the reading as value:

actions:
  - action: rest_command.sign_temperature
    data:
      value: "{{ states('sensor.outside_temperature') | round(0) | int }}"

Recovering a sign that has stopped responding

A sign mounted out of reach can wedge: a stray bit corrupts what its decoder is showing, it stops responding to writes, and there is no power switch within reach. There are two recoveries, and they are not interchangeable. Try the gentle one first.

A soft reset restarts the sign and erases nothing. The sign runs the same power-up diagnostics it runs when you plug it in, then carries on showing what it was showing. Its memory, its file table and its messages all survive; this was verified on the sign by reading them back either side of a reset.

curl -X POST http://localhost:5001/sign/command \
     -H 'X-API-Key: YOUR-KEY' -H 'Content-Type: application/json' \
     -d '{"command": "SOFT_RESET"}'

The call waits out the diagnostics before answering, so a 204 means the sign is listening again rather than that the bytes went out.

POST /sign/reboot is the escalation, and it is destructive. It clears the sign outright, waits for it to restart, then re-pushes every message and the run sequence from the service's own record, so the display still comes back to what it was.

curl -X POST http://localhost:5001/sign/reboot -H 'X-API-Key: YOUR-KEY'

Reach for it only when a soft reset was not enough. The sign is blank for about ten seconds while it resets. Neither is a way to clear messages: DELETE /messages does that without resetting anything. The client fronts the reboot with a warning-coloured confirmation for the same reason.

Configuration

Settings come from /etc/readerboard/config.toml, overridden by environment variables prefixed READERBOARD_. packaging/config.example.toml documents every one of them.

Every setting has both forms, and the container path relies on it: slot_count in the file is READERBOARD_SLOT_COUNT in the environment. Under Docker the file is optional and usually absent, which is not an error. READERBOARD_CONFIG_FILE moves the file if you want it somewhere other than the default.

The sign's address is a full pyserial URL in serial_url: socket://192.168.2.51:4001 for an Ethernet to RS-232 adapter, rfc2217://192.168.2.51:23 for one speaking the telnet serial protocol, /dev/ttyUSB0 or COM3 for a cable plugged straight in, or loop:// to run the service with no sign attached. There is no slash between the host and the port, and pyserial's answer to one that has a slash names neither the setting nor the value.

Four settings reallocate the sign's memory when changed, and that erases every message on it: slot_count, slot_capacity, variable_count and variable_capacity. The service will do it, and say so loudly in the log, but they are not settings to fiddle with. With variable_count at 0, variable_capacity allocates nothing, so changing it alone reallocates nothing either.

Security

An API key is required on every write, compared in constant time, and never logged. GET /sign/information needs one too: it is a read of the sign itself rather than of the service, so it sends a question over the wire, holds the sign until the answer arrives, and reports the hardware's firmware and how full its memory is. The service's own reads and GET /health need none, so a monitor can watch the slots without holding a key that could write to them.

The key is declared to the API description as a security scheme, so the Swagger UI at /docs has an Authorize button: enter the key once and everything on the page that needs it carries it. It is the same X-API-Key header a client sends, so nothing about a script or a Home Assistant rest_command changes.

That page is configured to remember the key, so it survives a reload or a browser restart rather than needing to be pasted in again. Convenient on your own machine, and worth knowing before you use Authorize on a shared or kiosk browser, where the next person to open /docs inherits it. Use the browser's Logout in the Authorize dialog, or just do not authorize there.

Message content reaches the sign as protocol bytes, so it is worth knowing what a client holding the key can do. The markup renderer emits bytes only for tokens it recognises and for characters in the sign's own table, so arbitrary control sequences cannot be injected through a message. What the holder of a key can do is display anything they like on your wall and set the sign's clock. There is nothing beyond the sign to reach: the service opens one serial link and touches nothing else.

Sensible precautions remain sensible:

  • Do not expose the service to the internet.
  • Keep /etc/readerboard/config.toml mode 0640. Anyone who can read it can write to the sign.
  • Give the key only to clients you trust, and prefer a firewall allow-list on top.
  • The service runs as a dedicated system user under a hardened systemd unit, which is worth keeping rather than running it as root for convenience.

Under Docker the same points apply, with different mechanisms:

  • The image runs as an unprivileged user, UID and GID 10001, not as root. A bind-mounted state directory has to be owned by that number on the host.
  • An API key passed as an environment variable is visible to anyone who can run docker inspect on the container, and to anything that reads the Compose file's environment. Mounting a config file mode 0640 keeps it out of both.
  • Bind the published port to the loopback address, -p 127.0.0.1:5001:5001, unless clients on other machines need to reach it.
  • Passing a serial device in with --device gives the container that device and nothing else. It does not need --privileged, and giving it that would hand it every device on the host.

Development

pip install -e ".[dev]"
pytest
ruff check .
mypy readerboard

No sign is needed. The tests run against a capturing fake transport and against pyserial's loop:// URL, so the real serial code path is exercised without hardware.

docs/protocol-notes.md records what the Alpha protocol actually says about the memory configuration, the run sequence and the priority file, with the quotations that back each claim. Read it before changing anything in readerboard/protocol/.

scripts/protocol_spike.py settles the few questions the document cannot answer about this particular sign. It is destructive and refuses to run without --confirm-erase.

tools/signsim/ is the sign simulator, a PySide6 stand-in for the sign, described above. tools/apiclient/ is the client, a PySide6 application for calling the API by hand: every endpoint, responses shown as text rather than JSON, and a vocabulary it loads from the service rather than one compiled into it. The tests of both are collected by the pytest run here and need no Qt installed; the applications do, and each is pinned separately so that nothing the service installs ever pulls Qt in.

scripts/run_with_simulator.py starts the service and the simulator together, and with --with-client the client as well, so the whole loop comes up from one command. Both editors carry it as a launch configuration under the same name, "readerboard and the sign simulator", in .vscode/launch.json and in .idea/runConfigurations/, beside configurations for running the pieces separately. Both carry the three way one as "readerboard, the sign simulator and the client" as well.

scripts/run_against_a_sign.py is the other one, for when the sign is real: the service and the client, no simulator, and the sign's address passed as an argument so that it can be edited in a run configuration dialog. Both editors carry it as "readerboard against the real sign and the client". The section above has the rest, including the one thing about it that is dangerous. The two launchers share their process supervision through scripts/_supervise.py and differ in what each child is given, which is the part that matters: the simulator launcher discards its state file on every run and this one never discards anything.

Every one of those that starts the service sets READERBOARD_OPEN_DOCS, so /docs opens in a browser once the port answers. The service does the waiting and the opening, which is why it lands on the port actually bound rather than one repeated in a launch file. The setting is off unless asked for, so an installed service opens nothing.

Licence

MIT. See LICENSE.

Credits

The protocol is documented in the Alpha Sign Communications Protocol, form 9708-8061, published by Adaptive Micro Systems. Every byte value in readerboard/protocol/constants.py is transcribed from that document, and tests/test_constant_values.py pins each one against it with a citation per assertion.

An earlier version of this project took that table from jonathankoren/readerboard, which is recorded here with thanks even though no code from it remains.

Metadata

Release files for readerboard 0.5.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for readerboard 0.5.0
File Size Uploaded
readerboard-0.5.0.tar.gz 175.5 kB Details

Built distribution (wheel)

Table of built distributions (wheels) for readerboard 0.5.0
File Interpreter ABI Platform
readerboard-0.5.0-py3-none-any.whl Python 3 none any Details

Total release size: 282.5 kB

Release files / readerboard-0.5.0.tar.gz

Download URL readerboard-0.5.0.tar.gz
Size 175.5 kB
Tags Source
SHA-256 checksum
How to use checksums
9c0a924e9a33d220683bc5c44b25bc4c7ed4be4f80e530495858e013745e331b
BLAKE2b-256 checksum
How to use checksums
90edfd9da59938cca44b7e08081099c3413dc994449b4a496e8860a651431413
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 11, 2026.

Transparency log

Release files / readerboard-0.5.0-py3-none-any.whl

Download URL readerboard-0.5.0-py3-none-any.whl
Size 107.0 kB
Tags Python 3
SHA-256 checksum
How to use checksums
b5aff55e5dd34ebc0ee4f07cfe558c70d3907bbc9b71e4f846cc5ae007e171bc
BLAKE2b-256 checksum
How to use checksums
1c3d2f6fcc7ea8f67c90fc0223d6e793e7ac8071709f98b28b38a167983e8e7b
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 11, 2026.

Transparency log

Release history Release notifications | RSS feed

0.7.0

2 release files

0.6.0

2 release files

This release

0.5.0 This release

2 release files

0.4.0

2 release files

0.3.0

2 release files

0.2.0

2 release files

0.1.4

2 release files

0.1.3

2 release files

0.1.2

2 release files

0.1.1

2 release files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page