On-demand local LLM inference on a PS5.
PS5LM runs llama.cpp on the GPU of a jailbroken PS5 and serves an OpenAI-compatible API (about 45 tok/s on Qwen 3.5 9B Q6_K). Keeping a console on all day for that wastes power and makes noise. This integration starts it when you need it and sends it back to rest mode when you are done.
- Wakes the PS5 from rest mode, signs the user in, starts PS5LM and waits until the model is loaded (about 2 minutes)
- Stops PS5LM, or puts the whole console back to rest mode
- Switches the loaded model: the dropdown lists the models on the console and restarts PS5LM with the one you pick
- Exposes PS5LM and console metrics as sensors
- Everything is a switch or a service, so automations can do it
- Wake packet over the Remote Play discovery port (UDP 9302)
- A Remote Play control connection signs the user in. A console woken from rest mode sits at the "press the PS button" screen and cannot launch apps until someone is signed in.
- ps5-payload-websrv (loaded via Payload Manager) launches PS5LM
- PS5LM's
/healthis polled until the model answers - Rest mode is sent as a Remote Play control message
A model switch writes PS5LM's app-args.txt through the console's web file manager (port 8888, loaded via Payload Manager) and restarts the app, because a model cannot be swapped in place. A tuned command line per model can be kept as /data/PS5LM/profiles/<model>.args and is then used as it is; otherwise the current one is reused with the new model and a smaller context.
No cloud and no extra Python packages. The Remote Play part is a small Python re-implementation of what chiaki-ng does; chiaki-ng is only needed once, to register the console.
- A jailbroken PS5 with Payload Manager installed
- PS5LM application installed on the console
- ps5-payload-websrv payload loaded via Payload Manager (port 8080)
- ps5-payload-filemanager payload loaded via Payload Manager (port 8888)
- The console registered for Remote Play in chiaki-ng
- On the PS5: Enable Turning On PS5 from Network (Settings > System > Power Saving > Features Available in Rest Mode) and Log In to PS5 Automatically (Settings > Users and Accounts > Login Settings)
HACS: add https://github.com/phoinixgrr/ps5lm_ha as a custom repository (category Integration), install PS5LM, restart.
Manual: copy custom_components/ps5_launcher/ to your custom_components/ and restart.
-
Get the Remote Play keys from chiaki-ng:
python3 tools/chiaki_keys.py -
Settings > Devices & Services > Add Integration > PS5LM: enter the PS5 host and the two keys
-
The second screen has defaults for everything; change them only if your setup differs
| Option | Default |
|---|---|
| PS5LM title ID | PPSA99575 (PPSA99581 for PS5LM Plus) |
| PS5LM API URL | http://<host>:8081 |
| websrv port | 8080 |
| ps5-exporter port | 9100 |
| Update interval | 30 s |
| Console metrics interval | 120 s |
| Model server / console metrics | on |
Everything except the host can be changed later under Configure.
| Entity | |
|---|---|
| Console switch | on = awake, off = rest mode |
| Server switch | on = start PS5LM (wakes the console if needed), off = stop PS5LM |
| Status | Rest mode, Waking the console, Signing in, Starting PS5LM, Loading the model, Model ready, Starting the fleet worker, Fleet worker ready, ... |
| Model ready | PS5LM answers its health check. Always off for a fleet worker, which has no model API |
| Model select | the models on the console; picking one restarts PS5LM with it (about 2 minutes) |
| Model metrics | model, context size, generation and prompt speed, active and queued requests, token counters, prompt cache hit rate |
| Console metrics | CPU and SoC temperature, fan speed, PS5LM memory, uptime, free space |
Services: ps5_launcher.wake, ps5_launcher.standby, ps5_launcher.start_app, ps5_launcher.stop_app.
Entity IDs below are for a console added as ps5.home; check yours on the device page. Replace notify.notify and the conversation agent with your own.
A script other automations can call: starts PS5LM if needed and returns once the model answers (fails after 5 minutes).
script:
ps5lm_ready:
alias: PS5LM ready
mode: queued
sequence:
- action: switch.turn_on
target:
entity_id: switch.ps5lm_ps5_home_server
- wait_template: "{{ is_state('binary_sensor.ps5lm_ps5_home_model_ready', 'on') }}"
timeout: "00:05:00"
continue_on_timeout: falseWakes the console, asks the local model, sends the answer, and puts the console back to rest mode only if it was asleep before.
- alias: PS5LM morning briefing
triggers:
- trigger: time
at: "07:30:00"
actions:
- variables:
was_asleep: "{{ is_state('switch.ps5lm_ps5_home_console', 'off') }}"
- action: script.ps5lm_ready
- action: conversation.process
data:
agent_id: conversation.your_ps5lm_agent
text: "Two short lines: today's weather and anything in the house that needs attention."
response_variable: answer
- action: notify.notify
data:
title: Good morning
message: "{{ answer.response.speech.plain.speech }}"
- if: "{{ was_asleep }}"
then:
- action: switch.turn_off
target:
entity_id: switch.ps5lm_ps5_home_consoleTwo hours after the last generated token, counting every client (Assist, opencode, the web chat).
- alias: PS5LM idle to rest mode
triggers:
- trigger: time_pattern
minutes: "/10"
conditions:
- condition: state
entity_id: binary_sensor.ps5lm_ps5_home_model_ready
state: "on"
- condition: numeric_state
entity_id: sensor.ps5lm_ps5_home_active_requests
below: 1
- condition: template
value_template: >
{{ now() - states.sensor.ps5lm_ps5_home_tokens_generated.last_changed > timedelta(hours=2)
and now() - states.switch.ps5lm_ps5_home_console.last_changed > timedelta(hours=2) }}
actions:
- action: switch.turn_off
target:
entity_id: switch.ps5lm_ps5_home_consoleIt checks every 10 minutes rather than waiting with for:, because a Home Assistant restart cancels a for: wait and an
idle model never changes the counter again. It only acts while the model is loaded: with PS5LM stopped the counter is
unavailable, which would otherwise count as idle and send the console to rest mode in the middle of a game.
- alias: PS5LM off when the house is empty
triggers:
- trigger: numeric_state
entity_id: zone.home
below: 1
for: "00:15:00"
conditions:
- condition: state
entity_id: switch.ps5lm_ps5_home_console
state: "on"
actions:
- action: switch.turn_off
target:
entity_id: switch.ps5lm_ps5_home_consoleStops PS5LM if the hottest SoC sensor stays above 85 °C, and tells you.
- alias: PS5LM too hot
triggers:
- trigger: numeric_state
entity_id: sensor.ps5lm_ps5_home_soc_temperature_hottest
above: 85
for: "00:02:00"
actions:
- action: switch.turn_off
target:
entity_id: switch.ps5lm_ps5_home_server
- action: notify.notify
data:
message: >-
PS5LM stopped: SoC at {{ states('sensor.ps5lm_ps5_home_soc_temperature_hottest') }} °C,
fan {{ states('sensor.ps5lm_ps5_home_fan_speed') }} %.- alias: PS5LM model ready
triggers:
- trigger: state
entity_id: binary_sensor.ps5lm_ps5_home_model_ready
from: "off"
to: "on"
actions:
- action: notify.notify
data:
message: "{{ states('sensor.ps5lm_ps5_home_model') }} is ready."
- alias: PS5LM start or stop failed
triggers:
- trigger: state
entity_id: sensor.ps5lm_ps5_home_status
attribute: last_result
conditions:
- "{{ (trigger.to_state.attributes.last_result or '').startswith('failed') }}"
actions:
- action: notify.notify
data:
message: "PS5LM: {{ trigger.to_state.attributes.last_result }}"PS5LM v0.11 can run as an RPC worker for another machine's llama.cpp (fleet docs): put /data/PS5LM/fleet-worker.txt on the console and the app serves no model API at all, only an RPC endpoint (the file's first line, ip:port). The integration reads that file and adapts:
- A start is ready when the RPC endpoint accepts a connection, not when
/healthanswers (it never does). Status goes Starting the fleet worker, then Fleet worker ready. - Model ready stays off, so automations that treat it as "the model API can be used or closed" (an idle rest mode, a job stop) leave a worker alone.
- The Model dropdown refuses to switch, with a message: a worker has no model of its own. Delete
fleet-worker.txtand restart PS5LM to go back to a normal server. - A
fleet-worker.txtwithout a validip:portfirst line fails the start at once instead of waiting 7 minutes. - Stop and Rest mode cancel a start that is still waiting for the model or the worker, instead of being refused until its 7 minute timeout ends.
An unauthenticated RPC port is probed (a plain TCP connect, no data) while a worker runs, about once per scan. Keep workers on a trusted network, as the fleet docs require.
- Tested with PS5LM v0.9.1, websrv v0.34, filemanager v1.12 and ps5-exporter v0.2.0 on a CFI-1216A, Home Assistant 2026.10
- The console handles one Remote Play session at a time. While chiaki-ng, the PS Remote Play app or a previous session is connected, rest mode (and the sign-in after a wake) is refused; the integration retries for up to 90 s and then shows "Failed: Remote Play is in use" in the Status sensor for 5 minutes. Close the other app and try again.
- Reading the console metrics makes the console stall the model for a few seconds (the exporter walks the whole process table), so they are read only every 2 minutes and never while a request is running. Lowering the interval makes answers stutter.
- A full power-off ends the jailbreak; etaHEN and websrv then have to be loaded again by hand
- websrv and the PS5LM API have no authentication. Keep the console on a trusted network.
Built on PS5LM, chiaki-ng, ps5-payload-websrv and ps5-exporter.
AGPL-3.0, because the Remote Play code is derived from chiaki-ng. Not affiliated with Sony Interactive Entertainment.