🧠 NexusAI

Local LLM inference on a PS5, with an authenticated model-control API, an agent orchestrator and an optional Hermes engine.

Source: ljcatherine/NexusAI (release 1.0.0) · agent-orchestrator · integrations/hermes

NexusAI is the AI side of Nexus. A custom PS5 app runs the model on the console’s GPU, the agent orchestrator on Debian turns chats into durable agent runs, and the gateway exposes both to the apps behind Cloudflare Access.

Apps ──► gateway (Debian) ──► orchestrator :8897 ──► NexusAI PS5 app :8081 (inference)
                       │                        └──► :8082 (model control)
                       └──────────────────────────► :8081 (direct chat)

The NexusAI PS5 app

A reproducible GPL-3.0-or-later fork of PS5LM Plus (itself built on Mert Cobanov’s PS5LM), kept in this ecosystem as patches plus new sources rather than a copied tree. It ships as a separate PS5 homebrew title with its own title ID, so it installs next to the original upstream app instead of replacing it.

Original upstream app NexusAI PS5 app
Inference llama-server on :8081 (OpenAI-compatible) Unchanged
Branding PS5LM NexusAI dashboard
Model control On-console only Opt-in authenticated HTTP API on :8082
Model metadata Basic Exact file size, architecture, GGUF quantisation type
Numeric JSON Locale-sensitive parse (could abort) Fixed

Model-control API (:8082)

Route Result
GET /v1/nexus/status State, active model, context used/capacity, measured speeds and memory, load progress
GET /v1/nexus/models Discovered models: stable 16-hex ID, label, drive, fits/current, size, architecture, quantisation
POST /v1/nexus/models/{id}/load 202: queues a GPU-safe restart into that model
POST /v1/nexus/models/unload 202: queues a restart with no model

Every call needs Authorization: Bearer <token>. The console stores only the token’s SHA-256 (nexus-control.token, sha256:<64 hex>, mode 0600). The raw token lives only in the Debian configuration and the iPhone Keychain.

The controller is off without the token file and binds 127.0.0.1:8082 unless the empty nexus-control.lan marker exists. A model switch restarts the app to free GPU memory, so poll /status until the new model is ready.

Build and install

You need a PS5 that can run homebrew and the upstream PS5_Vulkan toolchain and SDK. Nexus doesn’t provide console exploits or homebrew prerequisites.

git clone https://github.com/ljcatherine/NexusAI.git && cd NexusAI
./prepare-build.sh                                   # pinned upstream + NexusAI patches -> build/nexusai
python3 prepare-test-identity.py …                   # pick an unused PPSA##### title ID
# build inside build/nexusai with the upstream scripts/build-app.sh
python3 verify-test-title.py build/app-ps5/title/<TITLE_ID>
./test-control.sh                                    # host-side controller tests

Rules for every install:

  1. Back up first: the original title folder and the PS5LM state folder (/data/PS5LM, or /download0/PS5LM when sandboxed).
  2. Use a different title ID from the original, and never run both at once, because they share the state folder.
  3. Fix permissions: eboot.bin and sce_module/libc.prx must be 0755. Normalise before transfer and check again after install. The verifier records modes as well as hashes.
  4. Reinstall, don’t overwrite: close the title, remove the previous test install with supported operations, and install the new build fresh. Keep rollback copies and shared models and settings.

Turn on control

  1. Generate a 32–128 character token on a trusted machine (openssl rand -hex 32).
  2. Write only sha256:<digest> to nexus-control.token in the PS5LM state folder (0600).
  3. Create nexus-control.lan only once the Debian firewall and gateway are ready.
  4. Check: no token → 401; with token → /status and /models show real models.

:8081 and :8082 are plain HTTP. Keep them on your LAN, firewall :8082 to the Debian server, and never port-forward them.

Roll back

Stop the test title, keep app.log and llama.log, remove the token and LAN marker, restore the state folder from backup if needed, and launch the original title. Never delete the original title.

Agent orchestrator

A Python service on Debian (127.0.0.1:8897) that runs durable agent jobs against the PS5 and the workspace agents.

  • Jobs are stored owner-only and survive restarts.
  • Context capacity is read from the loaded model; NEXUS_CONTEXT_TOKENS can lower it.
  • Before each PS5 stream it writes an inference lease. A crash or a stream that ends without its completion marker quarantines inference until you confirm the PS5 is idle: clear-inference-quarantine.py "$RUNS_DIR" --confirmed-ps5-idle.
  • Patches and commands are proposals that need human review in the app. Workspace writes and command execution need separate grants.
  • Optional native model switching through NEXUS_PS5_CONTROL_URL and its own token.

Setup is in Debian → orchestrator.

Hermes engine (optional)

An alternative agent engine the orchestrator can use, built as an isolated container worker with a pinned base image. Build and verification steps are in integrations/hermes.

Using it from the apps

App What you can do
iPhone Chat (direct or via the gateway), switch PS5 models, start and review agent runs, web research
Linux / Windows Desktop Chat with NexusAI as a provider, see PS5 status
Nexus Online Chat, see models, status and agent runs (read-only)

For direct LAN use, point any app at http://<ps5-address>:8081/v1. For remote use, go through the gateway, which only allows models in NEXUS_AI_ALLOWED_MODELS.

Edit this page on GitHub ↗