opencode-neuron

OpenCode plugin for NeurOn, a lightweight control plane for shared self-hosted LLM capacity.

The plugin reserves NeurOn capacity before OpenCode sends a chat message. It waits until NeurOn reports the matching target healthy, then lets the request continue. After completions, it refreshes the same reservation to keep capacity warm without stacking long reservation tails.

Install

Install the package wherever OpenCode loads npm plugins:

npm install opencode-neuron

For project-local development, this repository also keeps the plugin at:

.opencode/plugins/neuron.js

Configuration

Required:

NEURON_API_KEY=sk-neuron-...

Optional:

NEURON_API_BASE_URL=http://localhost:8090
NEURON_RESERVATION_DURATION_MINUTES=2
NEURON_RESERVATION_KEEPALIVE_MINUTES=2
NEURON_WAIT_FOR_HEALTHY=true
NEURON_WAIT_TIMEOUT_SECONDS=600
NEURON_WAIT_POLL_SECONDS=5

Model Mapping

OpenCode model names are LiteLLM-facing names. NeurOn maps those names through configured model IDs, aliases, backend IDs, runtime IDs, and target litellmDisplayPrefix metadata.

If LiteLLM aliases a route prefix away, configure NeurOn with an empty display prefix for that target.

Runtime Warmup

Model warmup happens in NeurOn, not in this plugin. When configured, NeurOn keeps a target in provisioning until the requested reservation models have been warmed. The plugin simply waits for NeurOn's reservation status to become healthy.

License

AGPL-3.0-only.