Let your AI applications share the machine

Use the model files you already have

OA's inventory finds model files across supported local stores. The model bridge can lend a compatible file to another engine using a link, with permission. The original file stays in its store. A successful unlend removes the placement OA owns; uncertain or replaced paths require reconciliation.

openabstractions models lend <store>/<id> --to lmstudio
openabstractions models lends
openabstractions models unlend <entry>

The link uses the existing file. The weights occupy disk space once. The receiving engine still needs to support that model's format and architecture. See the recorded LM Studio example or read the model bridge setup and recovery guide.

Ask a loaded model to make room for another task

An image workflow can request a memory reservation while a chat engine holds a model. The resource service checks permission and asks a supported holder to unload. It verifies the available capacity before admitting the new reservation.

The model host participates directly: it holds a lease for each resident model and can release that model when asked. Existing engines participate through supported unload operations. Active calls and a holder's refusal affect whether a request succeeds. See the prepared ComfyUI workflow and generated image.

The ComfyUI and LM Studio experiment exercised both a permitted yield and a denied request. A separate demonstration waited until Lemonade was idle, then asked Lemonade to unload its model before Comfy rendered. The run remains under qualification. See the handover and its limits. Measurement and process attribution vary by platform; Windows currently retains conservative reservations when process attribution cannot be established. A reservation governs admission; it does not enforce a process's physical memory usage.

Choose models in OpenCode through one provider

OA's router lists the models that registered engines can serve. The OpenCode integration uses the runtime for chat and tool calls. The operator controls which model servers and credentials that program may use. A live TUI demonstration refreshed the local model chooser in the same running process. In a separate API demonstration, one saved preference sent the first turn to Lemonade and the next fresh turn to Ollama after Lemonade became unavailable. OA recorded the actual host and model for both answers; the preference alias has not yet been shown in the rendered picker.

node adopters/opencode/configure.js configure --dry-run
node adopters/opencode/configure.js configure

The first command previews the provider configuration. The second writes it from the router's available chat models. Install the development OpenCode provider and configure an eligible model server first. OpenCode setup and test evidence · Saved-preference example.

Serve a local model file under a memory lease

The optional model host starts an existing llama-server executable for an explicitly selected model. It keeps the engine resident for repeated calls and coordinates its memory reservation with the resource service. The runtime mediates application calls and checks their rights.

The operator supplies the engine and model. Installation alone starts no model. Model host configuration and measured behavior.

Give an assistant a clear, permitted action

The OA MCP gateway makes OA tools available to an assistant. Registered applications can advertise their presence and supported interfaces. An application-specific adapter supplies actions such as previewing and applying a ComfyUI parameter change against a particular document revision.

The gateway has its own verified caller identity and grants. The ComfyUI presentation example checks the target and revision before applying a proposed change. Application interaction example · MCP gateway guide.