What problem does it solve? Setting up Actual Computer (actual.inc) as an inference provider involves two distinct modes (hosted relay and local daemon), device authorization that requires a human in a browser, and several cryptic failure modes like HTTP 400s from unsupported reasoning_effort values and empty streams from context-window overflow. This Skill walks through the correct configuration flow and documents the known pitfalls. ## Core Features & Use Cases - Relay/API Mode Setup: Store an ac_ key in ~/.zeus/.env, discover models via GET /v1/models, and set the built-in actual provider with zeus config set. - Local Daemon Mode: Search, download (with explicit quantization), and load GGUF models through the actual CLI, then point ZeusAgent at http://127.0.0.1:8080 for no-auth loopback inference. - Troubleshooting Guidance: Covers the reasoning_effort clamp (xhigh→high, ultra→max), context-window overflow on small local models, download-id vs installed-name mismatches, and empty-content reasoning models. - Use Case: A user with an Actual account wants ZeusAgent routed through their private GPU cluster; the Skill verifies the key, selects a model, and confirms end-to-end with zeus chat -Q -q "Reply with exactly: ACTUAL_OK". ## Quick Start Ask the agent to set up actual.inc as the inference provider using your ACTUAL_API_KEY and verify it with a test chat message.
