Free models, no API key — and the glue code actually sweats

What it does It’s a plugin for dsh (DeepSeek Harness) that adds a model provider with exactly one upstream: OpenCode’s Zen gateway at https://opencode.ai. Once installed, models like Muse Spark 1.3 and MiMo V2.6 appear in the model selector — no account, no API key, and, per the project, no usage cap. It also opens an OpenAI-compatible local port (/v1/models, /v1/chat/completions, /v1/responses) so other tools on the same machine can call those models, plus a usage dashboard that never leaves the machine.

The interesting bit Most free-model wrappers are three lines and a prayer; this one read the gateway’s error taxonomy. Availability is probed from your own network egress and classified honestly — a model the gateway names but refuses to route vanishes from the selector, while gateway 5xxs, 429s, and timeouts are explicitly not counted as the model being dead. The same care shows elsewhere: streaming is recognized by body shape rather than the content-type header (the gateway sometimes mislabels SSE frames under load), and “effort” is a real output-token budget — 2048/8192/model-max, doubled for models like MiMo V2.6 that can’t stop thinking — rather than an effort string tossed upstream.

Key highlights

• One upstream, stated openly: OpenCode’s Zen gateway, no third-party relay; the model list, context lengths, and capabilities are re-fetched on refresh instead of hardcoded.
• Region-limited models get their own selector group and migrate between groups after a re-probe — useful around VPN toggles.
• The in-app upgrader verifies SHA-256 checksums, backs up, swaps atomically, and hot-reloads, rolling back to the previous version if any step fails — no restart.
• Runs headless: only the llm service is a hard dependency, so compositions without a web server still get models; the dashboard mounts later on its own fiber.
• Zero dependencies, no build step, MIT license; status is beta.

Caveats

• “Completely free, no cap” describes the upstream gateway’s current behavior, not a contract — the plugin’s own design assumes the gateway can refuse, rate-limit, or region-block at any time.
• The announcement center is a push channel: the repo owner can publish whitelisted HTML to every install within a polling cycle, and urgent items take over the full screen. Constrained, but still remote content into your app.
• The README spends real estate on desktop-install pitfalls — a profile gate that refuses to start on stray symlinks, plus pnpm lockfile surgery — so the frictionless story mostly belongs to the CLI profile.

Verdict dsh users who want frontier models without the account-and-key ritual get an unusually well-made on-ramp here. Anyone outside dsh, or anyone needing availability guarantees, should remember that “free” means “opencode.ai currently allows it.”