# Arka public artifacts

Public, anonymous files for Arka AI. Private product code stays in other Gitea
repos.

## OpenCode Arka-AI provider

`arka-ai.ru` is not in models.dev. One command installs the OpenCode provider
plugin, companion skill, and (optional) API key:

```bash
curl -fsSL https://git.arka-ai.ru/arka/public/raw/branch/main/install-arka-provider.sh | bash
```

With a key and a default model:

```bash
curl -fsSL https://git.arka-ai.ru/arka/public/raw/branch/main/install-arka-provider.sh \
  | bash -s -- --key "$ARKA_API_KEY" --set-default DeepSeek-V4-Pro
```

Then open a **new shell** and restart opencode. `/models` should list **Arka-AI**
(~116 chat models).

The installer is offline and idempotent. It does not call `api.arka-ai.ru`.
`python3` is required only with `--set-default`.

### Files

| Path | Role |
|------|------|
| `install-arka-provider.sh` | Self-extracting installer (the one-liner) |
| `plugins/arka-ai.ts` | Provider plugin source (embedded in the installer) |
| `skills/arka-provider/SKILL.md` | Companion skill source (embedded in the installer) |
| `catalog.json` | Filtered chat catalog, refreshed daily at **05:00 Europe/Moscow** from `https://api.arka-ai.ru/v1/models` |

Forced keep-list: `DeepSeek-V4-Pro`, `deepseek-v4-flash-0731`, `qwen3.8-max`,
`kimi-k3`, `MiniMax-M3`, `GLM-5.3`.

Inference uses `ARKA_API_KEY` against `https://api.arka-ai.ru/v1`. Listing the
catalog does not require a key.

After install:

1. New shell so `ARKA_API_KEY` is exported (GUI/desktop launches may need the
   key set in that session, not only `~/.bashrc`).
2. Restart opencode, run `/models`.
3. Refresh later: `rm -f ~/.cache/opencode/arka-ai-models.json` and restart.