BANSOS Free Models for VS Code / Copilot Chat
Free OpenAI-compatible models (OpenCode Zen + KiloCode gateway) added directly to the VS Code and Copilot Chat model picker — no API key, no external plugins (Continue / Cline / Roo not required).
Аналог pi-bansos (пакет для pi), перенесённый на VS Code Language Model Chat Provider API.
What it does
- Contributes a
bansos vendor to the chat model picker via contributes.languageModelChatProviders — the free models appear next to Copilot's own models.
- On startup it health-checks both upstream catalogs and registers only currently listed models (dead/promo-expired ones are skipped silently).
- Streaming responses (SSE), tool calling (agent mode), vision (image input where supported), Responses API for Muse models, Chat Completions for everything else.
- Relay egress (optional): route requests through your own Vercel/Cloudflare relay to dodge per-IP rate limits — toggle live, no restart.
- Local rate guards matching upstream quotas (KiloCode: 200 req/h/IP; OpenCode: daily guard).
Models
Up to 27 free models: 8 OpenCode Zen (incl. Muse Spark free via Responses API) + 19 KiloCode gateway (keyless). The exact list shown in the picker depends on which models are alive at startup — run BANSOS: Refresh free model list to re-check.
Install
From source (development)
npm install
npm run build
# then press F5 in VS Code ("Run BANSOS Extension" launch config)
As a .vsix
npm install
npm run package
code --install-extension vs-bansos-0.1.0.vsix
Restart VS Code, then open Chat and pick a model from the dropdown — entries are prefixed with OpenCode · or KiloCode ·.
Commands
| Command |
What it does |
BANSOS: Manage (relay / models) |
Menu: relay on/off/status, switch/remove saved relays, set URL, deploy a fresh Vercel relay, refresh models |
BANSOS: Refresh free model list |
Re-runs the catalog health check and updates the picker |
The status bar shows $(sparkle) bansos: free models (direct) or bansos: relay ON.
Relay (optional)
By default requests go directly to the upstreams. If your IP gets rate-limited:
Deploy fresh Vercel relay… — paste a Vercel API token (used once, never stored); the relay is deployed (~10–40 s), saved, and activated.Set relay URL… / Switch relay… — use any HTTP relay (Vercel, Cloudflare, Deno, your own). The x-relay-target / x-relay-path pattern is used; the deployed worker only allows opencode.ai and api.kilo.ai.
State is persisted in the extension's globalStorage (relay-state.json) and survives restarts.
Notes
- Free upstream models are best-effort: promos expire, model IDs change, rate limits apply.
- KiloCode gateway: keyless, 200 requests/hour per IP.
- OpenCode free tier requires the client fingerprint (UA/session/tool gates); the extension maintains it the same way pi-bansos does.
- The local rate guard counts requests per process (one VS Code window).
Comparison with pi-bansos
| pi-bansos (pi) |
vs-bansos (VS Code) |
Local proxy on 127.0.0.1:18080 |
No proxy — extension host calls upstreams directly |
pi.registerProvider + /model picker |
lm.registerLanguageModelChatProvider + Copilot Chat picker |
/bansos slash command |
BANSOS: Manage… command palette + status bar |
Relay state in ~/.pi/agent/… |
Relay state in extension globalStorage |
| Health check at startup |
Same, plus refresh command |
License
MIT
