开源AI工作流

💰 OpenCode Zen paid models (requires balance)

Add a payment method to your Zen account to unlock these models at pay-as-you-go rates. OpenCode passes provider price drops through at cost. Prices per 1M tokens.

Set opencodego.freeOnly: false to show paid Zen models in the picker (default shows only free models).

<details> <summary><b>🔬 How model metadata is resolved (3-tier fallback)</b></summary>

Limits and capabilities resolve in this priority order:

1. Live metadata from OpenCode /models endpoint
2. 6-hour models.dev snapshot cached in VS Code globalState
3. Bundled fallback catalog shipped with the extension (works offline)

Deprecated/unavailable models are filtered before registration. Per-provider limits tracked separately (Go vs Zen) so shared models (e.g. glm-5.1, qwen3.6-plus) use correct values for each.

</details>

<details> <summary><b>🛣️ Endpoint routing per model family</b></summary>

All Qwen models use /chat/completions because they use OpenAI-native tool-calling format. Routing to Anthropic /messages broke tool calls.

</details>

---

⚡ Quick Start (60 sec)

1.  Install or update VS Code 1.125+ ────────────────────────────── ✓
2.  Install this extension ──────────────────────────────────────── ✓
3.  Get an OpenCode Zen API key → opencode.ai/auth ─────────────── ✓
4.  Open Copilot Chat → click model → "Add Models" → OpenCode Zen ── ✓
5.  Paste API key → pick a free model → CHAT 🎉

<details> <summary><b>📖 Detailed step-by-step with screenshots</b></summary>

1. Install or update VS Code to version 1.125 or newer. OpenCode BYOK chat works without a GitHub sign-in or Copilot plan. 2. Install this extension from the VS Code Marketplace (or press F5 in this repo for dev mode). 3. Get an API key: - Free models: Sign up at opencode.ai → grab an OpenCode Zen key. Several models are truly free (Big Pickle, DeepSeek V4 Flash Free, MiMo-V2.5 Free, and others rotate). - Paid Zen models (optional): Add a payment method to your Zen account to unlock Claude Opus, GPT-5.5, Gemini, and other paid models at pay-as-you-go rates. Adding $20+ balance also improves rate limits on free models. - OpenCode Go (optional): Subscribe to OpenCode Go ($10/mo, $5 first month promo) for curated open models like DeepSeek V4 Pro, Kimi K3, GLM-5.3, Qwen3.8 Max, MiMo V2.5 Pro. 4. Open Copilot Chat (Cmd/Ctrl+Shift+I, or click the Copilot icon). 5. Click the model picker (current model name) → Add Models… 6. Select OpenCode Go or OpenCode Zen. 7. Press Enter to accept the default group name. 8. Paste your API key when prompted (stored by VS Code in your language-models configuration). 9. Pick the models you want enabled. 10. Select any OpenCode model from the picker and start chatting. 🚀

💡 Tips: - Go and Zen are separate provider groups: both can be active simultaneously. Switch anytime from the picker. - If a model shows in Language Models view but not the chat picker, hover its row and click the eye icon (👁) to enable it. - Set opencodego.freeOnly: false to reveal paid Zen models in the picker.

</details>

---

📊 Usage Tracking

- Go Usage Tracker: real-time burn-rate of OpenCode Go subscription: - Tracks 5-hour rolling ($12), weekly ($30), monthly ($60) tiers. - Server-synced meters: on startup and after each request the tracker pulls the official /zen/go/v1/usage endpoint with your Go key, so the Session/Weekly/Monthly percentages and "resets in" timers are account-wide and server-accurate (includes CLI, other devices, anything before the extension was installed). Falls back to local estimates when the endpoint is unreachable or the key/subscription is invalid. - Today / Yesterday and per-session spend stay device-local (the API does not return them). - Status bar: Go: 27%·62%·75%, with a ⚠ warning when any tier exceeds 80%. - Persisted in VS Code globalState; survives restarts. - Response usage bar: latest prompt/output/total/cache summary after each response. - Normalized usage DataPart: emits usage MIME so Copilot Chat's context widget shows accurate token counts.

Multiple Go accounts

You can use more than one OpenCode Go account in the same VS Code window. When a second Go API key is added via the Manage Language Models panel, the extension creates a new usage profile ("Profile 1", "Profile 2", etc.) on the first request made with each key.

- Auto-switch: the status bar and SVG hover card follow the model you last used. If you chat with a model from a different account, the panel switches automatically. - QuickPick: click the status bar to see which profile is active and switch to another profile. - Commands: OpenCode Go: Rename Active Profile and `OpenCode Go: Delete Profile` help you manage your profiles. The label you give a profile appears in the status bar and SVG title. - Storage isolation: each profile has its own globalState storage namespace (opencodego.usageLog.v1.<fingerprint>) so data never mixes. - SQLite: when 2+ profiles exist, the extension does not consult the shared opencode.db (which is per-machine and cannot be attributed to a specific API key). Accuracy falls back to in-memory entries for multi-account setups.

Upgrading from a single-account install? Your existing usage data is migrated into Profile 1 the first time a second profile is created. Your original stats are preserved; nothing is lost.

🎬 Demo

<p align="center"> <img src="docs/screenshots/model-picker.gif" alt="Copilot Chat model picker showing OpenCode models" width="480" /> </p>

Selecting an OpenCode model from the Copilot Chat model picker.

---

🔧 Settings

<details> <summary><b>📜 Full settings reference with descriptions</b></summary>

All settings live under the OpenCode namespace in VS Code Settings. Run Preferences: Open Settings (UI) and search opencode.

</details>

---

📊 GitHub Copilot vs This Extension

GitHub Copilot has four tiers now: Free, Pro ($10/mo), Pro+ ($39/mo), and Max ($100/mo). Here's how BYOK via OpenCode compares:

Model bridge, not completions replacement. OpenCode BYOK models work in VS Code Chat without a Copilot plan or GitHub sign-in. Standard inline suggestions, next-edit suggestions, semantic search, and embedding-backed features still require GitHub/Copilot support. OpenCode requests bypass Copilot billing entirely: you pay OpenCode directly (or nothing, on Zen free).

❓ FAQ

<details> <summary><b>Do I need Copilot Pro, Pro+, or Max?</b></summary>

No. On supported VS Code versions, BYOK chat works without a Copilot plan and without signing in to GitHub. Add OpenCode from Chat: Manage Language Models and select a model in Chat. Requests are billed only by OpenCode and do not consume Copilot requests.

Inline suggestions, next-edit suggestions, semantic search, and embedding-backed features still require GitHub/Copilot support. This extension provides chat, agent tool calls, and utility-model access; it does not replace standard Copilot autocomplete.

</details>

<details> <summary><b>Is it really free? What's the catch?</b></summary>

OpenCode Zen offers rotating free models, no balance required. Currently free: Big Pickle, DeepSeek V4 Flash Free, MiMo-V2.5 Free, Hy3 Free, Laguna S 2.1 Free, Ling-3.0-tiny Free, Nemotron 3 Ultra Free, Nemotron 3.5 Lightning Free (all limited-time, per the OpenCode docs). Without a balance, rate limits are low. Adding $20+ to your Zen balance significantly improves rate limits on free models. Paid Zen models (Claude Opus, GPT-5.6, Gemini, etc.) require adding a payment method; they're pay-as-you-go.

OpenCode Go is a subscription: $10/mo ($5 first month promo) with generous usage limits (5h/$12, weekly/$30, monthly/$60). It unlocks curated open models like DeepSeek V4 Pro, Kimi K3, GLM-5.3, Qwen3.8 Max, MiMo V2.5 Pro. When you hit the limit, you can continue using the free Zen models.

This extension is free and open source: you never pay us. You pay OpenCode directly (or nothing, on Zen free).

</details>

<details> <summary><b>Does Agent Mode / tool-calling work?</b></summary>

Yes, fully. The extension forwards VS Code tool schemas in the correct format for each endpoint (OpenAI tool_calls or Anthropic tool_use). Copilot Agent can read files, search, edit, and run terminal commands through any OpenCode model.

</details>

<details> <summary><b>Where is my API key stored?</b></summary>

In your VS Code language-models configuration: add it via Chat: Manage Language Models → Add Models… → OpenCode Go / OpenCode Zen. VS Code stores the key in its encrypted language-models storage, it never leaves your machine, and it is only sent to opencode.ai.

</details>

<details> <summary><b>Can I use Go and Zen at the same time?</b></summary>

Yes. They're separate provider groups. Add both via Language Models → Add Models…, enter each key separately, and switch between them from the chat model picker anytime.

</details>

<details> <summary><b>Can this import models from a local <code>opencode serve</code> instance?</b></summary>

Not safely through the VS Code language-model provider contract yet. opencode serve exposes OpenCode sessions and agents, whose tools execute inside OpenCode, rather than a Chat Completions, Responses, or Messages inference endpoint that can return VS Code tool calls. Registering only its model list would produce entries that lose Copilot's tool loop and permission UI.

For a local OpenAI/Anthropic-compatible inference server, use VS Code's Custom Endpoint provider directly. Native OpenCode-server support is tracked in #88.

</details>

<details> <summary><b>What should I include when the extension fails in an elevated/admin VS Code?</b></summary>

Run OpenCode Go: Diagnostics or OpenCode Zen: Diagnostics and attach the Runtime section plus the failing request summary. It identifies whether the elevated launch selected a different VS Code binary, extension installation, host, workspace trust state, or Windows integrity level. The extension requires VS Code 1.125 or newer.

</details>

<details> <summary><b>A model shows in Language Models but not the chat picker?</b></summary>

Hover its row in the Language Models view and click the eye icon (👁) to toggle visibility.

</details>

<details> <summary><b>Tool calls loop forever on Qwen?</b></summary>

Known issue with qwen3.6-plus-free on broad agent tasks (see issue #1). Workaround: set opencodego.thinking.qwen: "off" and use a narrower task scope, or switch to a paid Qwen model.

</details>

<details> <summary><b>How do I use OpenCode models in the Agents window (Copilot CLI)?</b></summary>

Agent models are enabled by default (opencodego.agentsWindow: true), and the extension auto-enables the two VS Code core settings they need (chat.agentHost.byokModels.enabled and extensions.supportAgentsWindow) on first activation, then offers a Reload Now button.

After reloading, open the Agents window, start a Copilot CLI session, and pick any OpenCode model from the picker. Only if you set opencodego.autoEnableAgentsWindow: false do you need to configure those settings manually. See the Agents Window section above for details.

</details>

<details> <summary><b>How do I report a bug or request a model?</b></summary>

Open an issue: pick the Bug Report or Feature Request template. Include the diagnostics report (OpenCode Go: Diagnostics or OpenCode Zen: Diagnostics).

</details>

---