31+ open coding models — GLM, Kimi, DeepSeek, Claude, Grok, Qwen — inside Copilot Chat's model picker, with vision, thinking mode, and a multi-agent swarm.
- Get a key — subscribe at opencode.ai/auth (Go or Zen) and copy your API key.
- Set it — run OpenCode: Set API Key; it lands in your OS keychain, never on disk.
- Chat — pick any of the 31+ models in Copilot Chat, or type
@swarmand let a team of agents do the work.
You already have Copilot's agent mode, tool calling, MCP, and skills. What you might not have is the models you want to run them on. This extension keeps the entire Copilot stack and swaps the brain:
- Don't replace Copilot — power it up. No sidebar, no new UI. Just a new model in the picker you already use.
- One API key, 31+ models. 14 OpenCode Go + 17 Zen models (5 free): GLM, Kimi, DeepSeek, Claude, Grok, Qwen, MiMo, MiniMax, Big Pickle, and more.
- Protocols handled for you. OpenAI
/chat/completionsfor GLM/Kimi/DeepSeek/Grok/MiMo, Anthropic/v1/messagesfor Claude/MiniMax/Qwen — routed automatically per model. - Everything Copilot gives you still works. Agent mode, tools, instructions, MCP, skills — untouched, now on your OpenCode subscription.
- Zero runtime dependencies. Pure VS Code API + Node.js built-ins. No Python, no Docker, no proxy process to babysit.
flowchart LR
subgraph VSCODE["VS Code · Copilot Chat"]
PICKER[Model Picker]
SWARM[Agent Swarm<br/>@swarm participant]
VISION[Vision Proxy<br/>describes images first]
end
PICKER --> ROUTER{Protocol Router}
SWARM --> ROUTER
VISION --> PICKER
ROUTER -->|OpenAI protocol| GO[OpenCode Go / Zen<br/>/chat/completions]
ROUTER -->|Anthropic protocol| ANTH[OpenCode Go / Zen<br/>/v1/messages]
ROUTER -->|GLM native| GLM[Zhipu · Z.ai · Coding Plan]
Every model is routed to the endpoint it speaks natively, images are described before they reach the model, and the Agent Swarm orchestrates parallel agents — all inside the model picker you already know.
Type @swarm and a team of autonomous agents takes over: research agents explore the codebase in parallel, review agents verify the plan before a single line is written, and an implementation agent — on the exact model you have selected — does the work and runs the tests.
flowchart TD
TASK[Your task] --> DECOMP[Decompose<br/>≤ 3 focus areas]
DECOMP --> R1[Research Agent · Area 1]
DECOMP --> R2[Research Agent · Area 2]
DECOMP --> R3[Research Agent · Area 3]
R1 --> F1[(Findings)]
R2 --> F1
R3 --> F1
F1 --> RV1[Reviewer A · verifies the plan]
F1 --> RV2[Reviewer B · verifies the plan]
RV1 --> IMPL[Implementation Agent<br/>chat-selected model]
RV2 --> IMPL
IMPL --> TESTS{runTests}
TESTS -->|fail · fix| IMPL
TESTS -->|pass| REPORT[Report · findings + review + diff + tests]
- Every agent is autonomous — its own tool loop and turn budget. Research and review are read-only; only the implementer edits.
- Resilient by design — a rate-limited or failed sub-agent degrades to a marked note; the swarm never sinks.
- Your model, your cost — research/review default to free models; implementation always runs on the model you picked.
Role models are configurable via
glm-copilot.agentRoles— see Settings.
Drop a screenshot into chat and the proxy describes it before the model sees it — GLM-5.2 focuses on code while GLM-4.6V-Flash handles the pixels. Falls back to any installed VS Code vision model automatically.
Full reasoning_content support on GLM-5.2/5.1 and Claude models. Pick none / high / max per model from Copilot Chat's native picker menu.
Every coding request carries a Ponytail-style system instruction that makes the model think like a lazy senior developer — efficient, not careless. Before writing any code it climbs a 7-rung ladder:
- Does this need to be built at all? (YAGNI)
- Does it already exist in the codebase? Reuse it.
- Does the standard library do this? Use it.
- Does a native platform feature cover it? Use it.
- Does an already-installed dependency solve it? Use it.
- Can this be one line? Make it one line.
- Only then: write the minimum code that works.
Three intensity modes — lite (reminder), full (complete ladder, default), ultra (strict, edge-case-first) — plus off. Injected only into real coding requests; utility calls stay lean and cheap. Switch live with OpenCode: Set Ponytail Mode.
An always-on refinement agent (inspired by the Claude Code Simplifier plugin) that reviews recently modified code and flattens nesting, kills redundancy, renames vague identifiers, and replaces chained ternaries — while never changing behaviour, API signatures, or safety checks. When active, Ponytail auto-downgrades to Lite for compatibility.
Because the extension plugs into Copilot's native provider API, you keep the full stack for free: agent mode, tool calling, instructions & skills, MCP, and prompt-caching stats.
- Per-turn list-price estimates in the status bar and logs — CNY on domestic GLM endpoints, USD on Z.ai.
- API key in VS Code
SecretStorage(OS keychain) — never insettings.json, never in Git history.
- VS Code 1.116 or later. This extension relies on non-public Copilot Chat APIs that may break on newer VS Code versions — report an issue if you hit one.
- GitHub Copilot subscription (Free / Pro / Enterprise — the free tier works)
- OpenCode account — Go subscription ($5 for your first month, then $10/month) or Zen pay-as-you-go. Subscribe at opencode.ai/auth and copy your API key. The GLM Coding Plan and Z.ai endpoints are also supported via the
endpointsetting.
Install from the registry used by your editor:
- Microsoft VS Code — install from VS Code Marketplace.
- Editors that use Open VSX — install from Open VSX.
- Subscribe to OpenCode and copy your API key from opencode.ai/auth
- Run OpenCode: Set API Key from the Command Palette (
Cmd+Shift+P/Ctrl+Shift+P) - Paste your OpenCode API key — it's stored in VS Code's secure SecretStorage (OS keychain)
- Open Copilot Chat, click the model picker, pick any OpenCode model (GLM-5.2, Kimi K2.7 Code, DeepSeek V4 Flash, Claude Sonnet 5, etc.)
- That's it — chat away!
All OpenCode Go & Zen models are available. The extension automatically routes each model to the correct endpoint protocol:
| Model | Best For | Source |
|---|---|---|
| GLM-5.2 | Flagship coding & reasoning, 1M context | Go |
| GLM-5.1 | High-quality coding & reasoning | Go |
| Grok 4.5 | Frontier reasoning (xAI) | Go |
| Grok Build 0.1 | Coding-tuned reasoning (xAI) | Zen |
| Kimi K3 | Frontier reasoning model | Go |
| Kimi K2.7 Code | Coding-tuned reasoning model | Go |
| Kimi K2.6 | General coding & reasoning | Go |
| DeepSeek V4 Pro | High-quality reasoning | Go |
| DeepSeek V4 Flash | Fast and economical coding | Go |
| MiMo V2.5 | Fast and economical coding | Go |
| MiMo V2.5 Pro | High-quality reasoning | Go |
| Model | Best For | Source |
|---|---|---|
| Claude Fable 5 | Frontier reasoning (Anthropic) | Zen |
| Claude Opus 4.8 | High-quality reasoning | Zen |
| Claude Opus 4.7 | High-quality reasoning | Zen |
| Claude Opus 4.6 | High-quality reasoning | Zen |
| Claude Opus 4.5 | High-quality reasoning | Zen |
| Claude Sonnet 5 | Balanced reasoning | Zen |
| Claude Sonnet 4.6 | Balanced reasoning | Zen |
| Claude Sonnet 4.5 | Balanced reasoning | Zen |
| Claude Haiku 4.5 | Fast economical model | Zen |
| MiniMax M3 | Coding agent work | Go |
| MiniMax M2.7 | Coding agent work | Go |
| MiniMax M2.5 | Coding agent work | Go |
| Qwen3.7 Max | Top-tier reasoning (256K context) | Go |
| Qwen3.7 Plus | Cost-effective reasoning (1M context) | Go |
| Qwen3.6 Plus | Cost-effective reasoning (256K context) | Go |
| Qwen3.5 Plus | Cost-effective reasoning (Anthropic) | Zen |
| Model | Notes |
|---|---|
| Big Pickle | Free stealth coding model (limited time) |
| DeepSeek V4 Flash Free | Free fast coding model (limited time) |
| MiMo V2.5 Free | Free fast coding model (limited time) |
| North Mini Code Free | Free coding model (limited time) |
| Nemotron 3 Ultra Free | Free NVIDIA trial model (limited time) |
Free models are available for a limited time and may collect data to improve the model. See Zen privacy docs for details.
All models support tool calling. GLM-5.2, GLM-5.1, and Claude (Fable 5, Opus, Sonnet) support thinking mode with reasoning effort control (none / high / max). Image attachments go through the Vision Proxy.
| Setting | Default | Description |
|---|---|---|
glm-copilot.endpoint |
opencode-go |
Single-value endpoint selector. opencode-go / opencode-go-anthropic serve OpenCode Go subscription models; opencode-zen / opencode-zen-anthropic serve OpenCode Zen pay-as-you-go models (including Claude and free models). Also supports Zhipu/Z.ai GLM endpoints: china-coding, china-standard, china-anthropic, international-coding, international-standard, international-anthropic |
glm-copilot.baseUrl |
empty | Optional API endpoint override. When non-empty, overrides the endpoint preset. |
glm-copilot.maxTokens |
0 |
Max output tokens (0 = no limit). Useful for cost control |
glm-copilot.modelIdOverrides |
prefilled OpenCode IDs | API model IDs to send for built-in or custom models. Change only for compatible endpoints with different model names |
glm-copilot.customModels |
[] |
Extra GLM-compatible models for the picker. Accepts string IDs or objects with id, optional name, token limits, toolCalling, and thinking. Custom IDs override built-ins. Images still go through the current Vision Proxy; custom models do not bypass it for native vision |
glm-copilot.debugMode |
minimal |
Diagnostic mode: minimal for token usage only, metadata for privacy-preserving logs, or verbose for full request dumps and pipeline snapshots under extension global storage. Full dumps may include sensitive prompt text, tool schemas, file snippets, and image descriptions. Use OpenCode: Open Request Dumps Folder to open the dump location |
glm-copilot.visionModel |
(auto) | VS Code vision model used as fallback when automatic vision is unavailable. Configure from OpenCode: Configure Vision Proxy; new saves use vendor/id, while legacy bare model IDs are still read |
glm-copilot.visionPrompt |
(built-in) | Prompt used to describe image attachments |
glm-copilot.ponytailMode |
full |
Ponytail coding-discipline system instruction level. off = no instruction; lite = brief reminder; full = complete 7-rung ladder with all rules; ultra = strict mode prioritizing edge-case correctness. Use OpenCode: Set Ponytail Mode to switch at runtime |
glm-copilot.codeSimplifier |
true |
Autonomous code refinement agent (on by default). Proactively reviews modified code and simplifies for clarity, consistency, and maintainability. When enabled, Ponytail auto-downgrades to Lite. Toggle with OpenCode: Toggle Code Simplifier |
glm-copilot.agentRoles |
{} |
Models per agent-swarm role: research (list, round-robin — defaults to free DeepSeek V4 Flash Free), implement (always the chat-selected model), and optional review (list, defaults to free Big Pickle). Each entry is { "vendor", "family", "id"? } |
glm-copilot.experimental.stabilizeToolList |
false |
Experimental. Tries to pre-activate VS Code/Copilot virtual tools so the API tools parameter is more complete and stable across turns. May improve context-cache hit rate when enabled tools change between turns. Can increase input tokens because more function definitions may be included; cache-hit input tokens are cheaper but still count toward usage. Usually leave it off with 64 or fewer enabled tools unless the tool list still changes across turns; do not enable it with more than 128 enabled tools |
Thinking Effort is configured from Copilot Chat's model picker for each thinking-capable GLM model.
Example settings.json for a custom API proxy:
{
"glm-copilot.baseUrl": "https://proxy.example.com/v1",
"glm-copilot.customModels": [
"my-model",
{
"id": "team-coder",
"name": "Team Coder",
"maxInputTokens": 200000,
"maxOutputTokens": 131072,
"toolCalling": true,
"thinking": true
}
],
"glm-copilot.modelIdOverrides": {
"glm-5.2": "your-glm-5.2-model-id"
}
}Recent VS Code versions gate custom providers from the background agent and the new agent window. If you can pick GLM in the editor chat but not in the agent window, add the extension to the allowlist in settings.json:
{
"extensions.supportUntrustedWorkspaces": true,
"extensions.supportAgentsWindow": {
"abbalochdev.opencode-for-copilot": true
}
}If the agent still refuses to start with No utility model is configured for 'copilot-utility-small' while the selected main model is BYOK, that is a known VS Code Copilot regression — see microsoft/vscode#324007. Switching the editor chat to an OpenCode model usually works while the upstream issue is open.
This extension targets the OpenCode (Go & Zen) endpoints and the official GLM endpoints (BigModel Coding Plan, Z.ai, and the documented BigModel/Z.ai standard API). VS Code/Copilot generates the tool schemas verbatim from its own tool definitions and forwards them as-is. Third-party relays or proxies (e.g. New API, OneAPI) often enforce stricter OpenAI-schema validation than the official endpoint and reject schemas that contain default: null, certain anyOf/oneOf shapes, or other minor deviations — the most common symptom is Invalid schema for function 'get_errors': null is not of type "array".
This is not something this extension sanitizes, by design:
- We forward exactly what VS Code/Copilot produces, so any compatibility fix that works on the official endpoint is preserved.
- Maintaining per-relay quirks would create an ever-growing patch surface that can mask real upstream bugs.
If you hit this on a relay, the supported options are:
- Switch
glm-copilot.baseUrlback to the OpenCode or official GLM endpoint (leave empty and useendpoint). - Open a request dump with OpenCode: Open Request Dumps Folder and inspect the offending tool schema, then report the strict-validation bug to your relay.
- The error is also written to the OpenCode output channel — you can copy the full server response from there.
| This extension | Local proxy (e.g. LiteLLM) | Standalone GLM extensions | |
|---|---|---|---|
| Works inside Copilot Chat | ✅ | ✅ | ❌ separate UI |
| Agent mode, tools, skills | ✅ | ✅ | |
| Vision support | ✅ proxied | ❌ | ❌ |
| No extra process to run | ✅ | ❌ | ✅ |
| One-click install | ✅ | ❌ | ✅ |
| API key in OS keychain | ✅ | ❌ |
This extension is a fork and rebrand of GLM for VS Code Copilot by ikaros, published under the MIT License. We thank the original author for the high-quality BYOK Copilot Chat provider implementation, vision proxy, and cost estimation infrastructure that make this OpenCode extension possible.
This project also references ideas and implementation patterns from Vizards/deepseek-v4-for-copilot, KiwiGaze/glm-for-copilot, and selfagency/z-models-vscode. Thanks to the original authors. Where applicable, redistribution and derivative work should preserve the original MIT License notices.