OmniCopilot

中文 | English

A VS Code extension that lets you use models from multiple LLM platforms in GitHub Copilot Chat as language model providers.

Supported Providers

[!NOTE] The 23 providers added in 0.5.0 (SiliconFlow ×2, Tencent Hunyuan ×3, StepFun ×2, SCNet, Alibaba Coding Plan ×2, MiniMax International, Moonshot international, Kimi kimi.ai, MiMo SGP/AMS/pay-as-you-go, Umans, LongCat, SenseNova, KUAE, Zhipu/Z.AI/Ark pay-as-you-go) source their model line-ups, context/output limits and thinking domains from models.dev (community-maintained catalog, data as of 2026-10-01). 13 of them are live-tested (see Tested & Working above); the remaining 10 — Tencent Hunyuan Coding/Token Plan/TokenHub, StepFun Step Plan (both regions), SCNet, Umans, LongCat, KUAE, Volcengine Ark pay-as-you-go — are models.dev-sourced and follow the same OpenAI-compatible contract, but endpoints and key regimes follow each vendor's official docs and 4xx/parameter errors are worth an issue.

Tested & Working

The following providers have been tested and confirmed working:

• DeepSeek Open Platform (platform.deepseek.com)
• Kimi Code Plan (kimi.com/code)
• MiniMax Token Plan CN (platform.minimaxi.com)
• GLM Coding Plan CN (open.bigmodel.cn Coding API)
• Volcengine Coding Plan / Agent Plan (console.volcengine.com)
• Alibaba Token Plan (platform.qianwenai.com)
• MiMo Token Plan CN (platform.xiaomimimo.com)
• Moonshot AI (International) (platform.moonshot.ai)
• Kimi Code Plan (kimi.ai) (kimi.ai/code)
• SiliconFlow / SiliconFlow CN (siliconflow.com / siliconflow.cn)
• MiniMax International (minimax.io)
• Alibaba Coding Plan / CN (coding-intl.dashscope.aliyuncs.com / coding.dashscope.aliyuncs.com)
• MiMo Token Plan SGP / AMS, pay-as-you-go (token-plan-sgp/ams.xiaomimimo.com, api.xiaomimimo.com)
• SenseNova (token.sensenova.cn)
• Zhipu AI pay-as-you-go (open.bigmodel.cn)
• Z.AI pay-as-you-go (api.z.ai)

[!NOTE] GLM Coding Plan billing: per Zhipu's docs, the Coding endpoint (open.bigmodel.cn/api/coding/paas/v4) only counts toward the Coding Plan quota when called from officially supported tools (Claude Code, Kilo Code, OpenCode, TRAE, CodeBuddy, etc.). VS Code Copilot Chat is not on that list — success is not guaranteed, usage may be billed at pay-as-you-go API rates instead of your plan's credits, and Zhipu's usage notes treat non-listed-tool calls as a violation that may lead to throttling or account restrictions. Keep an eye on your billing and account status. The same caveat applies to the international GLM Coding Plan on Z.AI (api.z.ai/api/coding/paas/v4): it is strictly limited to officially supported tools, and team plan members must use the team plan key (not interchangeable with other Z.AI API keys).

[!WARNING] Alibaba Token Plan terms and endpoint: the extension points at the Token Plan endpoint (token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1) and expects the subscription key (sk-sp-…) from platform.qianwenai.com — Token Plan and pay-as-you-go credentials/endpoints are fully isolated and must not be mixed (pay-as-you-go keys (sk-ws-…) belong to dashscope.aliyuncs.com/compatible-mode/v1). The key is also restricted to interactive coding/agent tools (Claude Code, Cursor, Qwen Code, Qoder, OpenClaw, etc.) — the docs explicitly forbid generic API usage and state that violations may suspend the subscription or ban the API key. VS Code Copilot Chat is not on the official tool list, so use at your own discretion and watch your account status.

[!WARNING] MiMo Token Plan CN terms and endpoint: the extension points at the Token Plan CN cluster (token-plan-cn.xiaomimimo.com/v1; Singapore/Europe clusters exist at token-plan-sgp/ams.xiaomimimo.com/v1) and expects the plan's dedicated key (tp-… personal / ttp-… team) from platform.xiaomimimo.com — plan keys and pay-as-you-go keys (sk-… on api.xiaomimimo.com/v1) are fully isolated and must not be mixed. The plan's quota may only be used inside coding tools (OpenCode, OpenClaw, Claude Code, etc.); the docs classify non-coding API usage as abuse that may suspend the subscription or ban the key. VS Code Copilot Chat is not on the official tool list, so use at your own discretion and watch your account status.

[!WARNING] Terms risk on the newly added plan providers: the new Coding/Token Plan subscriptions (Alibaba Coding Plan on coding.dashscope.aliyuncs.com, Tencent Hunyuan plans, StepFun Step Plan, SCNet, KUAE, Umans, Kimi international on kimi.ai, …) generally mirror the Qwen/GLM/MiMo Token Plans — quota is restricted to officially listed coding tools, VS Code Copilot Chat is not on those lists, and calls from unlisted tools may be billed pay-as-you-go, throttled, or treated as a violation. Plan keys and pay-as-you-go keys are not interchangeable (the Zhipu/Z.AI/Ark pay-as-you-go vendors need each platform's pay-as-you-go key). Read each vendor's terms before use.

TODO

• [ ] Test Kimi Open Platform (moonshot-open)
• [ ] Test GLM Coding Plan, Z.AI international (glm-coding-plan)
• [ ] Test the untested models.dev-sourced providers: Tencent Hunyuan Coding/Token Plan/TokenHub, StepFun Step Plan (both regions), SCNet, Umans, LongCat, KUAE, Volcengine Ark pay-as-you-go
• [ ] Sink vendor-specific serialization knobs into VendorConfig flags (outputLimitField, acceptsReasoningContent, kimiHeaders, mimoReasoningContract, maxCompletionTokens, toolStream), replacing the vendor-ID OR chains currently spread across api.ts / provider.ts / extension.ts (review #24, deferred from 0.5.0)
• [ ] Turn the ad-hoc serialization checks into automated tests (node --test): MiMo tool-call history backfill (thinking On/None, including history where VS Code dropped the thinking part), THINKING_BUDGET mappings, per-model effort clamping, and the 186-model menu catalog assertion
• [ ] Verify thinking effort levels (DeepSeek None/High/Max; others None/Low/Medium/High or None/On) actually take effect across providers
• [ ] Decide whether 0.5.0 ships via the marketplace pre-release channel (11 providers still untested)
• [ ] To be continued…

Features

• Multiple Providers: 30+ platforms/endpoints, 186 models in total
• Per-Model Thinking Effort: Hover any thinking-capable model in the Copilot picker to pick the effort level for the next turn — no need to flip a global switch
  • DeepSeek V4 menu: None / Low / High / Max (matches the V4 API's reasoning_effort domain; thinking is on by default, None disables it explicitly)
  • Kimi K3 (Code Plan k3 / k3-256k, Open Platform & international kimi-k3, Volcengine-hosted kimi-k3): Low / High / Max — no None option, thinking is always on; effort maps to reasoning_effort
  • Kimi for Coding (Code Plan kimi-for-coding, currently K2.8 Preview): None / Low / High / Max — None sends an explicit thinking disable
  • GLM-5.3 family menu: Low / High / Max — thinking always on (covers GLM-5.3-Highspeed / FlashX and the GLM-5.3 hosting on SiliconFlow, SCNet, Umans, SenseNova; the Coding endpoint auto-routes legacy GLM IDs like glm-5.1 / glm-4.7 to the 5.3 line)
  • GLM-5.2 family menu: High / Max — always-on thinking with a native high|max domain (Z.AI Coding Plan, Zhipu/Z.AI pay-as-you-go, SiliconFlow hosting)
  • MiniMax M3.1 menu: Low / Medium / High / XHigh / Max — always-on thinking over the full native reasoning_effort domain
  • SiliconFlow-hosted Kimi/DeepSeek/Qwen/MiniMax/LongCat/Hy3/gpt-oss: None / Low / High / Max — mapped to enable_thinking + thinking_budget (128–32768)
  • SCNet / SenseNova-hosted DeepSeek V4 Pro: None / High / Max (native high|max; low/medium clamp to high)
  • 4-level menu (None / Low / Medium / High) for Qwen reasoning models, StepFun Step 5 Preview / Step 3.7 Flash, Ark/Volcengine doubao models, Ark-hosted GLM-5.2, and the Umans/SenseNova effort-capable entries
  • StepFun Step 3.5 Flash: None / Low / High — its native domain has no medium tier
  • 2-level menu (None / On) for models that only expose a thinking on/off knob (Kimi K2.6, MiniMax-M3, pre-5.3 GLM, GLM-5V-Turbo / 5-Turbo / 4.7, Tencent hy3/hy4 and hosted entries, Alibaba Coding Plan reasoning models, the whole MiMo v2.6 line) — MiniMax-M3's and MiMo's None genuinely disables thinking
  • Thinking-locked models expose no menu at all: the K2.7 Code high-speed variants (kimi-for-coding-highspeed, kimi-k2.7-code(-highspeed)), MiniMax M2.x, and some SiliconFlow CN snapshots — their "None" would silently reroute the model or keep thinking on anyway
• Thinking UI: Models with reasoning capabilities show collapsible thinking sections via LanguageModelThinkingPart
• Context Gauge: Streams stream_options: { include_usage: true } and reports the real token usage back to Copilot Chat, so the context-window indicator shows actual usage instead of 0; falls back to CJK-aware token estimation (Chinese ≈ 1 token/char) before the first real usage arrives
• Vision Support: Every model flagged image-capable in the catalog (deepseek-flash, the GLM-5.3-Flash/FlashX/4.6V/5V-Turbo line, Kimi K2.5+/K3, MiniMax-M3/M3.1, the Qwen vision models, all doubao models, the MiMo v2.6 line, step-5/3.7, the Umans line-up, SenseNova 6.8, …) can read images attached in Copilot Chat
• Tool Calling: Function calling support for compatible models

Usage

1. Install the extension
2. Open Copilot Chat → Manage Models → Add Model
3. Select a provider and enter your API key
4. Start chatting with the selected model

Models added but not showing in the picker?

These models live inside the Copilot Chat session, so whether they show up is gated by VS Code/Copilot itself, not this extension. Also check the extension side first — a provider with a missing or empty API key returns no models at all (re-enter the key via Manage Models → the provider's gear icon):

• Signed in to GitHub without a Copilot entitlement (no subscription, lapsed, or org plan without a seat): this is a known VS Code bug — BYOK/extension models are blocked while a GitHub session without a Copilot token exists (microsoft/vscode#324310, #327078). Workaround: sign out of GitHub entirely and reload the window — the models come back.
• Not signed in at all: BYOK models are supposed to work without a GitHub account or Copilot plan, but the Chat view stays on the sign-in/BYOK dialog until you pick the "use your own key" path.
• Copilot (Agent Host) sessions hide BYOK/extension models by default — enable the chat.agentHost.byokModels.enabled setting.
• Workspaces in Restricted Mode only show Auto in the picker — trust the workspace first.
• If a freshly added model doesn't appear, reload the window.
• Without sign-in, utility features (title generation, commit messages, …) need chat.utilityModel / chat.utilitySmallModel pointed at one of these models.

Configuration

Thinking effort is now selected per model, per turn via the Copilot model picker's hover menu — there is no global thinking-effort setting.

Development

Prerequisites

• Node.js (LTS recommended)
• VS Code 1.108.0+
• GitHub Copilot extension installed

Setup

git clone https://github.com/LyaQanYi/OmniCopilot.git
cd OmniCopilot
npm install

Build & Run

# Compile TypeScript
npm run compile

# Watch mode (auto-recompile on changes)
npm run watch

To debug the extension, press F5 in VS Code to launch an Extension Development Host with the extension loaded.

Package .vsix

npx @vscode/vsce package --no-dependencies

Project Structure

src/
├── extension.ts   # Extension entry point, activation & commands
├── provider.ts    # Language model provider implementation
├── api.ts         # API call logic (streaming, thinking, vision)
├── models.ts      # Preset model definitions per vendor
└── types.ts       # Shared TypeScript interfaces

Contributing

Contributions are welcome! Here's how you can help:

1. Fork the repository
2. Create a feature branch: git checkout -b feat/my-feature
3. Commit your changes: git commit -m "feat: add my feature"
4. Push to the branch: git push origin feat/my-feature
5. Open a Pull Request

Guidelines

• Follow existing code style (TypeScript strict mode)
• Test with at least one provider before submitting
• Keep commit messages clear and descriptive
• One feature/fix per PR when possible

Ways to Contribute

• Test a provider — Pick an untested platform from the TODO list, test it, and report results
• Add a new provider — Add model definitions in models.ts and register in extension.ts
• Fix bugs — Check Issues for reported problems
• Improve docs — Help with documentation or translations

Requirements

• VS Code 1.108.0+
• GitHub Copilot extension

Changelog

Full release history lives on the Releases page.

0.5.0 — 2026-10-03

• 20 new vendors, 186 models total — Moonshot AI International, Kimi Code Plan (kimi.ai), SiliconFlow (intl + CN), MiniMax International, Alibaba Coding Plan (intl + CN), Tencent Coding Plan / Token Plan / Tokenhub, StepFun Step Plan (intl + CN), SCNet Token Plan, MiMo Token Plan (SGP + AMS), MiMo Open Platform, Umans AI Coding Plan, LongCat, SenseNova, KUAE Cloud, Zhipu AI (pay-as-you-go), Z.AI (pay-as-you-go), Volcengine Ark
• Architecture: effortMenu-driven catalog — each model now declares its picker menu and effort domain directly; the per-turn effort clamping that was scattered across vendor branches is now handled in a single upstream pass (resolveRequestedEffort), so vendor branches only translate field names and shapes
• New picker menus — none-high (SenseNova), none-low-high (StepFun step-3.5-flash), high-max (GLM-5.2 family), low-medium-high-xhigh-max (MiniMax M3.1-Flash-Preview)
• Calibrated against models.dev — SenseNova glm-5.2 / deepseek-v4-flash domain collapsed to none|high (low/medium/high indistinguishable server-side); Volcengine plan/agent/ark DeepSeek models corrected to minimal|low|medium|high (no max); StepFun step-5-preview / step-3.7-flash annotated with four-level menus
• VSIX packaging — .claude/ worktree directories excluded from the bundle (53 KB → was 3.8 MB due to a missing .vscodeignore entry)

License

MIT