Introfinity is a technology company delivering leading AI solutions and cybersecurity for modern enterprises — intelligence that acts, defenses that endure.

What does Introfinity offer?

Introfinity provides AI agents and intelligent automation, applied ML systems, decision support, threat detection, secure architecture, and continuous defense monitoring for enterprise organizations.

What is on the Introfinity Work page?

Work collects Introfinity products, client engagements, platforms, and engineering systems. Projects are labeled by format (case study, product story, or engineering deep dive) and maturity (live, open source, prototype, or in development). Flagship stories currently include Agent Platform. Client Work currently includes Vectenza, Fit & Flex, and CubeCom Pro.

How can I work with Introfinity?

Contact hello@introfinity.com or use the contact form at /contact to discuss AI solutions, cybersecurity posture, or systems that need to think and defend under real constraints.

How do I use custom LLMs in GitHub Copilot?

Copilot Custom LLM Provider is a VS Code extension that connects OpenAI-compatible endpoints to GitHub Copilot Chat, including DashScope, MiniMax, OpenRouter, Ollama, LiteLLM, and vLLM, with streaming, tool calling, and multi-provider model discovery.

What is Copilot Custom LLM Provider?

Copilot Custom LLM Provider is a VS Code extension that connects OpenAI-compatible endpoints to GitHub Copilot Chat, including DashScope, MiniMax, OpenRouter, Ollama, LiteLLM, and vLLM, with streaming, tool calling, and multi-provider model discovery.

Can Copilot Chat use Qwen or local models?

Copilot Custom LLM Provider is a VS Code extension that connects OpenAI-compatible endpoints to GitHub Copilot Chat, including DashScope, MiniMax, OpenRouter, Ollama, LiteLLM, and vLLM, with streaming, tool calling, and multi-provider model discovery.

Copilot Custom LLM Provider: Bring any OpenAI-compatible model into VS Code

Copilot Chat is useful. The model picker is a bottleneck.

Teams increasingly want agentic coding in VS Code without being locked to a single hosted model family. Alibaba Cloud Coding Plan, MiniMax, OpenRouter, local Ollama, LiteLLM proxies, and vLLM deployments all speak a familiar dialect: OpenAI-compatible chat completions. GitHub Copilot Chat did not natively treat those endpoints as first-class models in the picker.

Copilot Custom LLM Provider was built to close that gap. The primary use case was bringing Alibaba Cloud Coding Plan — including Qwen Coder models in agent mode — into Copilot Chat. The architecture stayed general: any OpenAI-compatible /v1/chat/completions endpoint should work.

What had to work on day one

A thin proxy is not enough for Copilot Chat. The extension needed:

models appearing directly in the Copilot Chat model picker
full streaming over Server-Sent Events
tool calling so agent mode, /fix, /edit, and @workspace keep working
multi-provider support with separate URLs and API keys
dynamic model discovery from /v1/models (or LiteLLM-compatible /model/info)
stable provider IDs so renaming a display label does not break the model list
secret storage for keys instead of writing credentials into settings JSON
hot-reload when settings change

Those requirements define the product. If streaming or tools break, the extension is a novelty. If discovery is manual, it becomes settings debt.

Multi-provider by default

Developers rarely stay on one backend. The extension lets Alibaba DashScope and OpenRouter sit side by side, or a local Ollama instance next to a remote LiteLLM proxy. Providers are added through a command-palette wizard: name, base URL, API key. Models from every healthy provider merge into one picker list under Custom LLM.

An optional @qwen chat participant routes a single turn through the custom model without permanently hijacking the picker selection. That keeps native Copilot models usable in the same session.

Reliability edges that matter in an IDE

IDE traffic is bursty and failure-prone. The extension retries with exponential backoff for network failures and rate limits, caps output tokens, and supports per-model parameter overrides such as temperature. Image input is supported for multimodal models. The runtime dependency footprint is intentionally empty — the extension should not drag a package tree into every VS Code workspace.

What this case study claims

Custom model support for Copilot is an integration problem, not a prompt problem. The durable work is provider identity, discovery, streaming, tools, secrets, and failure handling — the boring machinery that makes an alternate model feel native.

Extension: github.com/syedsadiq27/copilot-custom-llm-provider