LiteLLM Chat Provider for Visual Studio

Use 100+ LLMs in Visual Studio with a built-in chat window powered by LiteLLM. Multi-server support, streaming SSE, function calling, reasoning tokens, multimodal input, dashboard, status-bar context indicator.

Features

• Active editor context — attach unsaved document text and the current selection, with a preview before sending
• Solution tools for C# and Visual Basic — opt-in source-file discovery, literal search, Roslyn class/member lookup and semantic references across the open solution
• Reviewed edits across existing documents — inspect each proposal in a Before/After view; apply to an editor buffer only after confirmation, without automatically saving
• Inline change cards — while the model edits a file the chat shows which file it is working on, then a colored +N/−M diff of just the changed region, so a long edit never looks like a frozen window
• Markdown answers — headings, fenced code, lists and inline formatting rendered in the chat once a reply completes
• Image attachments — PNG/JPEG files, drag-and-drop and clipboard images with thumbnails; requires a vision-capable model
• Chat keyboard shortcuts — Enter or Ctrl+Enter to send, Shift+Enter for a new line
• 100+ LLMs through a unified OpenAI-compatible API (OpenAI, Anthropic, Google, AWS, Azure, Ollama, vLLM, sglang, and any OpenAI-compatible endpoint)
• Multi-server support — connect to multiple LiteLLM servers simultaneously and aggregate models across all of them
• Streaming chat completions via Server-Sent Events (SSE)
• Tool / function calling support, including a fallback parser for backends that stream raw tool-call syntax as plain text (Hermes/Qwen chat templates)
• Reasoning / thinking tokens (GLM / DeepSeek-R1 / o1-style reasoning_content)
• Automatic retry with smaller chunks when a response is truncated by the token limit
• Dashboard tool window for servers, models, and per-model enable/disable
• Status bar showing live server + model counts and a context-window indicator
• Per-model request parameter overrides (glob pattern → params, e.g. { "zai-org/GLM-5.2-FP8*": { "max_tokens": 50000 } })
• API key encryption via Windows DPAPI

Differences from the VS Code extension

This is the Visual Studio port of the litellm-vscode-chat VS Code extension.

Visual Studio (unlike VS Code) exposes no API for third-party extensions to register custom language model providers into the Copilot Chat panel. So instead of plugging into Copilot Chat, this extension ships its own WPF chat window that talks directly to LiteLLM — same streaming, tool-calling, reasoning, and retry behavior.

Requirements

• Visual Studio 2022 17.14+ or Visual Studio 2026, x64. The Marketplace build targets stable API 17.14 and Roslyn 4.14; VS 2026 supports these APIs through its extension compatibility model.
• .NET Framework 4.8
• C#/VB project support installed and a loaded solution for semantic solution tools
• A running LiteLLM proxy (self-hosted or cloud)

Quick Start

1. Open View → Other Windows → LiteLLM Chat.
2. Open View → Other Windows → LiteLLM Dashboard and add a server (label, base URL, API key).
3. Pick a model in the chat window's model dropdown and start chatting.
4. Use Attach active editor and Preview context to control the document context sent with a message.
5. To inspect related classes and callers, open your solution and enable Solution tools (C#/VB). Accept the source-access disclosure.
6. Select a proposed file change and click Review edit.... Approve each file individually; Ctrl+Z undoes the buffer edit and saving remains manual.

Version 0.3.8

Thinking blocks now show a live "Thinking… Ns" header with animated dots and an elapsed-seconds counter while the model reasons, switching to "Thought for Ns" when it finishes — a long reasoning phase visibly ticks instead of looking stuck.

Version 0.3.7

Thinking blocks no longer clip their text — the inner content is measured at full height so the bounded box actually scrolls. The chat message list now scrolls by pixels instead of whole messages, which makes the bottom of a long streaming answer (e.g. a comment below a code block) reachable with the sidebar scrollbar.

Version 0.3.6

The Review edit dialog now opens on a classic unified diff tab — line-number gutters, green added / red removed lines, ±3 lines of context and folded unchanged runs — with the full Before/After text still one click away. While the model generates a file edit, the chat shows a live progress card with the file name, a character counter and the code currently streaming in, so a long edit visibly progresses instead of looking like a frozen window.

Version 0.3.5

Streaming replies now render Markdown, thinking blocks stay collapsed and bounded, and UI updates are coalesced so a fast stream no longer stalls the window. Tool activity appears as a transient "Working: …" line, and each file edit produces a colored diff card with added/removed line totals.

Version 0.3.0

Adds a bounded model/tool/result loop so the model can inspect related source documents and receive errors when an edit proposal needs correction. Semantic references distinguish unrelated members with the same name. CRLF/LF differences are handled when matching proposed replacements, without guessing indentation or silently applying ambiguous edits.

This is a separate coding chat, not a complete autonomous Copilot replacement. It does not execute terminal commands, run builds/tests for the model, perform an automatic semantic rename, create/delete files, or provide inline ghost-text completions. Existing files can receive separate reviewed proposals; there is no automatic multi-file transaction.

Privacy

Prompts, attached editor text and images are sent to the server configured for the selected model. That server may route requests to upstream model providers; review its configuration and policies before sharing confidential code.

Solution tools are off by default and require confirmation. When enabled, source returned by tools is sent to the selected model server. Access is limited to indexed C#/VB source documents in the open solution; no arbitrary filesystem scanning or shell access is provided. Turning off context affects new attachments; clear the conversation to remove earlier context/images from future requests.

Images are normalized to PNG without source metadata, resized to a maximum 2048-pixel dimension, and limited to four images per message. Tool execution and returned context have explicit limits. Source edits require review and are not automatically saved. The extension does not add its own analytics telemetry.