Beta Stage — Back up your project before use.

DeepSeek v4 for Visual Studio

An AI coding assistant for Visual Studio 2022 and 2026

Overview

DeepSeek v4 for Visual Studio brings AI chat, code editing, solution-aware tools, terminal execution, and multimodal understanding directly into the IDE. It is designed for Visual Studio 2022 (17.14 or later) and Visual Studio 2026.

The extension combines a native-grade WebView2 chat experience with five cooperating agents, reusable Skills, MCP tool servers, Ghost Text completion, and persistent project memory.

Project Highlights

Highlight	Why it matters
Five-agent workflow	Ask, Explore, Plan, Edit, and Build collaborate through the Handoff protocol instead of forcing manual mode switching.
IDE-native interaction	Chat, inline editing, diff preview, build diagnostics, terminal commands, and Git operations stay inside Visual Studio.
Reliable code changes	Patch editing, four-tier matching, Healing repair, and per-hunk diff confirmation reduce accidental edits.
Long-context reasoning	A 900K token budget, Deep Reasoning, and automatic LLM summarization support large solutions.
Multimodal understanding	Images, screenshots, PDFs, webpage content, and OCR results can be reasoned about directly.
Extensible tools	Markdown-based Skills and MCP servers add custom workflows and external tools without changing the extension core.
Safety controls	Terminal approval modes and user/session/repository memory keep automation useful and controllable.

Core Features

Chat and reasoning: streaming responses, Deep Reasoning, Pro/Flash models, resumable streaming, and Prefix Cache.
Code editing: inline AI edit, exact/multi replacement, apply_patch, file creation, diff preview, and Ghost Text completion.
Project tools: solution exploration, file parsing, terminal commands, Git operations, and build verification.
Skills system: reusable workflows defined in SKILL.md, triggered by /skillname or semantic matching.
MCP integration: connect HTTP and stdio tool servers and classify their tools for agents.
Multimodal AI: image understanding, window screenshots, page-by-page PDF reading, web image reading, and OCR.
Search: Baidu Qianfan and DuckDuckGo with automatic fallback.
Memory: persistent user, session, and repository notes injected into new conversations.
Internationalization: automatic Chinese/English UI switching with custom translation overrides.

Requirements

Component	Requirement
Visual Studio	2022 (17.14+) or 2026
Operating system	Windows 10/11
Runtime	.NET Framework 4.7.2
Architectures	x64; ARM64 for the No-Local-OCR package

Installation

Download the latest .vsix from Releases, close Visual Studio, and install it.

Full package: x64 only, includes local OCR.
No-Local-OCR package: x64 and ARM64; local OCR is removed and other features remain available.

To build from source, install the .NET Framework 4.7.2 SDK and the Visual Studio extension development workload:

git clone https://github.com/zmy15/DeepSeek-v4-for-VisualStudio.git

Open .slnx in Visual Studio, build the solution, and press F5 to launch the experimental instance.

Quick Start

Get an API key from platform.deepseek.com/api_keys.
Open Tools > Options > DeepSeek Chat, enter the API key, and select a model.
Open View > Other Windows > DeepSeek Chat.
Ask a question, select code and press Ctrl+I for inline editing, or dispatch an agent command.

Recommended starting settings:

Setting	Value
Model	`deepseek-v4-pro`
Deep Reasoning	Enabled, effort `max`
Token budget	`900000`
Vision model	`deepseek-v4-flash-vision-exp` for image/PDF understanding

Shortcut	Action
`Ctrl+Shift+D`	Open or focus the chat window
`Ctrl+I`	Inline AI edit
`Enter`	Send chat message
`Ctrl+Enter`	Insert a newline

Agents and Skills

Agent	Responsibility
Ask	Answer questions, explain code, and summarize results.
Explore	Search the repository and analyze project structure.
Plan	Decompose tasks, design changes, and coordinate subagents.
Edit	Apply code and file changes with confirmation and review.
Build	Run builds, inspect diagnostics, and repair errors.

Use @ask, @plan, @edit, or @build to select an agent explicitly, or let Handoff route the task automatically. Skills are loaded from project-level directories, user-level directories, and built-in skills.

Roadmap / TODO

Item	Description	Priority
RAG code retrieval	Local vector storage, file indexing, BM25/vector hybrid search, and solution symbol indexing.	High
Code knowledge graph	AST-based relationships and semantic navigation across classes and methods.	Medium
Test generation skill	Generate xUnit tests from the built-in TDD workflow.	High
GitHub integration	PR descriptions, review assistance, and issue assignment.	Medium
More built-in skills	Debug analyzer, SQL optimizer, API design, and other workflows.	Medium
Local model support	Ollama and LM Studio offline inference.	Low
Session export	Export conversations as Markdown, PDF, or HTML.	Low
More UI languages	Japanese, Korean, and additional locales.	Low

Support

Usage questions and discussions: Discussions
Bug reports and feature requests: Issues

DeepSeek-v4-for-VisualStudio