cv4vs Agents

Claude Code inside Visual Studio. A rich chat pane and an interactive terminal, both wired into the editor, solution, debugger and build system.

It is not a fork of the CLI. It drives the real claude.exe; the binary is never bundled, and version differences are handled by feature detection rather than by pinning a version.

Before you start

This extension needs the Claude Code CLI installed separately — it drives it, and cannot work without it:

winget install Anthropic.ClaudeCode
# or: npm install -g @anthropic-ai/claude-code

Other platforms and methods are in Anthropic's setup guide. If it is missing, the pane says so and links you there instead of failing silently.

Then open View → cv4vs Agents → Claude. The IDE tools are wired up automatically — nothing to configure.

Two panes, one session

Chat — streaming replies, thinking blocks, collapsible tool output, inline diffs, clickable file references (ClientEvents.cs:208 opens the file at that line), image attachments, and a composer with slash commands, an @ file picker and prompt history. Any reply can be read aloud.

Under the message box, a row for the turn itself: thinking, effort, model and permission mode, each a click away. Which model is answering, and what it may do without asking, are the two things that change a turn the most — they belong in sight, not three levels down a menu.

CLI — the real claude.exe in an embedded terminal, connected to the IDE over the same channel the official VS Code extension uses.

Both read the same session store, so a conversation started in one opens in the other — or in VS Code. Panes can run on different working directories at the same time.

What it can see and do

Around 50 MCP tools expose the IDE to the agent, in process, so it reads the live state rather than the files on disk:

Tools are language-agnostic where Visual Studio allows it: they go through Roslyn's per-document language services or through EnvDTE, not through a C#-only path.

Built for long sessions

Nothing is built, read or started until you look at it. The chat holds nothing in memory — the transcript is read from the session file on demand, newest page first, older pages as you scroll, and heavy blocks (images, sub-agent transcripts, full diffs) only when you open them. Services, the MCP server and the panes themselves start on first use, not on solution load.

A long session opens as quickly as an empty one.

Context and cost, visible

A live gauge in the composer shows how full the context window is, and a stats dialog aggregates token usage and cost from your local session files — per day, per project, per model.

There is also a full-window Statistics tab: a navigable tree (All → Profile → Folder → Project → Days/Sessions) drives summary tiles, a GitHub-style activity heatmap and per-day/per-model charts — all aggregated locally, no telemetry. A companion Usage tab shows each profile's live plan and rate-limit windows, and a Context usage tab breaks down how any historical session fills the model's context window (a memory-map, category table and the loaded files/agents/skills/tools).

Profiles

Each pane can run against a different configuration: working directory, model, permission mode and environment. That includes Anthropic-compatible endpoints, so one pane can use a different provider without touching your global settings.

Requirements

Documentation and support

Full documentation, including MCP tools, options, sub-agents and architecture, is on GitHub.

• Report a bug
• Request a feature
• Release notes

Problems in claude.exe itself belong to the CLI's own tracker — this extension drives the CLI, it does not ship it.

Credits and legal

Artwork by filocorsa.

GPL-3.0-only — Copyright Corsinvest Srl. Made in Italy 🇮🇹

Claude and Claude Code are trademarks of Anthropic, PBC. Visual Studio is a trademark of Microsoft Corporation. This is an independent extension by Corsinvest Srl, not affiliated with or endorsed by either company; the names are used only to describe what it works with.