Claude Code Studio

A native Claude Code chat client for Visual Studio 2022 — a WebView2 chat panel driven by a persistent agent process that streams Claude's responses over stream-json, with inline permission prompts, native diffs, session history, and a usage dashboard.

Unlike terminal-embedding extensions, Claude Code Studio does not host a console window. It runs the claude CLI as a long-lived subprocess and speaks its streaming JSON protocol directly, so the whole conversation — text, tool calls, diffs, permission requests — is rendered as first-class UI instead of scraped terminal text.

• Visual Studio 2022 17.14 or newer (amd64)
• Windows
• Claude Code CLI installed and on PATH (or configured via the CLI-path setting)
• A Claude subscription (Pro or higher) or API access, signed in through the CLI

1. Build the solution (ClaudeStudio.slnx) in Visual Studio, or install the packaged VSIX by double-clicking it.
2. Restart Visual Studio.
3. Open the chat panel via Extensions → Claude Code Studio → Chat.
4. If you aren't signed in yet, use the in-panel sign-in to run claude login.

1. Open a solution or folder so the extension has a working directory.
2. Pick a model (and effort level) from the model button.
3. Type a prompt and send it — replies stream into the panel.
4. Attach context with @ (files), Ctrl+V (images), or the editor selection button.
5. When Claude wants to edit or run something, approve it inline (or switch permission mode to suit how hands-on you want to be).
6. Review edits via the diff chips; browse or resume past work from the session history.

• Streaming chat panel (WebView2) — Claude's replies stream in as rendered Markdown, with syntax-highlighted code fences and per-message model info.
• Live turn indicators — a presence line shows ✳ thinking…, 🗜️ compacting context…, and live TodoWrite / context-usage updates while a turn runs.
• Auto-compact feedback — when the context window is auto-compacted mid-session, a live indicator and an inline "compacted N → M tokens" divider explain what happened instead of a silent pause.
• Subagent trace — text and tool calls from spawned subagents (Task/Explore) are rendered nested under the parent tool chip, at an off/simple/full level you choose.
• Font zoom — Ctrl+Scroll zooms the whole chat; the level persists across sessions.

• Inline permission prompts — tool calls that need approval surface a modal with Allow, Allow for session, and Deny, showing the working directory the call runs in. Backed by a PreToolUse hook over a named pipe, not a terminal prompt.
• Permission modes — switch between ask, plan (Exit Plan Mode becomes an explicit "Approve plan" modal), bypass permissions (yolo), and don't ask (read-only: reads and searches run freely, all edits are blocked by the CLI).
• Permission rules — allow / deny / ask rules honored per session.

• Native diff viewer — Edit/Write chips open the real Visual Studio diff (IVsDifferenceService), inline or in a modal.
• Rewind (⟲) — roll a session back to an earlier point (dry-run preview, then revert).
• Branch (⎇) — fork the conversation from a chosen message into a new line.

• Per-workspace history — list past sessions for the current folder and resume any of them in one click.
• Rename & view — give a session a custom title inline, or open its full transcript as Markdown before resuming.
• Fork (↳) — resume a session as a native fork (--fork-session): a brand-new session seeded with the original's history, leaving the original untouched.
• AI-generated titles — sessions are named automatically, with custom renames always winning.

• @ file picker — type @ to search the workspace's files and folders with keyboard-driven drill-down, and insert a path.
• Slash commands & skills — / autocompletes built-in and custom commands plus installed skills; picking a skill drops a /name args chip and a "Use skill" card.
• Editor selection → prompt — send the current editor selection as a formatted, path-and-line-annotated snippet.
• Image & file paste — paste images (Ctrl+V) or attach files; binaries are shared with the CLI via a temp --add-dir.
• Prompt history — browse previous prompts with Ctrl+Up / Ctrl+Down, with an N/total indicator.

• Usage dashboard — token usage and computed cost per session, with auto-refresh, model breakdown, and a filter to a single session. Cache-write pricing is split by TTL (5-minute vs 1-hour) so long sessions are costed correctly.
• Cost limits — a soft spend warning plus an optional hard per-session budget (--max-budget-usd) that stops the turn when reached.

A single palette gathers on-demand tools that render as inline cards:

• 🩺 Doctor — run claude doctor and show the report.
• MCP status & reconnect — inspect MCP servers and reconnect them.
• Context usage — see how full the context window is.
• Build errors → agent — send the current build's errors (with warnings for context) to Claude on demand, or opt in to auto-send on every failed build; duplicate error sets are de-duped to avoid loops.
• Plus live TodoWrite, side-question cards, and system info.

• Working-directory cascade — resolves the cwd from the open Solution → Open Folder → active document → user profile, shown in a click-to-copy path bar.
• Trust management — workspace and MCP-server trust gates with a management UI, backed by a trusted-workspaces store and a file-system watcher.
• Integrated sign-in — claude login runs from inside the extension with a watcher and an overlay, and the titlebar shows your account name and plan.

• Settings panel — a searchable settings UI (search matches label, tooltip, and section) that passes through the full Claude Code settings surface, with info tooltips on individual controls.
• Model picker — Sonnet 5 / Sonnet 4.6, Opus 5, Opus 4.8, Opus Plan, Fable 5, Haiku 4.5, plus an effort level (low → max; "max" is session-only).
• Fallback model — an optional --fallback-model used when the primary model is overloaded, with a divider marking when it engages.
• Configurable CLI path — point at a specific claude executable, or let detection pick it (it prefers ~/.local/bin/claude.exe — the native installer's target — over a possibly-stale copy on PATH, and warns when the two diverge).
• Status line — an optional configurable status command rendered under the path bar.
• Theme aware — follows the Visual Studio light/dark theme, with an accent/override option.

Visual Studio (VSIX, .NET Framework)
        │  WebView2  ⇄  HTML/CSS/JS chat UI
        │
        ▼
ClaudeStudioAgent (subprocess, .NET 10)
        │  stdin/stdout: stream-json
        ▼
   claude  (Claude Code CLI)

The extension hosts the chat UI in WebView2 and launches ClaudeStudioAgent as a persistent subprocess. The agent spawns and talks to the claude CLI in streaming JSON mode, translating its event stream into chat chunks and relaying permission requests back through a PreToolUse hook. Pure, testable logic (event parsing, pricing, session-ordinal mapping, CLI discovery) lives in ClaudeStudioShared and is covered by the ClaudeStudioTests xUnit suite.

• Solution: ClaudeStudio.slnx (extension, agent, shared library, tests).
• Run the tests: dotnet test ClaudeStudioTests/ClaudeStudioTests.csproj (xUnit; covers the shared parsing/pricing/ordinal logic).

Released under the MIT License — free to use for personal, educational, and commercial purposes, including modification and redistribution.

A note on data: your prompts, attachments, and the code Claude reads are sent to the Claude Code CLI, which forwards them to Anthropic under Anthropic's data-usage policy. This extension stores only local state (settings, session titles, usage/perf logs) under %LocalAppData%\ClaudeStudio\ and Claude's own ~/.claude data; it sends nothing to any third party of its own.

"Claude" is a trademark of Anthropic. This is an independent, unofficial extension and is not affiliated with or endorsed by Anthropic.

Claude Code Studio was inspired by ClaudeCodeExtension by Daniel Carvalho (dliedke) — several UX ideas here (session rename/view, the @ file picker, the build-errors-to-agent flow) were adopted from that project, which is likewise MIT-licensed. It takes a different architectural path — an embedded terminal supporting many agents — and is well worth checking out. Thanks, Daniel.