Claude Desktop Wiki

Anthropic's official desktop app for Claude. macOS (Intel + Apple silicon), Windows. A wrapper around claude.ai with local niceties — system tray, keyboard shortcuts, drag-and-drop files — plus three differentiators: MCP server support, Claude Cowork (collaborative sessions), and Claude Code access (launch + drive the coding agent from the same app).

Desktop is the umbrella surface. From a single signed-in app you get: (1) the chat + artifacts + projects experience from claude.ai; (2) Claude Cowork collaborative sessions with all their features; (3) Claude Code with its full feature set — slash commands, hooks, skills, subagents, MCP, permissions, memory. Everything inherits one auth + one plan.

Same Claude models as the web app — Opus (deep reasoning), Sonnet (fast + balanced), Haiku (cheap + quick). Pick per conversation. Not the 1M-context Opus that Claude Code ships today; Desktop uses the standard 200K variants.

One chat thread. Independent context window. Can attach files + images. Lives under your Claude.ai account; synced across web + desktop. Threads under a Project inherit that project's knowledge + instructions.

Sign in with your Claude.ai account. Free / Pro / Team / Enterprise plans all supported. Plan tier determines message quota + model access + Projects availability.

Free (limited Sonnet/Haiku, no Projects). Pro (full model access + Projects + higher quota). Team / Enterprise (admin + SSO + audit log + shared Projects). Claude Max (specific model quotas for heavy users).

Collaborative Claude workspace exposed through Desktop. Lets multiple humans + Claude share a live session — same artifacts, same project knowledge, same running tool state. Desktop is the primary client; Cowork sessions open in-app rather than in a browser tab.

A single shared workspace. Multi-participant, multi-turn. Artifacts, files, and tool results are shared state — every participant sees the same view. Session outlives any one user; rejoin from Desktop with a link.

Invite by email or link. Role-gated (owner / editor / viewer). Team + Enterprise plans get workspace-wide access controls + audit trail. Free + Pro can share 1:1 or with a small group.

Project = your scoped knowledge + instructions, private threads inside. Cowork = shared live session, multiple humans + Claude in one room. They compose: a Cowork session can run inside a Project and inherit its knowledge.

Desktop surfaces the full Claude Code coding agent — not a stripped-down version. All features carry over: slash commands, hooks, skills, subagents, MCP, permissions, memory, the Agent SDK. Launch from Desktop; sessions share your Claude plan quota.

Everything documented in the Claude Code wiki: /slash commands (built-in + custom), PreToolUse/PostToolUse/UserPromptSubmit hooks, skills at .claude/skills/*, subagents (general-purpose, Explore, Plan, custom), MCP servers (stdio/HTTP/OAuth), permission modes, session + project memory, CLAUDE.md loading, checkpointing.

Same agent, different driver. The terminal CLI (claude) is still canonical for repo-local workflows. Desktop integration is for operators who want Claude Code's capabilities from a GUI, or to hand a coding task to the agent while staying in the chat surface. Both see the same ~/.claude.

Signing into Desktop signs you into Claude Code inside it — no separate keys, no separate subscription. Rate limits, model access, and Max quotas apply across chat, Cowork, and Code sessions.

Interactive content rendered in a side panel — code, HTML pages, SVG, React components, markdown docs, diagrams. Claude creates artifacts when the content is substantial + self-contained enough to warrant its own pane.

Code (syntax-highlighted, copyable) · HTML (preview renders live) · React (single-file component, renders with a small runtime) · SVG (renders inline) · Mermaid (renders diagrams) · Markdown (formatted preview).

Ask Claude to change a specific part of an artifact and it updates the existing artifact in place (versioned). You can browse earlier versions via the dropdown. No manual copy-paste between turns.

Publish an artifact to a public URL (Pro+). HTML artifacts become a hosted micro-site. Useful for prototypes, one-page demos, shareable calculators. Published artifacts stay editable by you and viewable by anyone with the link.

Save any artifact to disk as its native file type — .html, .jsx, .svg, .md, .py. No auth wrapper, no chrome; just the content.

Named container for conversations that share context. Attach up to 10 files / docs as knowledge; set custom instructions; conversations started under the project inherit both automatically. Pro+ feature.

Per-project file attachments. PDFs, Word docs, text, images, code files. Claude has these loaded as context in every conversation under the project. Max ~200K tokens of knowledge; prefer summaries of large sources over raw dumps.

Free-form system prompt scoped to the project. "Speak like a senior engineer. Always provide code with explanatory comments. Default to TypeScript." Applies to every conversation under the project; overrides your global user preferences.

Team / Enterprise plans can share a project across users — shared knowledge + shared instructions + separate individual conversations. Good for team-wide primers (codebase overview, brand voice, product docs).

Claude Desktop was the launch vehicle for MCP. The app's differentiator over the web Claude.ai experience is the ability to plug in local MCP servers — filesystem, git, Slack, your own custom integrations. Extends Claude's tool surface without needing API-level tool-use code.

User-global config at ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) / %APPDATA%/Claude/ (Windows). Contains the mcpServers map — one entry per server with transport, command, args, env.

Map keyed by server name. Each entry: command, args, optional env. Claude Desktop spawns each server as a subprocess on app start; tools become available in every conversation.

filesystem (local file access scoped to an allowlist) · git · GitHub · Slack · Gmail · Google Drive · Puppeteer (headless browser) · SQLite · Postgres · Stripe · Cloudflare. Hundreds of community-built servers. Implementations in TS + Python.

When Claude wants to call an MCP tool, Desktop shows an inline confirmation with the tool name + arguments. Approve once or "allow for this conversation." Per-server kill switches in settings.

Setting that unlocks MCP config UI + richer error messages + the ability to reload server config without restarting. Toggle in Settings → Advanced. Not strictly required — you can edit the JSON directly — but nicer during MCP development.

Live search for current info. Pro+ feature. On by default; toggleable per conversation. Claude cites sources. Good for "what's new in X?" or "what does this repo do?" when paired with a GitHub link.

JavaScript sandbox that runs in-browser. Claude writes + executes JS to analyse uploaded data, manipulate JSON, draw charts, run calculations. Output visible inline; state persists within the conversation.

Drop an image; Claude reads it. Works for screenshots, diagrams, charts, handwriting, photos. Not generative — doesn't create images. (For image gen, pair with an MCP server like Replicate or FAL.)

Attach PDFs, Word, plain text, spreadsheets, code files. Claude reads all of them natively. Large PDFs (100+ pages) work but slow; consider splitting for responsiveness.

Drop any supported file onto the chat window. Text, PDFs, images, code files (any extension Claude recognises), spreadsheets, JSON, Markdown. Multiple files at once. Max file size ~30MB each.

Text (any utf-8) · PDF · DOCX · XLSX (some caveats) · images (PNG/JPG/GIF/WEBP) · code files (any language — Claude infers syntax from extension) · Markdown · JSON · CSV · YAML. Unzips archives automatically in some cases.

Files attached to a single conversation persist only for that chat. Files attached to a Project's knowledge base apply to every conversation under that Project. Use conversation uploads for one-off docs, Project knowledge for persistent context.

Two ways Claude Desktop touches local files. Uploads: drag in a file, Claude reads the bytes into context. Filesystem MCP: configure a server with an allowlist, Claude reads + writes files in place. Uploads for review, MCP for ongoing edits.

Launch-at-startup · open-in-tray · notification preferences · default model · theme. Light / dark / auto. Keyboard shortcut to summon the app window (macOS).

Name Claude should call you · what you do · preferred response format (concise vs thorough) · any global preferences. Applied to every conversation unless overridden by a Project's custom instructions.

Developer Mode toggle. MCP servers UI. Raw config file path. Debug log access. Opt-out of training (some plans).

Chat retention · model training consent · export all data · delete account. Regional data residency options (Enterprise).

Same Claude models, same Projects, same artifacts, same chat history (synced via account). Desktop adds: system tray · native keyboard shortcuts · drag-and-drop · MCP server support. The MCP piece is the single biggest differentiator — the web app can't run local MCP servers.

Different surfaces for different work. Claude Code = terminal + hooks + skills + agent loop for coding. Claude Desktop = chat UI + artifacts + projects + MCP for general work. Share the underlying models; don't share config files. MCP servers can target either — same protocol.

Claude Desktop has no equivalent to Claude Code's PreToolUse / PostToolUse / Stop hooks. Tool calls are gated via inline confirmation UI instead. For programmatic interception of model actions, use Claude Code or the Agent SDK.

Projects are the closest analogue — per-project custom instructions + knowledge. But there's no directory of versioned procedures the way Claude Code's .claude/skills/ works. Reusable procedures on Desktop live in Project custom instructions or in an MCP server.

200K tokens (shared with Sonnet/Opus standard tier). Projects' knowledge counts against this. If you attach a lot, summarise; don't dump raw.

Human-readable operator guide. Tabs: Overview · Install · Artifacts · Projects · MCP · Tools · Files · Settings · Compare.

Claude's terminal-native sibling. Different surface, same Claude models. Use Claude Code for dev loops, Claude Desktop for everything else.

OpenAI's coding agent family. ChatGPT Codex is the closest analogue to Claude Desktop for agent work — sandboxed browser tasks. Codex CLI is the analogue to Claude Code.

The open protocol behind MCP. Small spec; implementation SDKs in TypeScript + Python. Write your own server in an afternoon.

Third-party desktop client for the OpenClaw Gateway. Similar shape — local desktop app + MCP surface. Useful for comparing non-Anthropic takes on the pattern.