Migrate from Warp to Claude Code.
2 documentation-derived translation patterns — what carries over and what to watch for. Cited to the Feature Parity Map; the audit tells you whether the move is worth it.
Both are conversational, in-terminal coding agents: you type plain English, the agent runs shell commands, reads and edits files, debugs, and self-corrects. A team using Warp Agent Mode for this can cut the Warp subscription and run Claude Code directly. Install Claude Code (`npm i -g @anthropic-ai/claude-code`), open the project, and run `claude` to get the same agentic REPL natively in any terminal (macOS/Linux/Windows/WSL) or terminal-of-your-IDE. The key reframe: Warp does NOT run Claude Code's brain — Warp only auto-detects a CLI agent launched inside it and adds a UI layer (rich input editor, vertical tabs with git metadata, agent notifications, the code-review panel). So if the team was already invoking `claude` inside Warp, dropping Warp keeps the exact same agent and only removes that surfacing layer; if they were using Warp's OWN native agent, switching to `claude` swaps the model harness but keeps the natural-language-in-terminal workflow. In Claude Code, approvals map to permission modes (`default` prompts per action; `acceptEdits` / `bypassPermissions` correspond to Warp's auto-accept toggle); for graphical diff review and inline plan editing, add the Claude Code VS Code or JetBrains extension and run `/ide` from the CLI.
- Warning: Loss of Warp's terminal UX: Blocks, the modern input editor (Ctrl-G), vertical tabs, and Warp Drive workflows/notebooks do NOT come along — Claude Code runs in whatever terminal emulator you already use. If the team valued the terminal itself (not just the agent), that polish is the real thing being given up.
- Warning: Slash commands differ: Warp Agent Mode commands and Warp's `/agent-*` skills are not the same as Claude Code's (`/clear`, `/compact`, `/context`, `/cost`, `/model`, `!bash`, Shift+Enter). Retrain muscle memory.
- Warning: Billing model flips: Warp gates agent usage by monthly AI credits (Free limited; Build/Max/Business have credit pools). Claude Code has no separate credit meter — it is included in Pro/Max/Team/Enterprise plans (shared with web Claude usage limits) or billed API pay-as-you-go via `ANTHROPIC_API_KEY`. Re-map budget from per-credit to per-plan/per-token before cutting.
- Warning: Model lineup narrows to Anthropic: Warp offers a multi-vendor dropdown (Anthropic/OpenAI/Google/open-weights) and Auto routing; Claude Code is Anthropic-only (Opus/Sonnet/Haiku via `/model`). Teams relying on non-Claude models inside Warp lose that choice.
- Warning: The native Warp Agent's persisted cloud conversation history and Oz Runs records are Warp-side; Claude Code session state lives under `~/.claude/` and resumes with `claude --resume`. Conversation history does not transfer between the two.
Both Warp and Claude Code are MCP hosts that connect Model Context Protocol servers (stdio/CLI and Streamable HTTP/SSE) to give the agent extra tools and data. Migrating MCP servers off Warp onto Claude Code is near-zero effort because Warp already reads Claude Code's own config files — Warp lists Claude Code as an MCP provider whose global config is `~/.claude.json` and project config is `.mcp.json`. So any server you configured for Claude Code-as-run-inside-Warp is already in Claude Code's native store and keeps working when you cut Warp. For servers you defined only in Warp's own location (`~/.warp/.mcp.json` or `.warp/.mcp.json`), port each entry with `claude mcp add` — e.g. `claude mcp add --transport http <name> <url> --header 'Authorization: Bearer ...'` for HTTP/SSE, or `claude mcp add <name> -- <command> <args>` for stdio. The `mcpServers` JSON shape (command/args/env or url/headers) is the same in both, so a Warp bulk-import JSON snippet maps field-for-field. Manage and confirm loaded servers in-session with `/mcp`.
- Warning: Auth re-handshake: OAuth-based servers store credentials locally per host, so a server authed via Warp's one-click OAuth must be re-authorized once in Claude Code (run the server, then `/mcp` to complete login). Env-var API keys and bearer headers copy directly.
- Warning: Warp's project-scoped servers require manual approval before starting; Claude Code likewise gates project `.mcp.json` servers — approve them once after the cut so they actually load.
- Warning: Warp's built-in `/agent-add-mcp` skill (agents editing file-based MCP defs in-conversation) has no identical Claude Code command; in Claude Code you add servers via `claude mcp add` / edit `.mcp.json` directly or via `/mcp`.
- Claude Code's built-in `ide` MCP server (127.0.0.1, per-activation token, `mcp__ide__getDiagnostics` / `mcp__ide__executeCode`) is specific to its VS Code/JetBrains extension and is unrelated to anything in Warp — do not try to recreate it.