| Import setup from another harness | Yes, from Claude Code and Cursor. In a local TUI session, type /import. If both are found, choose a source, then tick what to bring over. Supported items are CLAUDE.md/.cursorrules to AGENTS.md, selected settings to config.toml, skills to .agents/skills, plugins and marketplaces, MCP servers, subagents to Codex agent TOML, hooks, slash commands converted to skills, project folders, and recent chats. Existing Codex files and entries are not overwritten. Claude Code project memories are imported only when you turn on the under-development external_agent_memory_import feature (codex features enable external_agent_memory_import). There is no Cursor memory import. From Cursor, only the legacy .cursorrules file becomes AGENTS.md; .cursor/rules/*.mdc are not imported. /import can't run while a task is running. The docs say it is also unavailable in remote or daemon sessions, but that limit was lifted in 0.157.0.
Review imported hooks, MCP auth/headers and permissions afterwards; unsupported fields are dropped. Source | /import [codex|gemini|cursor] (or claude import <source> from the shell) brings OpenAI Codex, Google Gemini CLI or Cursor setup into Claude Code: instruction files such as AGENTS.md are appended once to the matching CLAUDE.md, and MCP servers, commands, subagents, skills and the permission mode come across. --dry-run previews and --yes skips the picker. Separately, claude mcp add-from-claude-desktop copies MCP servers from the Claude Desktop app.
Needs v2.1.213 or later (v2.1.265 for Cursor) and is turned on by a fetched feature flag, so it is unavailable on Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, Claude Platform on AWS, apps gateways, or with telemetry or nonessential traffic disabled. Hooks, plugins and other settings are not imported. claude mcp add-from-claude-desktop works on macOS and WSL only.
Source | Partly There is no import command or migration wizard. Instead the CLI reads Claude Code and Codex files in place every time it starts: Claude Code hooks from .claude/settings.json, .claude/settings.local.json and ~/.claude/settings.json; Claude Code plugins; skills; subagents; commands; and CLAUDE.md. This is governed by the Include Third-Party Plugins, Skills, and Other Configs setting, which is on by default. Claude Code settings such as permissions, model or env are not read. Configure those in ~/.cursor/cli-config.json or .cursor/cli.json. The built-in /migrate-to-skills skill only converts Cursor's own rules and commands to skills. Source | Partly Only from OpenCode V1. V2 reads V1 opencode.json(c) and .opencode/ agents, commands, skills, and MCP servers from the same locations and normalizes them in memory, migrates global tui.json into ~/.config/opencode/cli.json on first start, and you can ask OpenCode to rewrite your config into the native V2 shape. V1 plugins must be ported, and nothing is imported from other harnesses apart from the skills it reads directly. Source | Partly There is no import command for another harness's setup. The MCP docs explain how to move servers by hand: copy mcpServers entries from Claude Desktop, Claude Code or Cursor into ~/.pi/agent/mcp.json or .pi/mcp.json as they are, convert Codex TOML or VS Code and OpenCode entries, then run pi mcp list to check them. Source |
|---|
| Other harnesses' files it reads directly | Partly Mostly not. Codex reads AGENTS.md and skills in .agents/skills, a shared folder other tools also use. It does not read .claude/skills, .mcp.json or Cursor rules. To use CLAUDE.md directly, add it to project_doc_fallback_filenames in config.toml. It is then used only in directories without an AGENTS.md. Source | Partly Claude Code reads AGENTS.md (and .claude/AGENTS.md) as project instructions when there is no CLAUDE.md, .claude/CLAUDE.md or CLAUDE.local.md; setting Project instructions to claude-md-and-agents-md in /config loads both. It does not read other tools' skill folders such as .agents/skills, or MCP configs such as claude_desktop_config.json. /init also reads Cursor rules (.cursor/rules/, .cursorrules) and .github/copilot-instructions.md when it drafts a CLAUDE.md. Reading AGENTS.md directly needs v2.1.277 or later; on older versions put @AGENTS.md in your CLAUDE.md. With CLAUDE_CODE_NEW_INIT=1, /init also reads .devin/rules/, .windsurf/rules/, .windsurfrules and .clinerules. Source | It reads AGENTS.md, CLAUDE.md and CLAUDE.local.md; skills in .claude/skills, .codex/skills, .agents/skills and the matching home-directory folders; and subagents in .claude/agents, .codex/agents, ~/.claude/agents and ~/.codex/agents. It also reads Claude Code hooks from .claude/settings*.json and ~/.claude/settings.json, and enabled Claude Code plugins. In CLI 2026.10.12 it also loads commands from .claude/commands and ~/.claude/commands, and installed Claude Code plugins listed in ~/.claude/plugins/installed_plugins.json, honoring enabledPlugins. Those plugins bring their skills, agents, commands, hooks and MCP servers. Not read: ~/.claude/CLAUDE.md, .claude/rules/, GEMINI.md, project .mcp.json, ~/.claude.json MCP servers, and Codex config.toml. Add MCP servers to .cursor/mcp.json or ~/.cursor/mcp.json. Claude Code Notification and PermissionRequest hooks are not supported. When names clash, .cursor/ skills and subagents win over .claude/ and .codex/. Source | Partly Only skills: .claude/skills and .agents/skills in the project (walking up), and ~/.claude/skills and ~/.agents/skills globally. V2 does not read CLAUDE.md (only AGENTS.md), or other harnesses' agents, commands, MCP configs, or hooks. If you relied on the V1 CLAUDE.md fallback, move that guidance into AGENTS.md. Source | Partly Pi reads AGENTS.md and, in any folder without one, CLAUDE.md from the agent directory, the working folder and its parents. It also loads skills from the shared Agent Skills folders ~/.agents/skills/ and .agents/skills/. It does not read ~/.claude/CLAUDE.md, .claude/skills, .claude/commands, other harnesses' MCP configs, hooks or subagents. You can point the skills setting or --skill <path> at another harness's skills folder yourself. Source |
|---|
| Import sessions or history | Partly Yes, through /import. Pick recent chats from Claude Code (~/.claude/projects/...jsonl) or Cursor agent transcripts, up to 50 chats from the last 30 days. They become Codex sessions you can resume. Each one gets the source's custom or AI title when it has one, and otherwise a title from its first message. Imported chats are text replays. Thinking is dropped, tool calls and results become short text notes, and images become placeholders. Re-running /import appends new turns to chats it has already imported. Source | Not supported No. /import brings over configuration only, and there is no command to load another harness's conversations; Claude Code resumes only its own sessions (including ones moved from the cloud with --teleport). Source | Not supported No. The CLI can't import Claude Code, Codex or other harnesses' sessions; --resume only opens Cursor's own chats under ~/.cursor/chats. In CLI 2026.10.12 a hidden, undocumented -p-only flag, --conversation-history-file, takes history in Cursor's internal format, not other tools' transcripts. Source | Partly V1 sessions are copied automatically: on first start the V2 server migrates the V1 session tables in the same opencode.db (and any opencode-next.db beta database) into V2 format and shows a progress overlay in the TUI. opencode session import <file|url> accepts only V2 export JSON, and nothing imports other harnesses' sessions. Migrated V1 sessions lose their per-session permission overrides. Source | Not supported No. /import <path.jsonl> copies a session file into Pi's session folder and resumes it, but it only accepts Pi's own JSONL format; a file whose first line is not a Pi session header loads as empty. Source |
|---|
| Export for use elsewhere | Partly Codex has no export to other harnesses. You can export a chat as Markdown with /export (clipboard or file) or /export session.md. The raw session JSONL files stay under ~/.codex/sessions. Setup lives in plain files that other tools can read or that you can copy by hand: AGENTS.md, .agents/skills and ~/.codex/config.toml. Source | Partly /export [filename] copies the current conversation to the clipboard or saves it as plain text, and claude -p --output-format json|stream-json or the Agent SDK's getSessionMessages() give structured data for scripts. There is no command that converts Claude Code setup or sessions into another harness's format.
Other tools can pull Claude Code setup in themselves; for example, Codex's /import reads Claude Code's settings, skills, CLAUDE.md, MCP servers, subagents and hooks. Source | Partly There is no export command for setup or sessions. Each chat is mirrored as JSONL at ~/.cursor/projects/<slug>/agent-transcripts/<chatId>/<chatId>.jsonl, which you can copy yourself. Setup kept in shared formats (AGENTS.md, .agents/skills) is read by other harnesses directly. Prefixing a message with & hands the chat to a Cloud Agent, but that stays inside Cursor. Share links are created from the Cursor editor (Teams and Enterprise plans), not the CLI. Source | Partly opencode session export [id] writes a session as JSON (--sanitize redacts it) and /export in the TUI saves a transcript. These use OpenCode's own format, sharing is not available in V2 yet, and there is no exporter for another harness.
Source | Partly Sessions only. /export writes the current session as HTML, or as Pi JSONL when the path ends in .jsonl (active branch only); pi --export <file> [output] makes HTML from the shell, and /share uploads it as a private GitHub gist or Radius artifact. There is no export of settings, skills or MCP config for another harness. Exports can contain prompts, tool output, file contents and credentials, so review them before sharing. Source |
|---|