Commands

agentmove export <client> [-o dir] [--json]

Reads the client’s config, MCP servers, instructions, persona, memory, and skills into a portable bundle directory (default ./agentmove-bundle). Likely secrets are redacted to ${VAR} placeholders unless --include-secrets.

agentmove import <client> [-i dir] [--apply] [--replace-mcp] [--only layers] [--json]

Plans the file writes needed to apply a bundle to a client. Dry-run by default: prints each file that would be written. With --apply, existing files are first backed up to ~/.agentmove/backups/<timestamp>/.

Imported MCP servers are merged into the target’s existing server list (like the official mcp add commands) — servers the target already has are never removed. Same-name conflicts are won by the imported entry, with a warning. Pass --replace-mcp to replace the list entirely instead.

agentmove convert <from> <to> [--apply] [--replace-mcp] [--only layers] [--json]

export + import in one step, without leaving a bundle on disk. Uses the same MCP merge semantics as import.

Partial migration: --only

export, import, and convert accept --only <layers> — a comma-separated subset of mcp, skills, memory, instructions, persona — to migrate only those layers. For example, to copy just the MCP servers from Claude Code to Codex without touching instructions or skills:

agentmove convert claude-code codex --only mcp --apply

Unknown layer names fail with exit code 2.

Project-scoped migration: --project

export, import, and convert accept --project <dir> to operate on the client’s project-scoped files inside a repository instead of the user-scoped config under $HOME:

Client	Project files
claude-code	.mcp.json, CLAUDE.md, .claude/skills/
codex	AGENTS.md, .agents/skills/ (no project-scoped MCP config — warned)
gemini	.gemini/settings.json, GEMINI.md
cursor	.cursor/mcp.json, .cursor/rules/*.mdc
windsurf	.windsurf/rules/*.md (no project-scoped MCP config — warned)

# move a repo's Claude Code setup to Cursor, in place
agentmove convert claude-code cursor --project . --apply

Backups go to <dir>/.agentmove/backups/<timestamp>/. MCP merge semantics and secret redaction work the same as user-scoped migration. OpenClaw and Hermes have no project-scoped files — --project with them is a usage error (exit 2).

Memory interchange: --mif

export --mif <file> additionally writes the memory layer as a vendor-neutral MIF v2 document (mif_version, memories[] with id/content/created_at), so memories can be handed to any MIF-speaking memory system. import <client> --mif <file> imports the memory layer from a MIF document instead of a bundle:

agentmove export openclaw -o bundle --mif memories.mif.json
agentmove import gemini --mif memories.mif.json --apply

MIF fields with no portable equivalent (embeddings, knowledge-graph data) are dropped with a warning; a non-MIF file is a data error (exit 3).

Encrypted transport: pack / unpack

pack <bundle> [-o file] encrypts a bundle directory into a single portable .agentpack file (AES-256-GCM, key derived from the AGENTMOVE_PASSPHRASE environment variable via scrypt) so an agent can be carried across machines safely — including through untrusted channels like mail or cloud drives. unpack <file> [-o dir] decrypts it back into a bundle directory, and import -i accepts an .agentpack file directly:

AGENTMOVE_PASSPHRASE='...' agentmove export openclaw -o bundle
AGENTMOVE_PASSPHRASE='...' agentmove pack bundle -o agent.agentpack
# on the other machine
AGENTMOVE_PASSPHRASE='...' agentmove import claude-code -i agent.agentpack --apply

A missing passphrase is a usage error (exit 2); a wrong passphrase or tampered file fails authentication and is a data error (exit 3). Note that exported bundles redact likely secrets by default — encryption is for everything else (instructions, memory, persona, server lists). If you pack a bundle exported with --include-secrets, the ciphertext protects them in transit, but treat the passphrase accordingly.

agentmove diff <from> <to> [--json]

Layer-by-layer structural comparison between two clients, or between a bundle directory and a client. Output uses + (added), - (removed), ~ (changed).

agentmove clients [--json]

Lists the supported clients with their ids and default config locations (like pandoc’s --list-input-formats).

agentmove completion <bash|zsh|fish>

Prints a shell completion script covering commands, client ids, and flags. Enable it with:

eval "$(agentmove completion bash)"                    # bash / zsh
agentmove completion fish | source                     # fish

agentmove doctor [--json]

Detects which supported clients are configured on this machine and inventories what AgentMove can migrate from each (MCP servers, skills, memory entries, instructions, persona), including any format warnings.

Man page

A global install (npm i -g agentmove-cli) also links a man page: man agentmove-cli.

Global options

• --home <dir> — override the home directory (useful for testing and staging).
• --only <layers> (on export, import, convert) — migrate only the listed layers (mcp,skills,memory,instructions,persona).
• --project <dir> (on export, import, convert) — operate on the client’s project-scoped files in a repository instead of $HOME.
• --mif <file> (on export, import) — exchange the memory layer as a MIF v2 document.
• --debug (or AGENTMOVE_DEBUG=1) — print a full stack trace on unexpected errors; by default errors are a single readable line.
• --json (on export, import, convert, diff, pack, unpack, doctor, clients) — machine-readable JSON on stdout for scripts and CI: the migration plan, warnings, per-layer summary, and backup directory. After --apply, the human output also ends with a migrated: … per-layer summary line.

Exit codes

Code	Meaning
0	success
1	unexpected error (permission errors include remediation guidance)
2	usage error (e.g. unknown client)
3	bad input data (missing/corrupt bundle or config; the message includes the offending file path)