Painless migration from OpenClaw / Hermes / paperclip

duduclaw migrate-from moves an existing OpenClaw, Hermes, or paperclip setup into DuDuClaw with one command. It defaults to preview mode: it only prints what would be imported, what would be skipped, and why. Once the plan looks right, add --apply to actually write anything.

# Preview (writes nothing)
duduclaw migrate-from openclaw

# Apply after reviewing the plan
duduclaw migrate-from openclaw --apply

Command

duduclaw migrate-from <openclaw|hermes|paperclip> [--source <path>] [--apply] [--rename]

Flag	Effect
(none)	Preview the migration plan; writes nothing.
--source <path>	Source directory. openclaw/hermes have defaults; paperclip requires it.
--apply	Actually perform the write.
--rename	On an agent-id collision, import under an -imported suffix instead of skipping.

Every item is tagged with a status:

• IMPORTED — imported (or will be).
• PARTIAL — partially imported, or needs manual confirmation (e.g. a non-Claude model).
• SKIPPED(reason) — skipped, with a reason (source file missing, parse failure, security block, etc.).
• CONFLICT(reason) — the target already has a value; the existing setting is left untouched.

The overall result rolls up to COMPLETE / DEGRADED / PARTIAL. After an apply run, a full report is written to ~/.duduclaw/imported/<platform>/migration-report.md. Every token value is shown masked as “first 4, last 4” — never in plaintext on screen or in the report.

Per platform

OpenClaw (~/.openclaw)

duduclaw migrate-from openclaw            # defaults to ~/.openclaw
duduclaw migrate-from openclaw --source /path/to/.openclaw --apply

Reads openclaw.json (JSON5) and imports:

• Agents: agents.list[] (or a single default main), each with its workspace persona (SOUL.md) and memory (MEMORY.md / USER.md / bullet entries from memory/*.md).
• Channel tokens: channels.telegram.botToken, channels.discord.token, channels.slack.botToken + appToken (written encrypted into config.toml). WhatsApp is a linked device and technically cannot be transferred, so it is SKIPPED.
• Model: agents.defaults.model.primary (with the anthropic/ prefix stripped).
• Anthropic API key: read from the env section and ~/.openclaw/.env. Keys for other providers are SKIPPED.
• Cron: the legacy cron/jobs.json (parsed defensively). The newer SQLite cron schema is unvalidated and SKIPPED.
• Skills: SKILL.md folders located by OpenClaw’s own precedence order (scanned before install).

The legacy directory names ~/.moltbot and ~/.clawdbot are also supported.

Hermes (~/.hermes)

duduclaw migrate-from hermes --apply
# migrate a non-active profile:
duduclaw migrate-from hermes --source ~/.hermes/profiles/<name> --apply

Hermes is a single-agent platform, so this produces one DuDuClaw agent (id hermes). Imports:

• Model: model.default from config.yaml.
• Channel tokens (from .env): TELEGRAM_BOT_TOKEN, DISCORD_BOT_TOKEN, SLACK_BOT_TOKEN + SLACK_APP_TOKEN. EMAIL_* channels are not yet supported in v1 and are SKIPPED.
• Persona / memory: SOUL.md, memories/MEMORY.md, memories/USER.md.
• Cron: cron/jobs.json (parsed defensively).
• Only the active profile is migrated; other profiles are listed as SKIPPED, with a prompt to migrate each one individually via --source.

paperclip: via the official export

paperclip’s data lives in an embedded PostgreSQL instance, and DuDuClaw does not connect to that database directly. Export from the paperclip side first:

paperclipai company export <company-id> --out ./export \
  --include company,agents,projects,issues,tasks,skills

duduclaw migrate-from paperclip --source ./export --apply

--source is required (omitting it prints the instructions above). Imports:

• Agents: frontmatter (name/title/reportsTo/skills) from agents/<slug>/AGENTS.md becomes a DuDuClaw agent, and the body becomes SOUL.md. reportsTo maps directly to reports_to (agents are created in topological order by hierarchy; a detected cycle falls back to no superior for every agent involved and is marked PARTIAL).
• Tasks: tasks/<slug>/TASK.md becomes a Task Board entry; recurring becomes cron.
• Skills: skills/<slug>/SKILL.md goes into the agent’s SKILLS/ (scanned first).
• COMPANY.md becomes a shared wiki page at shared/wiki/imported/paperclip-company.md.
• The official paperclip export format contains no secrets (channel tokens, API keys, DB ids), so channels and keys are always SKIPPED.

Security and data preservation

• Skills are scanned before install: every SKILL.md first passes through duduclaw-security’s prompt-injection scanner (6 rule categories). A hit fails closed: nothing is installed, and it is marked SKIPPED(security). Imported skills keep skill_auto_activate at false (the safe default).
• Never overwritten: an existing agent with the same id is SKIPPED (or use --rename); a channel token / API key already present in config.toml is a CONFLICT, and the original value is left untouched.
• Tokens land encrypted: channel tokens and API keys are encrypted with AES-256-GCM before being written into config.toml — never in plaintext.
• No data loss: v1 does not parse conversation history into sessions.db, but --apply archives the original session / conversation files verbatim to ~/.duduclaw/imported/<platform>/raw/ for later reference.

v1 non-goals (honest boundaries)

1. Ingesting conversation history into the database (verbatim archiving only).
2. OpenClaw’s newer SQLite cron / auth-profiles (schema unvalidated).
3. WhatsApp linked-device credentials (bound to the device, cannot be transferred).
4. Hermes profiles other than the active one (migrate each individually via --source).
5. Reading paperclip’s Postgres directly (use the official export instead).
6. External memory backends (Honcho / Mem0 / QMD / LanceDB).

FAQ

Q: Does preview mode change anything? No. Without --apply nothing is written and no channel is started.

Q: I hit a CONFLICT partway through — what do I do? CONFLICT means the target already has a value, and it was left alone to protect your existing configuration. To replace it with the imported value, manually remove the old value from config.toml first and rerun, or import as a separate agent with --rename.

Q: What happens with a non-Claude model? It is kept as-is in [model] preferred and marked PARTIAL, prompting you to manually confirm which runtime it maps to (codex / gemini / openai_compat). DuDuClaw will not guess on your behalf.