Claude Portability Manual

Use this manual when the task involves Claude Web, Claude exports, Claude memory, Claude projects, Claude skills, or restoring neuDrive context back into Claude-compatible materials.

Treat Claude portability as a category-mapping task, not just a zip-upload task. Claude's real data surface is broader than memory + skills.

For Claude skills work, read this manual before choosing import_skill, import_skills_archive, or prepare_skills_upload. If the user says "all skills", /mnt/skills/user, public + examples, workspace zip, or any other multi-skill request, do not start with import_skill and do not reduce the migration to SKILL.md only.

This manual follows Claude's current public surfaces:

• Profile preferences
• Styles
• Memory
• Standalone chats
• Projects
• Skills
• Connectors / external sources
• Export packages

What Claude Actually Has

Profile preferences

• Account-wide response preferences.
• This is the closest Claude concept to durable neuDrive profile data.

Styles

• Claude styles change how Claude formats and presents responses.
• They are not the same thing as profile preferences or project instructions.
• neuDrive does not have a first-class style entity, so preserve exact style text when it matters.

Memory

• Claude has account-level memory and project-specific memory summaries.
• Claude also has a separate memory import/export flow.
• Do not mix memory summary, manual memory edits, and chat history into one bucket.

Standalone chats

• Non-project chat history lives separately from projects.
• Incognito chats are a special case: they are not in normal chat history, and only show up in org-level export/compliance contexts.

Projects

• A Claude project can include:
  • project instructions
  • project knowledge / uploaded files
  • project chats
  • a separate project memory summary
• Keep these subtypes separate when mapping into neuDrive.

Skills

• Skills are reusable cross-chat packages.
• They are not the same thing as project knowledge.
• One skill and many skills should not use the same import path.

Connectors / external sources

• Connectors give Claude access to external apps and files.
• Usually this is setup metadata plus externally hosted content, not a portable Claude-owned file bundle.

Export packages

• Claude exported data zip: the official account export from Claude settings.
• Claude memory export: the separate memory export/import flow.
• Claude Web skills workspace zip: a full zip of /mnt/skills/user from the Claude Web sandbox.

Interface Layering Rules

• Use update_profile for durable account-wide preferences, principles, and stable working rules.
• Use save_memory for dated notes, extracted facts, scratch material, and small derived memories.
• Use create_project, log_action, and get_project to rebuild project structure manually when the imported data really belongs to a Claude project.
• Use write_file for any imported file-like data that should be preserved in neuDrive, not just for projects. If some Claude material does not fit profile, memory, project, or skill cleanly, it may still be imported under a sensible custom directory structure.
• Use import_skill as the formal public MCP path for one skill whose files can be represented as a map[path]string. Nested relative paths are allowed, so text and code files such as SKILL.md, prompts, .py, .js, .ts, .sh, .json, .yaml, .xml, and .xsd can stay on this path.
• Even on the import_skill path, import the whole skill directory, not just SKILL.md. Include every text/code file the skill needs, such as scripts/, prompts/, config files, schemas, and helper sources.
• If the user asks for all Claude skills, a full skills workspace, or any multi-skill bundle, do not use import_skill as the primary transport. Package the full directories first and choose prepare_skills_upload or, only when already known small enough, import_skills_archive.
• Use import_skills_archive only for archives already known to be small enough for one MCP tool call. For Claude Web, that means the zip must already be known to be <= 64 KB. Do not cat base64(zip), paste archive base64 into chat, or otherwise emit long archive strings into the conversation, because that can crash the conversation session.
• The agent may use write_file to patch one file, archive unsupported exports, or preserve additional imported Claude data under custom paths when no first-class neuDrive domain exists.
• For Claude Web skills zips, check the zip size before reading contents. If the zip is larger than 64 KB, or if the size cannot be checked safely, do not read or base64 it into MCP args. Do not cat base64(zip) or otherwise print long archive strings into the conversation.
• Use prepare_skills_upload plus /agent/import/skills as the prepared upload path for Claude Web zips larger than 64 KB, zips whose size cannot be checked safely, or any archive that should not be inlined into one tool call.
• After prepare_skills_upload, first POST the returned connectivity_probe_url. If it returns 200, upload directly from the Claude sandbox with the returned curl_example and do not hand the zip off to the user.
• If the probe POST fails in Claude Web, treat it as an outbound-domain restriction. Tell the user to open Settings -> Capabilities -> Code execution and file creation -> Additional allowed domains and add the neuDrive host, such as neudrive.ai on the official service. Then retry the probe and direct upload.
• If the user has already added the domain and the probe still fails in the current Claude Web conversation, explain that Claude Web may require a new conversation before the new allowed-domain setting takes effect. Ask whether they want to start a new conversation and retry, or use the fallback now.
• If the user does not add the domain, does not want to start a new conversation, or direct upload is still blocked, fall back to the returned browser_upload_url or ask the user to run the returned curl_example manually.
• All skill import paths land under /skills/<name>/... in neuDrive. Do not ask the user for a separate destination directory.
• Keep Claude exported data zip separate from skills archive flows. It currently uses /api/import/claude-data and does not have public MCP parity.
• Keep Claude memory export separate from full account export. It currently uses /api/import/claude-memory and does not have public MCP parity.

Category Map

Import Checklist

1. Inventory the Claude-side categories first and mark each one available, missing, or blocked: profile preferences, styles, memory, standalone chats, project instructions, project knowledge, project chats, skills, connectors, official exports.
2. If the user already has the official Claude exported data zip, route that package to /api/import/claude-data first and explicitly note that there is no public MCP equivalent yet.
3. If the user has exported Claude memory text, prefer /api/import/claude-memory; otherwise split the content into durable profile rules for update_profile and smaller derived notes for save_memory.
4. Import profile preferences with update_profile. Do not bury stable account-wide rules inside scratch memory.
5. Import styles by extracting the stable formatting and communication rules into update_profile, and preserve the exact style text as an archive note when the exact wording matters.
6. Handle standalone chats separately from memory summaries. Use the full export path when available; otherwise preserve them as archive notes or files instead of pretending there is direct conversation parity.
7. Rebuild project instructions into /projects/<name>/context.md with create_project plus write_file.
8. Rebuild project knowledge into /projects/<name>/... when doing manual reconstruction. If using the official full export zip, note that the current importer writes project docs under /skills/claude-<project>/... rather than creating first-class neuDrive projects.
9. Preserve project chats and project memory summary as archive notes or conversation files unless the full export importer is being used. Do not claim first-class public MCP parity here.
10. For skills, choose the path by payload shape instead of by platform name alone:
    • one small or moderate skill, all files are text/code and can be represented as map[path]string -> import_skill(name, files)
    • text/code files can still include nested paths like scripts/run.py, prompts/review.txt, data/schema.xsd, or bin/tool.sh
    • do not simplify a skill to only SKILL.md; if scripts/, prompts, config, schemas, or helper files exist, include them too
    • if the user asks for all skills, /mnt/skills/user, public + examples, a workspace export, or any other multi-skill batch -> do not use import_skill; package one full archive first
    • for Claude Web zips, check the zip size before reading contents; if the zip is larger than 64 KB or the size cannot be checked safely -> prepare_skills_upload
    • many skills, binary assets, or any archive where exact bytes matter -> import_skills_archive only when the archive is already known small enough for one MCP tool call
11. All skill import flows land under /skills/<skill-name>/... in neuDrive. A fallback upload page or upload command should target the /skills root by default and should not ask the user to pick another destination directory.
12. Do not create or recommend markdown-only or SKILL.md-only skill exports. Preserve the entire skill directory and keep SKILL.md, scripts, prompts, config, schemas, and binary assets together.
13. If the user still wants pure inline MCP transport, split a Claude Web archive by top-level skill directories only when each resulting zip is known to be <= 64 KB. Do not split one skill directory into partial fragments unless a future chunked import flow exists.
14. If prepare_skills_upload is chosen, package one complete zip first, then try the direct sandbox upload path before involving the user.
15. First POST the returned connectivity_probe_url.
16. If that POST returns 200, use the returned curl_example from the Claude sandbox and finish the upload directly.
17. If the probe POST fails in Claude Web, tell the user to add the neuDrive host to Settings -> Capabilities -> Code execution and file creation -> Additional allowed domains, then retry the probe and direct upload.
18. If the user has already added the domain and the probe still fails in the current Claude Web conversation, tell the user that the new setting may only take effect in a new conversation. Ask whether they want to start a new conversation and retry, or use the fallback now.
19. If the user does not add the domain, does not want to start a new conversation, or direct upload is still blocked, use browser_upload_url for ordinary users or ask the user to run the returned curl_example manually.
20. When a Claude export contains useful files or metadata that do not fit a first-class neuDrive domain, preserve them with write_file under a clear custom directory structure instead of forcing them into a project.
21. Preserve connectors / external sources as setup metadata, selected-file manifests, or project notes. Do not claim that third-party service data has been imported unless those files were actually captured.
22. End with a report that lists imported, archived, and blocked items, with the exact interface used for each category.

Preferred Claude Web skills archive flow

When /mnt/skills/user exists and the goal is to move Claude Web skills into neuDrive, prefer a full archive:

cd /mnt/skills/user
zip -r /mnt/user-data/outputs/neudrive-skills.zip .

Then choose the transport path that matches the payload:

1. If this is really one text/code skill and the files can be represented as strings, skip the zip and call import_skill. The files map may include nested paths like scripts/main.py.
   • include the full skill directory contents, not just SKILL.md
   • do not use this branch when the user asked for all skills, /mnt/skills/user, public + examples, or any multi-skill workspace
2. Otherwise, stat the zip before reading contents.
   • if the Claude Web zip is larger than 64 KB, do not read the bytes
   • if the Claude Web zip size cannot be checked safely, do not read the bytes
3. If the Claude Web zip is larger than 64 KB, or the size is unknown, call prepare_skills_upload.
4. First POST the returned connectivity_probe_url.
   • if the probe returns 200, use the returned curl_example from the Claude sandbox and finish the upload directly
   • if the probe fails in Claude Web, tell the user to add the neuDrive host to Settings -> Capabilities -> Code execution and file creation -> Additional allowed domains, then retry
   • if the user has already added the domain and the probe still fails in the current Claude Web conversation, explain that they may need to start a new conversation for the setting to take effect
   • if the user does not add the domain, does not want to start a new conversation, or direct upload is still blocked, fall back to the returned browser_upload_url or ask the user to run the returned curl_example manually
5. Only if the Claude Web zip is already known to be <= 64 KB, read the zip bytes, base64-encode the archive, and call import_skills_archive with platform="claude-web" and the original archive name.
6. If the user insists on pure inline MCP transport, split the archive by top-level skill directories only when each resulting zip is known to stay within the same 64 KB limit.

All of these paths still import into the neuDrive /skills root.

Prepared upload flow

When prepare_skills_upload is used for Claude Web skill import, the agent should switch from "inline MCP import" mode to "prepared upload" mode.

Preferred flow:

1. Package one complete skills zip that preserves the original skill directories.
   • do not omit scripts/, prompts, config files, schemas, fonts, or other helper assets
   • do not replace a full skill with a SKILL.md-only shortcut
2. Call prepare_skills_upload.
3. First POST the returned connectivity_probe_url.
4. If the probe returns 200, upload the zip directly from the Claude sandbox with the returned curl_example.
   • this is the preferred path when Claude can reach the neuDrive host directly
   • no user download or browser handoff is needed in this case
5. If the probe fails in Claude Web, explain that Claude likely blocked outbound POST to the neuDrive host.
6. Tell the user exactly how to unblock direct upload:
   • open Settings -> Capabilities -> Code execution and file creation
   • add the neuDrive host, such as neudrive.ai, under Additional allowed domains
   • then retry the same probe and direct upload flow
7. If the user already added the domain and the same Claude Web conversation still cannot POST:
   • explain that the new allowed-domain setting may only take effect in a new conversation
   • ask whether they want to start a new conversation and retry, or use the fallback now
8. If the user does not add the domain, does not want to start a new conversation, or direct upload is still blocked:
   • use the returned browser upload link for ordinary users
   • or ask terminal-comfortable users to run the returned curl_example manually
9. When using the browser fallback, first make the generated zip available to the user as a downloadable file through the platform's file handoff/download mechanism.
10. Make it explicit that the browser upload page imports into the neuDrive /skills root by default. The user should not choose or type another destination path.

Suggested agent wording:

I packaged your Claude Web skills as a full zip, including all files under each skill directory, not just SKILL.md. I first tested whether Claude can POST directly to the neuDrive host. If that probe succeeds, I will upload the zip directly from the Claude sandbox with the prepared curl command. If Claude blocks that POST, please open Settings -> Capabilities -> Code execution and file creation -> Additional allowed domains, add the neuDrive host, and then I can retry the direct upload. If you already added it and this same Claude Web conversation still cannot POST, the new setting may require a new conversation before it takes effect. In that case, you can start a new conversation and retry, or I can fall back to the browser upload page or give you the curl command to run yourself.

If the token response includes both a browser upload link and a curl example, keep the direct sandbox curl path as the first choice after a successful probe. If direct upload stays blocked, use the browser path for ordinary users and the curl path for terminal-comfortable users. Do not ask ordinary users to manually build multipart requests if a browser upload link exists.

Export Checklist

1. Inventory neuDrive data by Claude category, not just by file location: profile preferences, styles, memory, standalone chats, project instructions, project knowledge, project chats, skills, connectors.
2. Export durable profile rules from read_profile into Claude-ready profile preferences.
3. If the user wants a Claude style, derive it from the formatting and communication rules in profile or archived notes, and mark it as a manual Claude style setup step.
4. Export memory as either:
   • a concise Claude memory import text block, or
   • manual notes the user can paste into Claude's memory import flow.
5. Treat /memory/conversations/*.md as chat archive/reference material unless the user explicitly wants manual conversation restoration notes.
6. Rebuild Claude project instructions from /projects/<name>/context.md.
7. Rebuild Claude project knowledge from /projects/<name>/... and related files.
8. Rebuild skills one directory at a time for single-skill cases, or as a full archive for many skills or asset-heavy skills.
9. Recreate connectors / external sources as manual setup instructions and selected-file manifests. Do not claim that neuDrive can restore third-party app connections natively.
10. Mark every manual restore step explicitly. Do not claim native Claude restore parity where it does not exist.
11. End with a report that lists generated materials, manual follow-ups, and remaining gaps.

Current Gaps

• There is no public MCP equivalent for /api/import/claude-memory.
• There is no public MCP equivalent for /api/import/claude-data.
• The full Claude export importer currently lands memory under /memory/claude/memory.md, conversations under /memory/conversations/*.md, and project documents under /skills/claude-<project>/...; this is useful but not full first-class project parity.
• The current full Claude export importer does not map users.json into a first-class neuDrive domain.
• There is no first-class public MCP importer for standalone chats, project chats, or project memory summaries.
• import_skill is the right path for one text/code skill whose files can be represented as strings, including nested paths such as scripts/*.py. Even then, import the whole skill directory, not just SKILL.md.
• For Claude Web, do not inline a skills zip into import_skills_archive unless it is already known to be <= 64 KB. If the size is unknown, default to prepare_skills_upload.
• prepare_skills_upload plus /agent/import/skills is the prepared upload path for Claude Web archives that should not be inlined into one tool call.
• When prepare_skills_upload is used, direct sandbox upload via the returned curl_example is preferred after the returned connectivity_probe_url succeeds.
• If the probe fails in Claude Web, the user may need to add the neuDrive host to Settings -> Capabilities -> Code execution and file creation -> Additional allowed domains before retrying direct upload.
• If the user already added the domain and the current Claude Web conversation still fails the probe, the new setting may require a new conversation before retrying direct upload.
• If the user still wants pure inline MCP transport, split by whole skill directories only when each resulting archive is known to stay within the same 64 KB limit. True chunked import for one oversized skill does not exist yet.

Prompt Template

Use or adapt this prompt when another agent needs to execute Claude portability work:

Read /skills/portability/claude/SKILL.md first. Inventory the Claude-side categories as profile preferences, styles, memory, standalone chats, project instructions, project knowledge, project chats, skills, connectors, and official exports. Map each category to the nearest neuDrive domain instead of mixing them together. Use update_profile for durable account-wide rules, save_memory for smaller derived notes, create_project for true Claude project reconstruction, and write_file for any additional imported Claude files or metadata that should be preserved even when they do not fit a first-class neuDrive domain. The agent may design a sensible custom directory structure for those files. Use import_skill only for one text/code skill whose files can be represented as strings. If the user asks for all skills, /mnt/skills/user, public + examples, or any multi-skill workspace, do not use import_skill; package the full directories first. Never simplify a skill to SKILL.md only; include the whole skill directory and all needed scripts/, prompts, config, schemas, and assets. For Claude Web skills zips, check zip size before reading contents. If the zip is larger than 64 KB, or if the size cannot be checked safely, do not read or base64 it into MCP args; switch to prepare_skills_upload instead. After prepare_skills_upload, first POST the returned connectivity_probe_url. If it returns 200, use the returned curl_example to upload directly from the Claude sandbox. If that POST fails in Claude Web, tell the user to add the neuDrive host to Settings -> Capabilities -> Code execution and file creation -> Additional allowed domains, then retry. If the user already added the domain and the probe still fails in the current Claude Web conversation, explain that the new setting may require a new conversation before it takes effect, and ask whether they want to start a new conversation and retry or use the fallback now. If the user does not add it, does not want to start a new conversation, or direct upload is still blocked, fall back to the returned browser upload link or ask the user to run the returned curl command manually. Only use import_skills_archive when the zip is already known to be <= 64 KB. All skill imports land under /skills/<name>/... in neuDrive, and the browser upload page targets the /skills root by default. Preserve unsupported structures as archive notes, structured metadata, or custom file trees instead of dropping them, and finish with imported, archived, and blocked items plus the exact interface used for each category.

{{CURRENT_USER_SNAPSHOT}}