Issues Accessing DuckDB Python Extension in Claude.ai Due to Domain Allowlist Restrictions in Anthropic Cloud | MotherDuck Community

MotherDuck Community Icon

Issues Accessing DuckDB Python Extension in Claude.ai Due to Domain Allowlist Restrictions in Anthropic Cloud | MotherDuck Community

Bev T.
·
·
Hi Nick. You are correct in that you cannot use the MotherDuck extension from within the claude.ai. As to why? I asked Claude itself: The why, briefly: outbound network from Claude's sandbox is locked down by design — it's a safety boundary so code Claude runs can't reach arbitrary hosts to exfiltrate data or call paid APIs without explicit user authorization. Adding MotherDuck would mean carrying credentials, which the sandbox is intentionally engineered not to do. May I ask where these csv/parquet files sit or would be exported to?
Nick C.
·
(edited)·
I am fine with giving this claude session my read-only motherduck token. I understand the security implications. The files are uploaded by my coworker for analysis. eg a messy table of people data. I want to allow them to enrich this csv with data that lives inside our MD, and then re-export the CSV. I was helped by Connor C. who gave the helpful workaround of telling claude to use the MCP to CREATE TABLE with the resulting data, and then the user can download the csv from the md web ui. But, that has several downsides: 1. it only works for exporting data 2. normy users aren't going to know to request this 3. it is not a great ux to ask a user to exit their claude session to go login to motherduck.
Bev T.
·
·
Have you tried doing this through Claude Code? Unfortunately the restrictions in place make the upload unworkable any other way.
Nick C.
·
(edited)·
yes, claude code works fine as I said because it downloads the pypi package/CLI/md extensions just fine from duckdb.org because there are no network restrictions. I don't have all the context here, but I THINK what motherduck needs to do is:
host the various built motherduck extension binaries as github releases. Then all downloads come from the github.com domain are on the sandbox allowlist. I think there will be one extension for each of the various builds, eg linux_amd, linux_arm, macos_x86, etc
publish an agent skill that describes to the agent how to 1. find which build it needs 2. how/where to download it 3. how to load/install it into duckdb manually. This will avoid duckdb going with the default downloading algorithm and attempting download from duckdb.org
somehow get this skill into user's context windows:
add this to your existing skills repo?
publish somewhere else?
then also this ideally should be more discoverable. If you added something to your MCP (eg a install_api() tool??), then any agents already using that MCP would autodiscover this skill.
Even more ideal would be if we actually sidestepped the whole problem by doing what I suggest in https://github.com/duckdb/community-extensions/issues/1887 and helped move the entire duckdb ecosystem so the default extension download domain was on pypi or github or npm or one of the allowlisted domains. This requires the most buy in from the entire duckdb community, but solves it for everybody and for all extensions.
Leonardo V.
·
·
Hey Nick C.! Also looking into this with our AI team. If we would publish the extension via an accessible method (e.g. pypi), and it'd be downloaded in the sandbox, would the sandbox be able to connect and talk to MotherDuck?
Nick C.
·
·
Oh, I see what you mean, probably not. The motherduck extension would need to make network requests to a motherduck domain to actually make queries. Sooooo I think we are dead in the water there. I guess the only way to salvage this is to improve the experience around using the MCP, so that it can handle more of the usecases that I want a CLI for. eg improve it's tool description so when the user says "give me a CSV" then the LLM chooses to CREATE TABLE in motherduck and tell the user about it instead of trying to fetch all the data through the context window.
👍1
Leonardo V.
·
·
DMing you to talk more about this
Nick C.
·
(edited)·
wait, hold the phone, this might be solved. It looks like in the settings on claude.ai I can select "all domains" in the allowlist. I know I looked here in the last week, and I SWEAR that wasn't an option then. Not sure if anthropic is shipping like mad, or if my account didn't have the right permissions, etc. But I can now download the duckdb cli, python package, motherduck extension, and then use those to query motherduck. I think my advice around adjusting the MCP descriptions is still good advice though. Unless I suggest that to the LLM in a custom skill, it still LOVES to try to download huge CSVs using the MCP. Thanks for the time and effort here everyone, I really appreciate it!
👀1
Leonardo V.
·
·
This is great Nick C.! We will be adding to our docs as well!
❤️1
Nick C.
·
·
excellent, thanks!