What is the Weld MCP server?

A Model Context Protocol server that connects AI assistants to your Weld workspace. Once it is added, an assistant like Claude, ChatGPT or Cursor can list and debug your syncs, write and publish SQL models, set up reverse ETL and query your warehouse, all from a conversation.

Which AI clients does it work with?

Any client that supports MCP over Streamable HTTP. Claude Desktop, claude.ai, Claude Code, Codex, ChatGPT Desktop and VS Code with GitHub Copilot log in with OAuth. Cursor, Windsurf and your own scripts use an API key in the x-api-key header.

What can an agent do with it?

There are 66 tools covering connections, ELT syncs and streams, custom reports, transforms, reverse ETL, orchestrations and warehouse queries. The full list, with what each tool does, is in the MCP server docs.

Should I use OAuth or an API key?

OAuth wherever your client supports it. There is no key to store, and the agent acts with exactly the permissions your Weld user has. An API key is scoped to one workspace but currently has full admin rights whoever created it, so keep it in an environment variable or a secrets manager. Role-scoped keys are on the way.

Can an agent change or delete things?

Yes, the tools include creating, updating and deleting syncs, models and reverse ETL syncs, within the permissions it has. Most MCP clients ask you to approve a tool call before it runs, and you can turn off the tools you do not want an agent to use.

Where does my data go?

Your data stays in your warehouse. When an agent runs a query, up to 1,000 result rows go back to your AI client, so they reach whichever model provider that client uses. Syncs, models and credentials stay in Weld.

Which plans include it?

The MCP server is included on the Premium plan and above. On the Basic plan you can try it free for two weeks, and after that add it for €99 a month.

Are there rate limits?

Yes, 60 requests a minute per account and 60 per credential, plus 600 a minute per IP address. Every JSON-RPC call counts, including the initial handshake, and a warehouse query takes at least two calls. Past the limit the server returns 429 until the minute is up.

We have more than one workspace. How does that work?

With OAuth, connect with the standard URL and tell the agent which workspace to use. It calls list_accounts to see the workspaces you can reach and works in the one you pick. An API key already belongs to one workspace.

Let your AI agents run the data stack

The Weld MCP server gives Claude, ChatGPT, Cursor and any MCP client 66 tools for your syncs, models, reverse ETL and warehouse, acting with the permissions you already have.

Start for free Read the docs

OAuth, no key to store
Acts with your permissions
SOC 2 Type II

Claude CodeOpus 5

SessionsConnectedWeld MCP60+ toolsWeld APIRESTGitHub2 repos

shopify.orders has not landed since Tuesday. Work out why, fix it, and open a PR for the model change.

Two syncs are failing. Shopify dropped a column the model still selects.

●weld·list_failed_syncs

⎿shopify.orders schema drift

stripe.charges auth expired

●weld·request_elt_stream_full_refresh

⎿shopify.orders queued, 30 days

●github·create_pull_request

⎿weld/analytics #482 · drop discount_code

✓Sync re-running, PR open for review

Weld MCPWeld APIGitHub

>Ask Weld to move, model or activate your dataMCP

Trusted by data and AI teams

Just ask

Ask in plain language. The agent picks the tools.

No IDs to copy and no API requests to write. Your agent finds the right sync, table or model, calls the tools it needs and tells you what it did.

Debug a failing sync

“Why did the Stripe numbers stop updating?”

1list_failed_syncs
2get_elt_source_stream_history
3request_elt_stream_run

Charges failed on an expired token. Rerun queued once you reconnect.

Build a model

“Model monthly revenue by channel from Shopify and Google Ads.”

1search_available_references
2get_reference_schema
3create_transform
4publish_transform

analytics.revenue_by_channel published to the warehouse.

Activate data

“Send customers with a lifetime value over €500 to HubSpot.”

1get_reverse_etl_object_types
2get_reverse_etl_destination_schema
3create_reverse_etl_sync
4activate_reverse_etl_sync

1,284 contacts mapped, syncing every hour.

Answer a question

“How many orders did we take last week, by country?”

1search_available_references
2run_query
3get_query_result

3,912 orders. Denmark leads, Germany up 18%.

Tools

66 tools across the whole pipeline. From the first sync to the last query.

Everything you can do in Weld, an agent can do through the MCP server: set up and debug syncs, write and publish models, push data back to your tools and query the warehouse.

ELT syncsCreate, schedule and debug the syncs that land your sources in the warehouse.

list_failed_syncsFailing syncs, with the error and last success
list_elt_syncsEvery sync, filterable by connection
get_elt_syncOne sync's details
create_elt_syncA new sync for a connection
update_elt_syncChange the interval or settings
delete_elt_syncRemove a sync, and its tables if you want
enable_elt_syncTurn scheduling on
disable_elt_syncTurn scheduling off
get_elt_sync_statusCurrent state, per stream
get_elt_sync_available_streamsStreams that can be added
add_elt_sync_source_streamsAdd streams to a sync
get_elt_sync_historyLatest result for each stream
get_elt_source_stream_historyUp to 100 runs of one stream
create_custom_reportsGAQL and ShopifyQL reports on a sync

StreamsRun, refresh and inspect individual tables inside a sync.

list_elt_streamsEvery stream, filterable by sync
list_elt_stream_runsRuns from the last three months
get_elt_stream_runOne run's details
request_elt_stream_runRun a stream now
request_elt_stream_full_refreshMake the next run a full refresh
request_elt_stream_full_refresh_with_history_tableFull refresh and reset history
cancel_elt_stream_full_refreshBack to incremental
delete_elt_streamRemove a stream

TransformsWrite, version and publish the SQL models in your warehouse.

list_transformsFilter by status, name or folder
get_transformSQL and materialization settings
list_transform_versionsRevision history, newest first
create_transformA new model, published if you like
bulk_create_transformsUp to 50 models in one call
update_transformChange a model
publish_transformPublish to the warehouse
delete_transformRemove a model and its view

QueriesRead the warehouse, with the schema checked before the SQL is written.

search_available_referencesFind tables, with columns and freshness
list_available_referencesEvery source view and published model
get_reference_schemaColumn names, types and nullability
run_queryRun SQL, up to 1,000 rows back
get_query_statusPoll until it is done
get_query_resultFetch the rows

Reverse ETLSend models back to your CRM and ad platforms, and clean up what failed.

list_reverse_etl_syncsEvery reverse ETL sync
get_reverse_etl_syncOne sync's details
create_reverse_etl_syncFrom a model to a destination
update_reverse_etl_syncMappings, schedule or settings
delete_reverse_etl_syncRemove a sync
activate_reverse_etl_syncStart it on its schedule
deactivate_reverse_etl_syncPause it
request_reverse_etl_sync_runRun it now
get_reverse_etl_object_typesContact, Company and the rest
get_reverse_etl_available_operationsWrite modes for an object
get_reverse_etl_destination_schemaFields you can map to
get_reverse_etl_settings_schemaSettings a destination needs
list_reverse_etl_sync_runsRuns, newest first
get_reverse_etl_sync_runOne run's details
get_reverse_etl_sync_failed_record_errorsRecords that did not land
get_reverse_etl_sync_failed_record_countHow many failed
retry_failed_reverse_etl_sync_recordsRetry chosen records
retry_all_failed_reverse_etl_sync_recordsRetry them all

OrchestrationsSee what ran in what order, and start a run.

list_orchestrationsEvery orchestration
get_orchestrationOne orchestration's details
list_orchestration_runsRuns, newest first
get_orchestration_runOne run's details
request_orchestration_runRun it now

WorkspaceWhich workspace you are in, what is connected and what your plan includes.

list_accountsWorkspaces you can reach
get_accountWarehouse and plan details
list_connectionsEvery connected source
get_connectionOne connection's details
get_connection_elt_settingsSettings a new sync needs
delete_connectionRemove a connection
list_integrationsConnectors on your plan

Setup

Connected in two minutes. One URL, every client.

Point your client at the endpoint and log in with your Weld account. Clients without OAuth take an API key in a header instead.

Streamable HTTP. No WebSockets, no local server to run.

Claude

1In Claude Desktop or claude.ai, open Settings → Connectors
2Click Add custom connector and name it Weld
3Paste https://connect.weld.app/mcp and click Add
4Log in with your Weld account in the window that opens

Governance

Your agent, your permissions. Nothing it could not do in the app.

Acts as you

With OAuth the agent has exactly the permissions your Weld user has. Nothing more.

Every workspace you can reach

Connect once with the standard URL and tell the agent which workspace to work in. It looks them up with list_accounts.

Keys for headless use

API keys are for scripts and clients without OAuth. Each key is scoped to one workspace and managed in Settings → API keys.

Rate limited

60 requests a minute per account and per credential. Past that the server answers 429, and the window resets each minute.

Stateless

No session to manage or expire. Every request is authenticated on its own, by its OAuth token or its API key.

Audited

Weld is SOC 2 Type II audited and GDPR compliant. Your data stays in your warehouse, and the agent reads it through Weld.

Security at Weld

FAQ

Questions we get asked. Before the first prompt.

Still have a question?

Talk to us

Connect your agent.
Ask it about your data.

Add the Weld MCP server to Claude, ChatGPT, Cursor or any MCP client and it works with your syncs, models and warehouse in minutes. Free for 14 days.

Start for free Talk to us