Duckle MCP Server: Build and Run ETL Pipelines from Claude and Cursor

• Browses the 400+ pipeline building blocks

  Lists the sources, transforms, sinks, control, quality and custom-code components Duckle has. You can filter by kind or search by a word, so the assistant picks a real component instead of guessing one.

  Try: “Which Duckle components can read from Postgres?”

  Read · list_components

• Says which data assets are out of date

  Checks each asset in a workspace against the freshness limit you declared for it. An asset with no limit shows as unknown, not fresh, so a missing rule is never read as good news.

  Try: “Which assets in my Duckle workspace are stale, and since when?”

  Files · asset_freshness

• Answers what a component can do

  Looks up which sources can run custom SQL or read incrementally, which sinks offer which write modes, and which components send rejected rows aside. The answer comes from the same manifests the editor uses.

  Try: “Which sinks support upsert?”

  Read · component_capabilities

• Shows the settings one component takes

  Returns the form fields and the input and output ports of one component id. The assistant uses it to fill in the right property names before it builds a pipeline.

  Try: “What settings does src.csv take?”

  Read · get_component_schema

• Checks a pipeline and saves it as a file

  Compiles a pipeline first and writes it only if it compiles. With a workspace it saves pipelines/<id>.json and lists it in the workspace so it shows in the app; with a directory it saves a loose file.

  Try: “Create a pipeline in ./etl that reads orders.csv and writes a Parquet file.”

  Write · create_pipeline

• Changes part of a saved pipeline

  Merges a partial change into a pipeline already on disk, matching nodes and edges by id, so the assistant can change one setting without resending everything. It validates the result and then overwrites the file.

  Try: “In my orders pipeline, change the CSV sink to overwrite mode.”

  Write · update_pipeline

• Compiles a pipeline without running it

  Turns a pipeline into SQL and returns the SQL for each stage, or a structured error that says what failed. Nothing is executed.

  Try: “Validate the orders pipeline and show me the SQL.”

  Read · validate_pipeline

• Runs a pipeline on your machine

  Runs a pipeline with the DuckDB engine and returns each node's status, row counts, errors and a small preview. Setting a target node stops the run there, so later sinks do not write.

  Try: “Run the orders pipeline up to the join node and show me the rows.”

  Runs code · run_pipeline

• Runs the pipeline tests

  Runs the *.test.json files in a workspace's tests folder, or only the files you name. Each case checks the rows one node produces against a fixture, which catches a step that compiles but computes the wrong thing.

  Try: “Run all tests in my Duckle workspace and list the failures.”

  Runs code · run_tests

• Traces columns back to their sources

  Maps every output column of every node back to the source columns it comes from. It reads the pipeline and writes nothing.

  Try: “Where does the revenue column in the final sink come from?”

  Read · pipeline_lineage

• Gives a pass or fail check of a pipeline

  Compiles the pipeline, resolves lineage and flags structural risks such as joins without keys, unconnected nodes, a sink with no input, or no sink at all. It returns one verdict the assistant can act on.

  Try: “Verify my orders pipeline before I deploy it.”

  Read · verify_pipeline

• Finds SQL errors before a run

  Binds the SQL of each node against its upstream columns without running it, and reports DuckDB's message, the position, and the column it suggests for a typo. A source query that goes to a remote system is reported as not checked.

  Try: “Check the SQL in the join node of my pipeline.”

  Read · check_node_sql

• Runs, retries or cancels a partitioned backfill

  Creates, lists, retries or cancels a backfill that re-runs a pipeline over partitions. Each slice is a normal run with its own id, and a dry run shows the partitions without queueing anything.

  Try: “Do a dry run of a backfill for January 2026 on the orders pipeline.”

  Runs code · backfill

• Suggests what to type next in a node's SQL

  Returns ranked suggestions for a cursor position in a node's SQL: upstream columns with types, readable relations, declared parameters, and DuckDB functions and keywords. It never runs the SQL or touches a source.

  Try: “What columns can I use at the end of the filter node's WHERE clause?”

  Read · complete_node_sql

• Suggests data contracts and flags personal data

  Profiles a pipeline's columns and proposes personal-data tags, based on column names, plus required-column checks for sources. The result can be merged back with the update tool.

  Try: “Which columns in this pipeline look like personal data?”

  Read · suggest_contracts

• Shows what breaks if a column changes

  Inverts column lineage so each source column lists the downstream columns built from it. You can narrow it to one column and one node.

  Try: “What depends on the email column from the customers source?”

  Read · pipeline_impact

• Shows what breaks across all pipelines

  Joins the pipelines in a workspace through the tables, files and topics they name. It returns the pipelines that read an asset, what they write, everything downstream, and who to tell when an owners file exists.

  Try: “What breaks if I rename the analytics.events table?”

  Files · workspace_impact

• Compares two versions of a pipeline

  Reports nodes added, removed or changed, by component, settings or compiled SQL, plus edges added or removed, with a summary. It only compiles and reads.

  Try: “Diff my pipeline against the version in the other branch.”

  Files · diff_pipelines

• Scores how trustworthy a pipeline is

  Gives a 0 to 100 score and a letter grade, and lists every deduction: compile errors, structural risks and columns that look like personal data but are not tagged or masked. Turning on drift checking also reads each source's live schema.

  Try: “Give me a trust report for the billing pipeline.”

  Read · trust_report

• Detects changed columns in your sources

  Reads each source's live schema and compares it with the schema the pipeline declared. It reports missing columns, new columns and type changes, with a breaking or non-breaking verdict.

  Try: “Has the schema of any source in my orders pipeline drifted?”

  Files · schema_drift

• Lists the pipeline files in a folder

  Lists the pipeline .json files in a directory with the number of nodes and edges in each.

  Try: “What pipelines are in ./etl?”

  Files · list_pipelines

• Opens a pipeline file

  Reads one pipeline .json file and returns its content so the assistant can explain or edit it.

  Try: “Explain what the pipeline in ./etl/orders.json does.”

  Files · read_pipeline

• Lists a pipeline's saved resume points

  Shows the state a pipeline resumes from: incremental watermarks, DuckLake snapshot ids, Kafka offsets, spool positions and tumbling-window buffers. Each entry says whether it can be set by hand or only cleared.

  Try: “What watermarks does the orders pipeline hold?”

  Files · backfill_list

• Moves a watermark or snapshot

  Sets an incremental watermark or a DuckLake snapshot id so the next run starts from there. It refuses nodes that hold Kafka offsets, spool positions or windows, because overwriting those destroys what they hold.

  Try: “Set the orders watermark back to 2026-10-01.”

  Write · backfill_set

• Erases a node's saved resume point

  Removes a node's saved state so it starts over on the next run. This is not always a full reload: a Kafka source set to start from latest skips the backlog that is already in the topic.

  Try: “Clear the saved state on the orders source node.”

  Delete · backfill_clear

• Lists quality baselines in a workspace

  Shows every quality baseline node: how many profiles it accepted, what the last run concluded, and whether a new measured profile is waiting to be accepted.

  Try: “Which baselines are waiting for approval?”

  Files · baseline_list

• Compares one baseline with the last run

  Shows what a baseline node treats as normal against the last measured run: the accepted median per metric, the observed value, the percentage change and any violation that stopped the run.

  Try: “Why did the row-count baseline fail on the last run?”

  Files · baseline_inspect

• Accepts the last run as the new normal

  Promotes the profile the last run measured to the accepted baseline, for a source that changed shape on purpose. It cannot invent numbers, the change is written to the audit log, and a workspace policy can refuse it.

  Try: “Accept the latest profile for the customers baseline.”

  Write · baseline_accept

• Forgets a baseline's accepted history

  Wipes a node's accepted history so the next run starts it over. The last observation is kept as evidence, the action is audited, and a workspace policy can refuse it.

  Try: “Reset the baseline on the events source.”

  Delete · baseline_clear

• Reads a pipeline's run log

  Returns the last lines of a pipeline's run log, which records events for each component.

  Try: “Show the last run log for the orders pipeline.”

  Files · read_run_logs

• Builds one deployable executable

  Packages a pipeline into a single self-contained executable for server deployment, written to the path you give it. It uses the duckle-runner binary and a secrets mode of environment variables or a passphrase.

  Try: “Build the orders pipeline into ./dist/orders for my server.”

  Runs code · build_pipeline

• Lists saved connections

  Lists the connections saved in a workspace. Secret fields are masked.

  Try: “Which database connections does my workspace have?”

  Files · list_connections

• Saves a database or service connection

  Writes a connection file into the workspace connections folder, and lists it in the workspace when one is present, so pipelines can refer to its fields.

  Try: “Create a Postgres connection called warehouse that reads its password from the environment.”

  Write · create_connection

Watch out for

• It can run any pipeline, including custom code

  run_pipeline runs pipelines headlessly, and the component catalog includes code.shell, code.python and code.javascript. Only run pipelines you have read, and keep the assistant in a project folder you trust. github.com

• Pipelines can write to your databases and files

  A pipeline run can write to its sinks, and run_pipeline, backfill and build_pipeline all start runs or write files. Point connections at test data first, and use run_pipeline with a target node to stop before the sink. github.com

• It can erase saved resume points

  backfill_clear and baseline_clear remove a node's saved state, so the next run starts over. For a Kafka source set to start from latest, clearing skips the backlog instead of replaying it. github.com

• Older Duckle versions had security fixes

  The project lists seven advisories, rated medium to critical, fixed in 0.7.3 and 0.7.4. They are about the serve web console, the container image, self-extracting artifacts and the archive extraction component. None of them names the MCP server. Use 0.7.4 or later. github.com

Security advisories