Skip to main content
This walkthrough builds a company list, checks the results, limits the rows before optional paid work, and exports a CSV. You need a Landbase account and an authorized MCP client. The quoted prompts are instructions to your assistant. Dataset IDs, account IDs, and run IDs come from your actual tool results. For more copy/paste patterns, see the MCP prompt cookbook.

​
1. Choose the account

Ask:
Use Landbase to list my accounts. If there is more than one, ask me which account to use before doing anything else.
The assistant calls list_accounts. Choose an account and keep its account_id for every account-scoped call in the walkthrough. Pass it explicitly on searches, dataset reads, transformations, qualification quotes and confirmations, workflow polls, publishing, and both download calls. The server does not remember the selection between calls.

​
2. Build the list

Ask:
Find B2B SaaS companies in Texas with 50–200 employees. Preview the results before any paid follow-up work, and keep the session and run IDs so we can resume if the connection drops.
The assistant calls search in default mode. If the result has a running handle, it continues with get_search_result using that handle. If the result includes a workflow_run_id, it waits for that workflow with get_workflow_run before reading the dataset. A completed search is not enough on its own: check for a dataset_id or a workflow that produces one. If neither exists, the assistant should explain answer rather than claiming it built a list. For interrupted work, see Continue a search or workflow.

​
3. Inspect and limit the rows

Ask:
Show the dataset’s row count, its available field names, and a preview of 10 rows. Then create a separate dataset containing at most 50 rows for the next steps. Do not enrich or qualify it yet.
The assistant uses get_dataset, get_dataset_fields, and preview_dataset. It should use the returned field names when discussing columns, rather than assuming every dataset has the same schema. Searches can return more rows than requested. To cap the working set, the assistant calls transform_dataset on the search dataset with:
This selects up to 50 rows; it does not rank them. Ask for an explicit sort using available fields if you need a ranked selection. Wait for the transform workflow to reach SUCCEEDED, then use its output_dataset_id as the working dataset. Keep the original search dataset separate so later steps cannot accidentally process the full list.

​
4. Optionally qualify the working set

If you want a more specific fit check, ask:
For the limited dataset, check whether each company sells software to sales teams. Use the relevant name and website fields from its schema. Show me the credit quote or row count and wait for my approval before starting.
The assistant calls qualify_dataset without confirm, using the working dataset ID and actual context columns. A needs_confirmation result can include an estimate. If there is no exact estimate, it should say so and show the row count rather than inventing a price. If you approve, it repeats the same call with confirm: true, then polls the returned workflow. On SUCCEEDED, switch to that workflow’s output_dataset_id and preview the result. If you decline, skip qualification and keep the limited dataset from step 3. See Limits and credits for approval rules, including the separate research-search confirmation flow.

​
5. Publish and download a CSV

Ask:
Export the current working dataset as CSV. Tell me how many rows the download will include and ask before spending download credits.
The assistant calls publish_dataset with format: "csv", polls the publish workflow, and uses its output_dataset_id for get_download_url. All three calls keep the selected account_id, as does each follow-up poll. The first download call omits confirm and returns a confirmation request with the row count. After you approve the cost, the assistant repeats the call with confirm: true and returns the temporary download URL. Links expire in about 30 minutes, so download the file promptly.

​
Check the result

The assistant should give you the final dataset ID, row count, and download link, and state whether qualification ran. A failed or cancelled workflow does not establish that a finished output is available. Keep the IDs if you need to inspect dataset lineage or resume the conversation later.