Rules for agents using Landbase MCP
- Call
list_accountswhen account context is ambiguous. - Preserve the selected
account_idon every account-scoped call. - Use fields returned by
get_dataset_fieldsinstead of assuming a schema. - Poll returned handles and run IDs before retrying a request.
- Preview and quote paid work before asking for confirmation.
- Never claim a dataset or workflow completed without checking its terminal status.
Choose an account safely
Prompt
Expected behavior
The assistant callslist_accounts, asks the user to choose when necessary, and passes the selected account_id on later account-scoped calls. list_accounts itself does not require an account ID.
Cost and safety
This is an inspection call. It should not spend credits or change data.Search companies without spending credits
Prompt
Expected behavior
The assistant usessearch in default mode, keeps the returned session and search handle, and polls get_search_result if the search is still running. It should inspect the resulting dataset before proposing paid work.
Cost and safety
Searching and previewing should remain read-only. The assistant must not start enrichment or another paid workflow from this prompt.Search people using returned fields
Prompt
Expected behavior
The assistant callsget_dataset_fields before constructing the follow-up request, uses the returned field names, and previews the result before contact enrichment.
Cost and safety
Do not assume that every dataset has the same schema. Contact details and enrichment may spend credits and require a separate approval.Match one company or person
Prompt
Expected behavior
The assistant usesmatch_company or match_person, prefers a website, LinkedIn URL, email, or company context over an ambiguous name, and explains whether the match is verified or only plausible.
Cost and safety
Matching is an inspection step. Do not silently chain it into enrichment or contact discovery.Limit a dataset before paid work
Prompt
Expected behavior
The assistant callsget_dataset, get_dataset_fields, and transform_dataset with a read-only SELECT, polls the transform workflow, and uses its output_dataset_id for later work. It keeps the original dataset separate.
Cost and safety
Transforming a dataset helps control the rows processed by later paid workflows. It does not rank rows unless the SQL explicitly sorts them.Enrich after approval
Prompt
Expected behavior
The assistant callsenrich_dataset without confirm, shows the returned estimate or row count, repeats the same call with confirm: true only after approval, and polls the returned workflow. It uses the workflow’s output_dataset_id after success.
Cost and safety
Enrichment spends credits. Preserve the selectedaccount_id on the quote, confirmation, and every workflow poll.
Qualify or research with explicit context
Prompt
Expected behavior
The assistant usesqualify_dataset or research_dataset, supplies the relevant returned context columns, follows the quote/confirmation flow, and polls the resulting workflow.
Cost and safety
Qualification and research can spend credits. Research-modesearch uses a confirmation turn in the original session rather than a confirm boolean.
Recover a running operation
Prompt
Expected behavior
Useget_search_result for a search handle, get_agent_run or wait_for_agent_run for an agent run, and get_workflow_run for a workflow run. A wait expiration means the remote work may still be running.
Cost and safety
A timeout or lost response does not prove that work stopped. Never repeat a paid start, upload, publish, or download call until the original operation has been recovered or its outcome is known.Publish and download after review
Prompt
Expected behavior
The assistant publishes when required, polls the publish workflow, uses itsoutput_dataset_id, requests a download URL without confirmation first, and repeats with confirm: true only after approval.