Before you start
- Connect a client first. The Quickstart takes about five minutes, or pick your client from Connect a client.
- You never give the agent a password or key. For a new connection, the agent sends you a one-time link and you type the secret in the dashboard yourself (see Creating Connections). Existing connections you just refer to by name.
- Use names, not ids. The agent looks up ids itself, so “the
prod-snowflakeconnection” works as well asconnection_id: 42.
Level 1: Ask questions about your account
Start here. These prompts each use one or two read tools and return in seconds, so they are a good way to get a feel for what the agent can see. Get an overview (Read)Prompt
Prompt
Prompt
Example response
Example response
Based on a real job log, with names changed:Run 4815162 of salesforce_opportunities_to_snowflake failed at 06:07 UTC, about 20 seconds in. The dashboard only shows “Package failed to execute”, so here is what the log says.What broke: the Snowflake destination couldn’t write to its table. Snowflake returned:
Table 'OPPORTUNITY' already exists, but current role has no privileges on it.The table exists, but it’s owned by a different role than the one your Snowflake connection uses. That usually happens when someone recreates the table by hand or from another tool. The Salesforce side read fine, and the package itself is valid.How to fix it: ask your Snowflake admin to give the connection’s role access to the table, for example:GRANT ALL PRIVILEGES ON TABLE ANALYTICS.SALESFORCE.OPPORTUNITY TO ROLE INTEGRATEIO_ROLE;Then re-run the package. No changes to the package are needed. Want me to retry the run once the grant is in place?Prompt
Prompt
Prompt
Prompt
Level 2: Document and audit your pipelines
The agent can read the full definition of every package, so it can describe your pipelines in plain language and cross-reference them. This is one of the most common first uses: getting a clear map of what you have before changing anything. Map everything that touches one system (Read)Prompt
Example response
Example response
From a demo account, lightly trimmed:I found 6 packages that touch HubSpot. None of them is on a schedule.
Things to look at: there are two copies each of “SFTP to HubSpot” and “HubSpot Contacts to Salesforce”, and one “SFTP to HubSpot” has no components at all, so it’s safe to archive. Nothing here has completed a run, so if HubSpot data is landing somewhere today, it isn’t coming from these packages. Want me to compare the duplicates?
Prompt
Prompt
Prompt
Prompt
Prompt
Level 3: Troubleshoot and fix
Here the agent moves from reading to changing things. It diagnoses a problem with read tools, then proposes an edit. Edits land in your draft of the package first. Nothing runs differently until the draft is saved as a new version. Diagnose, then propose a fix (Read, then Write)Prompt
Prompt
Prompt
Prompt
Prompt
Prompt
Level 4: Build new pipelines
The agent can build a complete source-to-destination package from a description. For a straight “load X into Y”, it does this in a single call and validates the result. For anything with filters, joins, or several sources, it builds component by component. Set up a connection (Write)Prompt
Prompt
Prompt
Prompt
Prompt
Prompt
Prompt
Prompt
Prompt
Level 5: Change many pipelines at once
This is where the MCP server saves the most time. Work that would take days of clicking through the designer becomes a prompt plus a review. One customer used it to regenerate a large set of packages and estimated it cut about three months of manual work down to three days. These prompts work best in a client that can run long tasks, such as Claude Code, Claude Desktop, or Cursor. Apply one change everywhere (Write)Prompt
Prompt
Prompt
Prompt
Prompt
Teams that use the MCP server heavily tend to do the routine work (cloning, troubleshooting, bulk edits) through the agent and then open the dashboard for the last mile: a final review of the canvas, or a tricky component. Knowing the platform well makes it much easier to check what the agent did.
Writing better prompts
- Be specific about names. Name the package, connection, table, or workspace. The agent will search if you don’t, but it may pick the wrong match.
- Ask for a plan first on any write. “Show me what you would change and wait for my go-ahead” is the single most useful phrase for write prompts.
- Ask it to validate. Mention
validate_packageor “validate before saving” and the agent will catch wiring and expression errors before anything runs. - Say what output you want. A table, a doc, a list of ids, or a yes/no answer.
- Never paste credentials into the chat. No tool returns credentials, and new connections are finished through a one-time dashboard link where you enter the secret yourself. Point the agent at existing connections by name.
Working safely with write prompts
- Keep write tools on approval in your client. Claude, Cursor, and most other clients ask before each tool call unless you choose “Always allow”. Allowing the read tools saves clicks, but leave the write tools set to ask, so you approve every change before it happens.
- Edits go to a draft first. Component, edge, and variable edits change your draft only. Runs and schedules keep using the saved version until the draft is saved with
save_package_version. Ask the agent to show you the draft before it saves. - Use dry runs. Many write tools, including
clone_package,update_package_components,save_package_version,create_schedule,update_schedule, andrun_package, acceptdry_run. Ask “do a dry run first” to see what would change without changing it. - Everything is versioned. Saved changes appear in the package version history and the account audit log, and
revert_packagecan roll a package back. Unsaved drafts that are discarded cannot be recovered. - Clusters cost money.
create_clusterprovisions billed infrastructure. If you don’t want the agent to create one, say so in the prompt. - Control access with permissions. The agent can only do what the signed-in user can do. Give people who only need to look a read-only role.