claude mcp add) so your agent can read it with that app’s own MCP tools. Dippr never sees that app’s
credential.
Where it works
Small complete inline results
If the app already returned a complete, unpaginated result of at most 200 small rows, your agent can useworkspace_db_write_sql to copy it into a new table in your personal database.
With the load-data skill installed, the agent keeps source and pulled_at columns. In a chat
client without that skill, ask your agent to include those columns explicitly. The skill checks
your personal database with dip db status --json without a name, then omits the tool’s optional database argument to select your personal database, checks the returned
database, and escapes quotes in app text before making SQL literals. It checks that the table’s
count(*) equals the number of rows received before reporting success. The rows pass through the model on this path.
A table copied this way prevents a later saved-file import from publishing into your personal
database until you agree to remove it, because that importer accepts only earlier imports’ tables.
The tool gives no numeric byte or per-row size cutoff; 200 rows does not make large text or blob
fields suitable. Previews, truncated results, unknown completeness and results with a next page
do not qualify. The agent must not write inline rows to a local file or divide a large pull into
small copies.
For large or paginated pulls, use the saved-file steps below. If their results arrive inline,
restart Claude Code with MAX_MCP_OUTPUT_TOKENS=500 claude and try again. That setting cannot
guarantee saving or override an app tool’s own size limit; each response still needs checking.
Import saved pages from an app
1
Connect the app
If the app is not connected yet, add it once. In Claude Desktop, go to Settings → Connectors. In Claude
Code, your agent can add it with
claude mcp add, and you finish signing in with /mcp. You sign in to
the app, not to Dippr.2
Ask for the import
Tell your agent what you want, for example: “Import the AI project’s issues created from July 1 to
September 30 into a jira_issues table in my workspace database.” Use fixed dates rather than “the last
90 days”, so the search does not move while the import runs.
3
The agent learns the page shape
The agent asks the app for a small first page and has
dip db import inspect describe its structure,
which shows field names but no values. From that it writes a recipe naming the columns to keep, the
next-page cursor and the app’s total. dip db import learn checks the recipe against two real pages.4
The agent collects every page
After each page,
dip db import page prints the next request to send to the app. The agent sends
exactly that request, then hands the new page to dip, until the app reports no more pages. The pages
stay private on your computer.5
dip publishes the table
dip db import finish publishes only the recipe’s typed columns to a new table in your personal
database. It checks first that pagination finished, that no row appears twice and that the row count
matches the app’s total. If the app reports no total, the agent shows you the row count and asks
before publishing.What stays on your computer
- Saved pages. Claude Code’s saved result files and
dip’s private copies stay on your machine. Only the recipe’s typed columns are published, with an import control table. - Values in reports.
dip’s reports carry structure, counts and the next request, never row values. Field names can themselves be data, for example an email address used as a key, so keepinspectreports private. - The app’s credential. Your app connection belongs to your AI client. Dippr never receives it.
Limits
- An import is a snapshot. It does not refresh on its own. To update it, ask your agent to import again into a new table.
- Saved-file destination.
finishpublishes to your own personal database, and only while it holds nothing but earlier imports’ tables. The agent checks this before it asks the app for any page. - Sharing covers the whole database.
dip db sharegives a colleague read access to everything in your personal database: every imported table, and any table you import later, not only this one. Before you share, check that you mean to share every table in it. - Previews are refused. If an app’s tool declares its own result size limit, Claude Code may show a
preview of the rows instead of only saving them.
diprefuses such a result, and the agent stops, because those rows were already shown to the model. - Jira’s total is not checked. Jira returns the number of matching issues with the first page only, so
diphas no total to compare the rows with on later pages. A Jira import reports its count as unavailable, and the agent shows you the row count and asks before publishing. - The connected app’s limits apply. Page sizes, rate limits and what the app’s tools return are set by the app.
- Your access in the app decides what you get. The import contains what your own app account can see. Once the rows are in your workspace database, the database’s sharing rules apply, not the app’s.
Related
Load data
Load files into a workspace database, and the import checks in detail.
Workspace databases
What a workspace database is and who can read it.