Analyze local data with Capy Agent

Capy can explore workspace tables, write and run read-only SQL, and save editable charts, dashboards and notebooks in your RawQL workspace.

Choose context and sharing

Import orders.csv and customers.csv through Databases → Add local sources. Open Capy Agent in the workbench navigation, then open Capy settings. All current and new workspace objects are available by default. Use the edit icon under Workspace access to select objects in the explorer. Checkboxes sit on the right. Save, leave the settings tab or switch windows to save your choices. Excluding a folder excludes its current and future descendants. Source tables remain read only.

Capy sends your question, authorized names, types and business definitions directly to Google AI Studio for Gemini 3.5 Flash-Lite, or through OpenRouter for other models. File content sharing and result sharing are enabled by default. You can turn them off in Capy settings. They let Capy read SQL, notebook text, chart sources and raw or aggregated result pages after provider consent. These values can contain sensitive information. A sample never proves a source total.

Access, sharing choices and business definitions are saved with each workspace. The provider consent checkbox is in Capy settings and stays valid in the current workspace, including new conversations and page reloads. Switching workspaces clears it. Access changes stop the current task. Completed actions and previously transmitted content remain.

Reference files and folders with @

Type @ anywhere in your question to search workspace SQL files, notebooks, dashboards and folders. Select an entry with the mouse, or use the arrow keys and Enter. You can add several mentions, continue typing after each one, or insert a mention earlier in the question. Backspace removes a selected mention; Escape closes the picker.

Choose Read for reference material or Read & edit for files Capy may change. A folder includes its nested files when you send the question. An explicit file mention overrides its folder's access; the last explicit mention of that file wins. Full paths distinguish files with the same name. New documents are available by default on the next request, except under excluded folders. Mentions never override exclusions.

Update @Reports using @Reference.sql as a read-only source.

Pick Read & edit for Reports and Read for Reference.sql. Review the file count and access summary, then review the sharing choices in Capy settings and allow the provider consent. Mentions do not grant sharing consent. Deleted or excluded references block sending. Metadata exceeding the byte limit produces an explicit error without silently omitting documents.

Use the Marketplace

Capy can search, describe and import Marketplace datasets autonomously. It follows the same source availability and import rules as the human interface. Importable entries create or reuse a local table and open an executed preview. SQL starters open a worksheet without running it. Excluded tables cannot be imported through Capy.

Build your first analysis

In local, preview and production, the model button in the composer opens thinking controls and the model list. Gemma 4 31B is the default. Gemini 3.5 Flash-Lite uses the Gemini API directly, with Auto, Light, Standard and Deep reasoning. Google free-tier quotas or API charges apply to the configured key. The available controls depend on the model. Switching model starts a fresh model context; completed files stay in your workspace. Free providers may log prompts and outputs, so use non-confidential context. The model badges describe provider capabilities; Capy currently sends text context.

Calculate paid revenue by customer segment. Create a SQL file, execute it and save a bar chart. Explain the result.

Capy describes the authorized tables, creates a SQL file and executes a guarded read query in DuckDB on your device. Action cards show what actually succeeded. Open the SQL file or saved chart directly from its card.

Compare the SQL and values before relying on an answer. If result sharing is off, inspect the local result yourself. Capy can still create the chart, but the model receives no result values.

Dashboards and files

A chart request saves the visualization in the SQL file's Chart tab. It creates no dashboard. For a combo chart, specify the bar and line measures and their axes, for example revenue as bars on the left axis and order count as a line on the right. Click a file name in Capy's answer or action list to open it. Answers support Markdown lists, tables and code. Use Copy below an answer to copy its Markdown. Jump to latest returns to the newest reply after you scroll up.

Build a paid-sales dashboard with a revenue KPI, a bar chart by segment, a results table and a note explaining how revenue is calculated. Save the SQL files too.

Each visualization uses the full SQL source and keeps its execution provenance. SQL that already aggregates data uses result mode, so the chart does not aggregate those values a second time. Dashboard refreshes run locally and make no AI call.

Create a notebook named Paid sales report with a Markdown explanation and SQL cells for the paid total and revenue by segment. Do not run the notebook cells.

Supporting SQL files appear in the _capy_artefacts folder; the final notebook appears at the workspace root. A notebook can contain SQL and Markdown; Capy does not execute Python or JavaScript. Ask for visualizations in a notebook to copy completed charts, their SQL and local results into editable cells. Export files or download a workspace snapshot through the usual RawQL controls when you want a portable copy.

Edit, stop and recover

Activity opens while Capy works and shows elapsed time, comments and actions. Click its time summary to collapse or expand it. When the task ends, activity collapses automatically and the final answer stays visible. Reopen activity to inspect diagnostics, file links and undo controls. Successful actions reset the consecutive repair counter. The final recap links verified workspace files. If notebook creation fails, Capy reports that failure instead of confirming an absent file. Follow-up notebook requests can reuse saved SQL files; Capy re-executes them locally to refresh their results.

Include a SQL file and ask for a specific change, such as adding a status filter. Capy checks the file ID and revision before replacing its SQL. If you edit the same file while Capy works, the action reports stale and preserves your edit. Changing tabs does not redirect an action to a different file. Undo edit reverses a successful Capy SQL edit while that revision is still current. Include notebooks and dashboards to revise their cells, titles, notes and chart bindings. Charts shared with another dashboard stay unchanged there. Dashboard changes use the existing definition history.

Stop interrupts the request and prevents subsequent actions. Completed files and charts remain available. An in-flight model request may still be charged. If DuckDB cannot stop promptly, the engine resets; reload a saved snapshot to recover SQL-created data if needed. Closing and reopening the drawer keeps the in-memory conversation. Reloading the tab clears it. Follow-up questions use an encrypted in-memory history for up to thirty minutes. Starting a new conversation clears that history.

Limits and privacy

Tasks have no fixed limit on model steps or total actions. Each task has a three-minute deadline, and SQL queries have a twenty-second deadline. The server reserves a conservative spending allowance before each model call, with a default task ceiling of $0.05 and separate daily account and environment limits. The displayed reserved amount is an upper allowance, not a provider invoice.

Capy accepts one read-only SELECT or WITH statement at a time. It cannot modify source data, import arbitrary remote sources, execute arbitrary code or browse your disk. Unknown functions, excluded sources and replaced import views are rejected. A truncated result is marked explicitly. Totals must come from SQL over the source, not from counting a preview.

RawQL keeps conversation text and result handles in browser memory. The server forwards the approved context and returns an encrypted continuation; Redis stores only run control metadata and budget counters. It does not store prompts, SQL, results or encrypted conversations. Free-model providers may log prompts and outputs after your explicit consent, in every environment. Read the Capy privacy details before sharing private context.

If Capy is unavailable, your ordinary SQL workbench stays usable. Retry after checking your connection, wait for an active task to finish, or ask the instance administrator to check Capy's server settings. A stopped or interrupted task must be restarted explicitly; completed actions are not replayed automatically.