# Search PDFs and spreadsheets from Claude Desktop without uploading the whole collection

> Use Central Brain to index PDFs and XLSX files locally and retrieve source context through Claude Desktop MCP without uploading the whole source collection.

- URL: https://neuroaigent.ai/central-brain-answers/search-pdfs-spreadsheets-claude-desktop/
- Updated: 2026-10-18

**Central Brain prepares and indexes supported PDFs and spreadsheets locally, then lets a connected Claude Desktop session retrieve relevant project context through MCP.** You can avoid manually uploading the whole source collection to a cloud file library. The context returned to Claude still reaches the connected AI provider; local indexing does not mean that no document text leaves your computer.

Central Brain by NeuroAIgent is built for recurring questions over project files: which contract contains a clause, which report supports an assumption, or where a workbook mentions a cost category. It combines semantic and exact text search with source references where available, rather than requiring the entire corpus to be attached to every conversation.

Start with [Central Brain plans](https://neuroaigent.ai/central-brain/pricing/) and the installer currently available in [getting started](https://neuroaigent.ai/central-brain/getting-started/). Supported preparation formats include PDF, DOCX, TXT, Markdown, CSV and XLSX. OCR is optional and build-dependent. Central Brain helps locate spreadsheet evidence through prepared content; validate formula calculations and workbook interpretation in your spreadsheet workflow.

## 1. Build a small, representative local project

Begin with a source you understand and inspect the prepared text alongside it. This gives you a clear view of extraction quality, scan readability and workbook layout before expanding the searchable collection.

1. Choose approved files for one project. Include a text PDF and a spreadsheet whose expected contents you can verify.
2. Select the source folder and separate prepared-output storage in Central Brain. Keep both in the project backup plan.
3. Run preparation and indexing. Review errors and inspect the prepared output for the files that matter.
4. Search locally for an exact phrase in the PDF and a distinctive label from the workbook. Open the sources to confirm the match.
5. Add a scanned PDF only if the installed build includes the OCR capability you need. Check the recovered text before relying on it.

## 2. Connect Claude Desktop to that project

Open the project MCP setup in Central Brain and generate the Claude Desktop entry. Use the actual command and root supplied by your app, then apply it through the client configuration supported by your installed version and restart as directed. Do not substitute a remote connector URL for a local stdio launcher.

Use the project-specific connection generated by Central Brain. Anthropic documents local desktop servers separately from remote connectors, with tool inspection and diagnostic logs for local connections. Your organization settings determine which local integrations are enabled; follow those settings when connecting the project.

Ask Claude to call central_brain_ping and show the project root. Confirm it is the intended project and that the tools include semantic_search, search_text, read_file and list_files. That makes the evidence boundary visible before the first question.

## 3. Ask for evidence, then open the source

A useful request is: Search this project for the renewal notice requirement, quote the relevant text, and include the source path and page where available. For a workbook, ask for the retrieved labels and file reference before asking the assistant to interpret the business meaning. Central Brain returns path, prepared-file path, page and chunk where available; format-specific location detail varies.

| Need | Tool strategy | What to verify |
| --- | --- | --- |
| Find a concept without knowing its wording | semantic_search over the selected project. | Does the retrieved passage actually answer the question? |
| Locate a clause, identifier or known label | search_text for the exact string. | Is this the intended version and source file? |
| Read the surrounding content | read_file within the permitted project. | Does surrounding context qualify the apparent answer? |
| Check whether a document was included | list_files and preparation review. | Was the expected file prepared and indexed? |

## 4. Treat spreadsheets as structured source evidence

A search hit is not a recalculated workbook. Open the original spreadsheet to verify which sheet, period, units and assumptions apply. Check whether the prepared text captures the relevant labels and values. Do not assume formulas, hidden sheets, formatting, merged cells or external workbook links have been interpreted as a human spreadsheet reviewer would interpret them.

For a question requiring a fresh financial model or formula evaluation, use an appropriate spreadsheet workflow and validate its results separately. Central Brain can help locate relevant source context; it does not establish that a retrieved number is current, complete or suitable for a decision.

## 5. Keep the retrieval boundary deliberate

In the default Central Brain workflow, source files, prepared output, vectors and memory stay on storage you choose. Claude receives the context its tools retrieve to answer the question. A broad read request can expose more context than a narrow search hit, so select approved source folders and ask focused questions.

Check optional cloud connections, synced folders, backups and the provider settings that apply to the Claude account. Refresh changed files before answering from an updated report. When a search misses, inspect preparation, OCR availability, the project root and the query wording before concluding the information is absent.

## Test extraction before trusting a document answer

Use a small source set with answers you already know: a PDF paragraph, a workbook label and value, and a revised document. Check each returned source against the original. A search hit is a retrieval result, not a guarantee that a table, formula or scanned page was interpreted correctly.

For a spreadsheet question, name the workbook and the relevant period or sheet in your request, then compare the retrieved value with the original workbook. For PDFs, inspect page references where available. Scans depend on the OCR capability and build you use. Review preparation errors rather than assuming a file was indexed because its folder was selected.

- Confirm a known phrase is retrievable before trying an open-ended summary.
- After updating a source, refresh preparation and indexing and verify the revised answer.
- Keep confidential test material out of a connected cloud assistant until its provider and workspace settings are approved.

## Common questions

### How does Central Brain reduce repeated file uploads?

It prepares and indexes the source collection locally by default, so supported clients can retrieve project context as needed. Returned search context and file content still reach the connected assistant when retrieved.

### Can I search scanned PDFs?

OCR is optional and depends on the Central Brain build. Confirm it is present and inspect extraction quality on representative scans.

### Does Central Brain recalculate XLSX formulas?

This workflow makes no recalculation claim. Verify values, formulas, units and sheet context in the original workbook or a suitable spreadsheet tool.

### Why is a page number missing?

Source metadata is returned where available. Location fields vary by format and extraction; use the path and surrounding source content to verify the result.

## Related answers

- [One memory for ChatGPT, Claude and Copilot](https://neuroaigent.ai/central-brain-answers/one-memory-chatgpt-claude-copilot/): Use Central Brain as one local project memory for supported ChatGPT desktop, Claude and GitHub Copilot clients. Set up and verify cross-client recall.
- [Local RAG desktop app](https://neuroaigent.ai/central-brain-answers/local-rag-desktop-app/): Evaluate Central Brain as a local RAG desktop app for Windows and Apple silicon Mac, without confusing retrieval with the AI that writes the answer.
- [Private business-file AI search](https://neuroaigent.ai/central-brain-answers/private-business-files-ai-search/): How Central Brain lets AI search selected business files locally by default, what leaves your computer, and a checklist for a careful private-file pilot.

## Next steps

- [Get Central Brain file search](https://neuroaigent.ai/central-brain/pricing/): Prepare one approved collection and connect your supported client.
- [Prepare your first project](https://neuroaigent.ai/central-brain/getting-started/)

## Sources and review notes

- [Anthropic: Local MCP servers on Claude Desktop](https://support.claude.com/en/articles/10949351-getting-started-with-local-mcp-servers-on-claude-desktop) — Local tool inspection and troubleshooting; checked October 17, 2026.
- [Anthropic: Desktop and web connectors](https://support.claude.com/en/articles/11725091-when-to-use-desktop-and-web-connectors) — Local versus remote connection boundary; checked October 17, 2026.
- [Central Brain security and data boundaries](https://neuroaigent.ai/central-brain/security/)

## Explore the complete Central Brain workflow

Bring local file preparation, source-traceable search and readable project memory together. Explore [Central Brain features](https://neuroaigent.ai/central-brain/), follow the setup guide and choose a plan for your project.