A PDF tool can process a document on your computer while an AI client sends extracted text elsewhere. Check both parts of the workflow before describing it as private.
A recent prerelease with several interfaces
PaperPilot is a document toolkit written in Rust. Its v0.6.0 release was published on October 24, 2026, and is marked as a prerelease. The package list includes command-line, MCP, REST, desktop, and WebAssembly options.
The attraction for an agent workflow is direct access to document operations. MCP, or Model Context Protocol, lets a compatible client call those tools. The README describes current operations such as merging PDFs, changing page order, compression, and text extraction.
These are documented capabilities, not results from our own installation test. Check the specific release package and operation you need before building a workflow around it.
Follow the document through the whole request
The maintainer describes the engine as running locally or in a browser tab. That describes where the document operation runs. The connected AI client has a separate role: it receives instructions and may receive tool results.
Suppose you ask an agent to merge two files. A useful result may be an output path and a success message. Ask it to summarize the merged content, and extracted text may become part of the model request. Actual behavior depends on the client, selected tool, and provider configuration.
Before using sensitive material, inspect the tool request and returned fields with a synthetic PDF. Check which text appears in the conversation or request logs available to you. A local executable alone cannot establish the data handling of every connected service.
Check one operation with a synthetic document
I would start with a disposable PDF containing a made-up name, a table, and a distinctive test sentence. This proposed evaluation separates file correctness from how the AI connection handles content.
- Record the original page count and expected page order.
- Run one supported operation and open the resulting PDF.
- Check the page order, visible text, and any required formatting.
- Inspect the MCP response for extracted text, paths, and other returned fields.
- Repeat the task through the intended client and review the content visible to that client.
For extraction, a fluent summary cannot confirm that every number survived. Compare the extracted table with the original document. Our AI21 LIFT-Bench coverage explains the related problem of checking requested content independently from presentation.
Keep roadmap features out of the buying decision
The current README puts PDF-to-Word, Excel, and PowerPoint conversion in a later roadmap stage. It also lists signing, redaction, and forms as future work. Distinguish those plans from currently listed operations such as validating or removing signatures.
The license file identifies an Apache 2.0 Community Edition and a separately licensed Enterprise Edition. Enterprise access controls and audit features should not be assumed to come with the community package.
For reports built from extracted material, keep an evidence trail back to the original page. Our OmniDR-Bench article examines that citation problem in research reports using video sources.
Inspect the release before connecting it
- PaperPilot v0.6.0 prerelease and packages
- Current operations and roadmap
- Community and Enterprise license boundaries
Does your document workflow need file transformations, extracted text for an agent, or both?
