Skip to main content
Glama
README.md
# pdf-mcp

[![PyPI version](https://img.shields.io/pypi/v/pdf-mcp)](https://pypi.org/project/pdf-mcp/)
[![Python 3.11 to 3.14](https://img.shields.io/badge/python-3.11--3.14-blue.svg)](https://www.python.org/downloads/)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
[![GitHub Issues](https://img.shields.io/github/issues/jztan/pdf-mcp)](https://github.com/jztan/pdf-mcp/issues)
[![CI](https://github.com/jztan/pdf-mcp/actions/workflows/ci.yml/badge.svg)](https://github.com/jztan/pdf-mcp/actions/workflows/ci.yml)
[![codecov](https://codecov.io/gh/jztan/pdf-mcp/graph/badge.svg)](https://codecov.io/gh/jztan/pdf-mcp)
[![Downloads](https://pepy.tech/badge/pdf-mcp)](https://pepy.tech/project/pdf-mcp)

**Agentic RAG over your PDFs, with the evidence to check every answer.**

The agent decides when to search; pdf-mcp does the retrieval and hands back excerpts, each with the page it came from. When the extracted text might be wrong, the agent can render that page and check it against what's actually printed. Hidden text, including injected instructions, is flagged, not silently passed on. It's an [MCP](https://modelcontextprotocol.io/) server that lets Claude Code and other AI agents search one PDF or a whole folder, read only the pages that matter, and pull out tables, images, and scanned text, even from multi-column and Japanese layouts.

**mcp-name: io.github.jztan/pdf-mcp**

## Try it in your browser

**[See what your AI agent sees →](https://pdf-mcp.jztan.com/)**

Drop in any PDF, or a whole folder of them, and watch an agent triage the corpus, search across every document at once, and read only the pages that matter, using a fraction of the tokens. 100% client-side, no install required.

<p align="center">
  <a href="https://pdf-mcp.jztan.com/"><img src="https://raw.githubusercontent.com/jztan/pdf-mcp/develop/docs/images/demo.gif" alt="pdf-mcp browser demo: an AI agent warms a 6-PDF corpus, triages it, searches across all six documents, and reads only the matching page, with 97.3% of the corpus never entering the context window" width="760"></a>
</p>

## Why pdf-mcp?

| | Without pdf-mcp | With pdf-mcp |
|---|---|---|
| Large PDFs | Context overflow | Read only the pages you need |
| Hidden text, including injected instructions | Silently ingested | Flagged as untrusted, nothing stripped |
| Checking extracted text | Trust it blindly | Render the page and compare against what's printed |
| Finding content | Read page after page until it turns up | Hybrid search: BM25 keyword + semantic |
| Folders of PDFs | No index: open documents until one matches | Warm, triage, and search a whole folder |
| Warming a big folder | Minutes of CPU embedding | Length-sorted small-batch CPU encode; optional CUDA embedding, one to two orders of magnitude faster on an NVIDIA card |
| Tables and charts | Lost in raw text | Structured rows, and `(x, y)` data from vector charts |
| Multi-column and vertical layouts | Columns interleaved | Correct reading order, including Japanese tategaki |
| Scanned PDFs | No text at all | OCR via Tesseract, parallel across pages |
| Repeated access | Re-parse every time | SQLite cache that survives restarts |

### From a researcher using pdf-mcp

> "This MCP is genuinely useful, and useful in a way that is unique because it holds through real scientific validation."
>
> [(Markson, P 2026)](https://go.dataacorns.com/the-record), on re-checking Fisher's 1936 iris table

## Installation

### Claude Desktop: nothing to install first

1. **[Download pdf-mcp.mcpb](https://github.com/jztan/pdf-mcp/releases/latest/download/pdf-mcp.mcpb)**.
2. In Claude Desktop, open **Settings > Extensions**, drag the file onto
   that page, and click **Install**.
3. Ask Claude about a PDF by its location, for example "Use pdf-mcp to
   summarize C:\Users\me\Downloads\report.pdf", or about a whole folder.

The first start downloads pdf-mcp's components (about 250 MB) and can take a
few minutes; later starts take seconds. OCR for scanned pages is included:
the first scanned page downloads an English-only Tesseract (about 14 MB).
Needs Windows 10 or later, or macOS 13 or later (14 on Apple Silicon), and
works in Claude Desktop's Chat. Updating, uninstalling and other details are
in [docs/clients.md](docs/clients.md).

### Claude Code and other MCP clients

Needs Python 3.11 to 3.14. On 3.15 the installer falls back to a very old
release, because a dependency has no 3.15 build yet. Install the `pdf-mcp`
command with [uv](https://docs.astral.sh/uv/) or
[pipx](https://pipx.pypa.io/):

```bash
uv tool install pdf-mcp     # or: pipx install pdf-mcp
```

`pip install pdf-mcp` works inside a virtual environment; Homebrew's Python
and recent Debian and Ubuntu refuse a system-wide pip install.

Then add it to your client. For Claude Code:

```bash
claude mcp add pdf-mcp -- pdf-mcp
```

For VS Code, Cursor, Codex CLI, Kiro or any other MCP client, see
**[docs/clients.md](docs/clients.md)**. Then ask your agent to read a PDF.

Search, the corpus tools, tables and multi-column and CJK reading order work
out of the box. For non-English documents, see
[language support](docs/configuration.md#language-support). OCR on scanned
pages also needs Tesseract:

```bash
brew install tesseract                             # macOS
sudo apt install tesseract-ocr                     # Ubuntu/Debian
winget install -e --id UB-Mannheim.TesseractOCR    # Windows
```

Optional: CUDA embedding on an NVIDIA card warms large folders one to two
orders of magnitude faster; see
[docs/configuration.md](docs/configuration.md#gpu-embedding-nvidia-cuda).

### From Python

pdf-mcp's tools are also plain Python functions, so you can import them and
hand a PDF to the Anthropic SDK without running a server. Two runnable
scripts, for a question and for a whole document: **[examples/](examples/)**.

Why this exists, and what broke along the way: [Claude's 100-page PDF limit and how I got around it](https://blog.jztan.com/how-i-built-pdf-mcp-solving-claude-large-pdf-limitations/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=quickstart-how-i-built-pdf-mcp-solving-claude-large-pdf-limitations)

## Tools

13 specialized tools rather than one monolithic one. Typical pattern:
`pdf_info` to plan, `pdf_search` to locate (its paragraph excerpts often
answer the question outright), `pdf_read_pages` when you need more. For a
folder, `pdf_corpus_overview` to triage, then `pdf_corpus_search`.

| Tool | What it does |
|------|--------------|
| `pdf_info` | Page count, metadata, TOC summary, scanned-page detection. **Call first.** |
| `pdf_search` | Hybrid search (keyword + semantic) by page or section; hits carry an excerpt, source coordinates, the enclosing outline section and a table's lead-in sentence |
| `pdf_read_pages` | Read specific pages or ranges, with OCR on demand, tables, and embedded images |
| `pdf_read_all` | Read a whole document in one call, byte-capped |
| `pdf_get_toc` | Full table of contents for documents with many bookmarks |
| `pdf_render_pages` | Render pages as PNG for vision models: diagrams, handwriting, scans |
| `pdf_extract_chart` | Chart data as exact `(x, y)` tables, read from plot geometry |
| `pdf_corpus_warm` | Warm a folder of PDFs into the cache within a time budget; OCRs scanned pages between calls (`ocr`) |
| `pdf_corpus_overview` | Per-document triage cards for a folder |
| `pdf_corpus_search` | Search across a folder; hits carry document and page provenance |
| `pdf_cache_stats` | Per-document cache breakdown and total size |
| `pdf_cache_clear` | Clear expired or all cache entries |
| `server_info` | Which optional features and config are active |

Text returned by any of these is untrusted content extracted from a PDF.
`pdf_info(content_trust=True)` reports hidden text a human reader cannot
see, and the read tools flag it per page.

Example prompts:

```
"Read the PDF at /path/to/document.pdf"
"Which pages discuss supply chain risks?"
"Find sections about the training process"
"Show me what page 5 looks like"
"OCR pages 3-5 of the scanned PDF"
```

Full reference, every parameter and response shape:
**[docs/tool-reference.md](docs/tool-reference.md)**. Embedding model
selection: **[docs/embedding-models.md](docs/embedding-models.md)**.

## Example Workflow

For a large document (e.g., a 200-page annual report):

```
User: "Summarize the risk factors in this annual report"

Agent workflow:
1. pdf_info("report.pdf")
   → 200 pages, TOC shows "Risk Factors" on page 89

2. pdf_search("report.pdf", "risk factors")
   → Matches with structural paragraph excerpts: each excerpt
     is the bullet, paragraph, or heading that matched, not a
     fixed-width window. Often enough to answer directly.

3. If excerpts are sufficient → synthesize answer

4. If more context needed:
   pdf_read_pages("report.pdf", "89-95")
   → Full page text for deeper reading
```

## Remote / HTTP transport

STDIO is the default and is what every example above uses. `pdf-mcp-http`
serves the same tools over HTTP, for clients that cannot spawn a process
(the Anthropic API MCP connector, claude.ai custom connectors) and for a
warm corpus shared by several clients.

```bash
export PDF_MCP_AUTH_TOKEN="$(openssl rand -hex 32)"
pdf-mcp-http
```

Paths resolve on the server, so an HTTP agent reads what is already there:
files under an allow-listed root, or a URL the server fetches. It cannot
hand over a file from its own machine. It is single-tenant and fails
closed: with no auth token and no `[paths]` allow list, the process exits
rather than serving an open endpoint.

Docker images are published to GHCR for amd64 and arm64, with everything
baked in, so every tool works on the first request:

```bash
./deploy.sh              # token, image, start, health-check
cp your.pdf documents/   # this folder is the server's /data/pdfs
```

Read **[docs/remote-access.md](docs/remote-access.md)** for the trust
boundary and threat model before deploying, and
**[docs/configuration.md](docs/configuration.md#http-transport-setup)**
for setup, client config, and token rotation.

## Configuration

pdf-mcp works out of the box. To restrict which paths and URL hosts the
server may touch, tune cache and worker settings, or add your own
content-trust phrases, see **[docs/configuration.md](docs/configuration.md)**.

## Roadmap

See [ROADMAP.md](docs/ROADMAP.md) for planned features and release history.

## Contributing

Contributions are welcome. See **[docs/contributing.md](docs/contributing.md)** for setup, checks, the coherence eval harness, and quality-loop guidelines.

## Contributors

Thank you to everyone who has helped improve this project through code, reviews, testing, and feature requests:

<!-- contributors:start -->
[@Summer907](https://github.com/Summer907) · [@ebbsanchez](https://github.com/ebbsanchez) · [@VooDisss](https://github.com/VooDisss) · [@DerDennisOP](https://github.com/DerDennisOP) · [@deepdmk](https://github.com/deepdmk) · [@TheSOV](https://github.com/TheSOV) · [@janLo](https://github.com/janLo) · [@bbulkow](https://github.com/bbulkow) · [@HillsCloud](https://github.com/HillsCloud) · [@phinnphace](https://github.com/phinnphace)
<!-- contributors:end -->

<a href="https://github.com/jztan/pdf-mcp/graphs/contributors">
  <img src="https://contrib.rocks/image?repo=jztan/pdf-mcp" alt="Contributors" />
</a>

Per-release contributor credits are listed in the [Changelog](./CHANGELOG.md).

## Security

Found a vulnerability? See [SECURITY.md](SECURITY.md) for the threat model, reporting channel, and expected response timeline. Please do not open a public GitHub issue for unpatched security reports.

## License

MIT. See [LICENSE](LICENSE).

## Links

- [pdf-mcp on PyPI](https://pypi.org/project/pdf-mcp/)
- [pdf-mcp on GitHub](https://github.com/jztan/pdf-mcp)

## Blog posts

**The story behind the releases.** Building pdf-mcp keeps surprising me: benchmarks that go the wrong way, formats that break everything, features I had to remove. I write about that thinking in [The Dispatch](https://blog.jztan.com/newsletter/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-newsletter). Come along if that's your kind of thing.

Background, benchmarks, and design notes from building pdf-mcp:

**Getting started**

- [How I Built pdf-mcp](https://blog.jztan.com/how-i-built-pdf-mcp-solving-claude-large-pdf-limitations/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-how-i-built-pdf-mcp-solving-claude-large-pdf-limitations): The problem with large PDFs in AI agents and a working solution
- [How to Send PDFs Over 100 Pages to Claude's API](https://blog.jztan.com/llm-api-pdf-page-limits/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-llm-api-pdf-page-limits): Measuring the real page and token ceilings on the Claude API, and the two ways around them: search the PDF on disk for questions, window the text and render only the picture pages for whole-document tasks
- [How Claude Code Actually Reads PDFs](https://blog.jztan.com/how-claude-code-actually-reads-pdfs-lessons-from-building-an-mcp-server/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-how-claude-code-actually-reads-pdfs-lessons-from-building-an-mcp-server): How AI agents use pdf-mcp tools to read and navigate PDF documents
- [How AI Agents Should Read PDFs: 5 Patterns That Survived Production](https://blog.jztan.com/ai-agent-pdf-reading-patterns/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-ai-agent-pdf-reading-patterns): Five production-tested patterns for how agents should navigate PDFs at scale

**Corpus & multi-document search**

- [A Knowledge Base Is Just a Folder](https://blog.jztan.com/ai-agent-pdf-knowledge-base/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-ai-agent-pdf-knowledge-base): Turning a folder of PDFs into an agent knowledge base with the corpus tools, no ingestion pipeline or vector store
- [Cross-Document Retrieval for AI Agents Without a Vector Database](https://blog.jztan.com/rag-without-vector-database/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-rag-without-vector-database): Why BM25 scores don't merge across per-document indexes but ranks do, and how two-stage RRF puts a gold document in the top 3 on 84.8% of 184 graded queries over a 100-PDF corpus
- [How Amazon Bedrock Helped Me Make My RAG Better](https://blog.jztan.com/how-bedrock-improved-my-rag/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-how-bedrock-improved-my-rag): Benchmarking pdf_corpus_search against Bedrock Knowledge Bases at an equal 2,000-token budget for four cents surfaced two bugs eight months of self-testing missed: an excerpt picker discarding answers from pages it had already retrieved, and one embedding per page hiding short answers
- [Does a PDF Search Tool Still Matter for AI Agents?](https://blog.jztan.com/pdf-mcp-vs-builtin-pdf-reading-benchmark/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-pdf-mcp-vs-builtin-pdf-reading-benchmark): Claude Code with and without pdf-mcp on 53 multi-paper questions over 120 papers: answer quality tied, but the search tool read 7.5 pages per question instead of 187 and answered in 25 seconds instead of 92

**Search & retrieval**

- [Semantic vs Keyword Search for AI Agents](https://blog.jztan.com/semantic-vs-keyword-search-ai-agents/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-semantic-vs-keyword-search-ai-agents): Benchmarks and a dual-search routing pattern: FTS5 for exact identifiers, embeddings for natural language
- [Hybrid Search vs Query Routing for AI Agents](https://blog.jztan.com/hybrid-search-vs-query-routing-ai-agents/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-hybrid-search-vs-query-routing-ai-agents): Why pdf-mcp uses hybrid RRF instead of query routing: benchmarks showing RRF wins across query types
- [Section Chunking vs Page Chunking for AI Agents](https://blog.jztan.com/section-chunking-vs-page-chunking-ai-agents/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-section-chunking-vs-page-chunking-ai-agents): Why section-aware search delivers full section content in one call while page-mode costs 2–6 extra tool calls per query
- [Section-Level RAG: Why BM25 Beat Hybrid Search in My Benchmark](https://blog.jztan.com/bm25-vs-hybrid-search-section-rag/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-bm25-vs-hybrid-search-section-rag): Why pdf-mcp's section-grain search is BM25-only: hybrid RRF caused a 33% lexical regression at section grain, so granularity decides the search technique
- [How One Search Change Eliminated an Entire Agent Step](https://blog.jztan.com/how-paragraph-excerpts-changed-agent-behavior/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-how-paragraph-excerpts-changed-agent-behavior): Switching pdf_search from fixed-width snippets to paragraph excerpts turned it from a pivot tool into a terminal tool: 97% vs 80% answer containment across a 30-query benchmark

**Engineering & security**

- [MCP Server Security: 8 Vulnerabilities](https://blog.jztan.com/mcp-server-security-8-vulnerabilities/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-mcp-server-security-8-vulnerabilities): What we found when we audited an MCP server for security holes
- [Your LLM Is Free QA for Your MCP Server](https://blog.jztan.com/llm-free-qa-mcp-server/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-llm-free-qa-mcp-server): Four Payload UX bugs in pdf-mcp that schema tests missed but Claude Desktop surfaced during real use
- [Why Multi-Column PDFs Scramble Reading Order in RAG](https://blog.jztan.com/multi-column-pdf-reading-order/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-multi-column-pdf-reading-order): Fixing two-column extraction (0.564 → 0.816 fidelity), the title-page author-grid regression it caused, and the aggregate metric that stayed blind to both
- [How I Fixed Vertical Japanese PDF Extraction](https://blog.jztan.com/vertical-japanese-pdf-reading-order/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-vertical-japanese-pdf-reading-order): Tategaki pages extract scrambled because reading order is geometric, not stored; rebuilding it from glyph positions (columns right to left, characters top to bottom), with no OCR and no new dependency
- [My AI Agent Stopped Using My Own PDF Tool: The 72 DPI Bug](https://blog.jztan.com/ai-agent-phone-scan-legibility/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-ai-agent-phone-scan-legibility): Phone scans silently downsampled to 72 DPI to fit the response byte budget made handwriting unreadable, so the agent routed around pdf_render_pages; the fix spends JPEG quality before resolution and makes the response say when a render is illegible and which clips to retry
- [Extract Chart Data From PDFs: An MCP Tool That Won't Guess](https://blog.jztan.com/extract-chart-data-from-pdfs/?utm_source=github&utm_medium=referral&utm_campaign=pdf-mcp&utm_content=list-extract-chart-data-from-pdfs): Reading exact chart values straight from a PDF's vector geometry, then handing the ambiguous judgment calls back to the LLM instead of guessing

TDQS

A3.6/5.0

Scored across 5 tools

Disambiguation5/5

Each tool serves a distinct purpose: cache clearing, corpus overview, server info, page rendering, and chart extraction. There is no functional overlap, so an agent can clearly differentiate between them.

Naming Consistency3/5

Most tools use the 'pdf_' prefix, but 'server_info' breaks this pattern. Also, there is inconsistency in verb/noun ordering: 'pdf_cache_clear' uses noun_verb, while 'pdf_render_pages' and 'pdf_extract_chart' use verb_noun, and 'pdf_corpus_overview' uses noun_noun.

Tool Count5/5

With 5 tools, the count is well-scoped for a PDF processing server covering essential operations: cache management, overview, diagnostics, rendering, and chart extraction. Neither too few nor too many.

Completeness3/5

While the set covers core viewing and analysis features, it lacks a dedicated text extraction tool. The overview tool provides text content, but a direct 'extract text' function is missing, which is a notable gap for a PDF server.

Maintenance

ActivityActive
ResponsivenessUsually handled