TableShot

The only MCP server for PDF table extraction. Give any AI assistant the ability to read tables from PDFs -- no other tool does this.

PyPI License: MIT Tests Python 3.10+

Camelot, Tabula, and Table Transformer are Python libraries -- they require a developer to write code. TableShot is an MCP server: Claude Desktop, Cursor, and Windsurf can use it directly with zero code.

~33MB install. No model downloads. No API keys. Results in <100ms.

The Problem

Ask any AI assistant to read a table from a PDF. It can't -- you get word soup:

Sales Report Q1 2024 Product Price Quantity Total Widget A $10.00 100
$1,000.00 Widget B $25.50 50 $1,275.00 Widget C $5.99 200 $1,198.00

TableShot gives you this:

Product	Price	Quantity	Total
Widget A	$10.00	100	$1,000.00
Widget B	$25.50	50	$1,275.00
Widget C	$5.99	200	$1,198.00
Widget D	$149.00	10	$1,490.00

Quick Start

Claude Desktop / Cursor / Windsurf

Add to your MCP config:

{
  "mcpServers": {
    "tableshot": {
      "command": "uvx",
      "args": ["tableshot"]
    }
  }
}

Then just ask: "Extract the tables from /path/to/report.pdf"

pip

pip install tableshot

Run as a standalone MCP server:

tableshot              # stdio transport (for MCP clients)
python -m tableshot    # same thing

Tools

Tool	What it does
extract_tables	Extract all tables as Markdown, CSV, JSON, or HTML
list_tables	Quick scan -- preview tables before extracting

extract_tables

source: str           # File path or URL to a PDF (or image with [ml] extra)
pages: str = "all"    # "all", "1", "1-3", "1,3,5"
format: str = "markdown"  # "markdown", "csv", "json", "html"

list_tables

source: str           # File path or URL to a PDF
pages: str = "all"    # "all", "1", "1-3", "1,3,5"

Returns table count, dimensions, headers, and a preview row for each table found.

Examples

Financial report (bordered table)

Input: BlackRock-style quarterly earnings PDF

Output (markdown):

|                                      | Q3 2023    | Q3 2022    | 9M 2023    | 9M 2022    |
| ------------------------------------ | ---------- | ---------- | ---------- | ---------- |
| Total revenue                        | $4,522     | $4,311     | $13,228    | $13,536    |
| Total expense                        | 2,885      | 2,785      | 8,538      | 8,578      |
| Operating income                     | $1,637     | $1,526     | $4,690     | $4,958     |
| Operating margin                     | 36.2%      | 35.4%      | 35.5%      | 36.6%      |

Extracted in 25ms.

Multi-table document

Input: PDF with employee directory + budget summary on the same page

Output: Both tables extracted separately with correct headers:

Table 1: 3 rows x 3 cols (Name, Department, Email)
Table 2: 4 rows x 2 cols (Category, Amount)

Wide table (8 columns, landscape)

| ID  | Name  | Q1  | Q2  | Q3  | Q4  | Total | Status |
| --- | ----- | --- | --- | --- | --- | ----- | ------ |
| 1   | Alpha | 100 | 150 | 200 | 250 | 700   | Active |
| 2   | Beta  | 90  | 110 | 130 | 170 | 500   | Active |
| 3   | Gamma | 0   | 0   | 50  | 80  | 130   | New    |

All 4 output formats (Markdown, CSV, JSON, HTML) available for every extraction.

Benchmarks

Tested on 10 PDFs covering bordered tables, multi-table pages, multi-page documents, special characters, wide tables, and real financial statements.

Metric	Result
Bordered table accuracy	8/8 exact match
Speed (bordered tables)	4-25ms per extraction
Speed (3-page financial PDF)	182ms
Output format validity	36/36 pass (9 PDFs x 4 formats)

Test Data

Generated fixtures — click Source to see the input PDF, Output to see what TableShot extracts:

Fixture	Description	Source	Output	Speed
simple_bordered	4-column sales report (Product, Price, Quantity, Total)	PDF	Extracted	10ms
multi_table	Two tables on one page: employee directory + budget summary	PDF	Extracted	10ms
single_row	Minimal table — header + one data row	PDF	Extracted	4ms
multi_page	One table per page across 2 pages	PDF	Extracted	9ms
empty_page	Page 1 text only; page 2 has a table	PDF	Extracted	6ms
special_chars	Cells with $, :, ", &, <>	PDF	Extracted	6ms
wide_table	8-column landscape table (Q1–Q4, Total, Status)	PDF	Extracted	11ms

Real-world PDFs (not included in repo due to size/licensing):

PDF	Description	Tables	Speed
BlackRock mock	Generated mock of a BlackRock quarterly earnings statement (5 columns)	1 table, 11 rows	25ms
Sample Financial Statements	3-page financial statement with complex visual formatting (155KB)	3 tables, 75 rows	182ms
NHM table	Large 56-page document with 55 tables (25MB)	55 tables, 2321 rows	5.8s

Full machine-readable results in benchmarks/results.json. Detailed before/after comparisons in benchmarks/results.md.

vs Other Tools

	TableShot	Camelot	Tabula-py	Table Transformer
Install	~33MB, nothing else	Needs Ghostscript	Needs Java (100-300MB)	Needs PyTorch (700MB-5GB)
Speed	~10ms/table	>20s worst case	Variable (JVM startup)	2-5s/page
Bordered tables	Excellent	Excellent	Good	Excellent
Borderless	Good (text fallback)	Poor	Better detection	Best
MCP support	Native	None	None	None
Maintained	Active	~5 years stale	Active	Active

Competitor data from Adhikari & Agarwal 2024, OpenNews 2024 review, and published GitHub metrics. Full results in benchmarks/results.md.

Need Scanned PDFs or Images?

The base install handles native PDFs with text layers (90%+ of real-world use cases). For scanned documents and images:

pip install tableshot[ml]     # Table Transformer for image-based tables
pip install tableshot[ocr]    # OCR for scanned documents (ONNX, no PyTorch)
pip install tableshot[all]    # Everything

With [ml] installed, TableShot automatically detects whether a PDF has a text layer:

• Text layer present -- uses pdfplumber (fast, ~10ms)
• Scanned / no text layer -- uses Table Transformer for detection, pdfplumber for text extraction
• Image files (PNG, JPEG) -- uses Table Transformer + OCR (requires [ocr])

You can also force the ML backend: extract_tables("/path/to/scan.pdf", backend="ml")

How It Works

PDF/Image ──> Smart Router ──> Table Detection ──> Cell Extraction ──> Formatted Output
                  |                                                        |
                  |  PDF with text layer:                                  |  Markdown
                  |    pdfplumber (lines → text fallback)                  |  CSV
                  |                                                        |  JSON
                  |  Scanned PDF / Image (with [ml]):                      |  HTML
                  |    Table Transformer → pdfplumber text / OCR           |

• pdfplumber handles PDF parsing and table detection (MIT)
• pypdfium2 renders PDF pages to images for ML backend (Apache-2.0)
• Table Transformer (optional [ml]) detects tables in images (MIT)
• MCP SDK exposes tools to AI assistants via stdio transport (MIT)

Total base install: ~33MB. No model downloads. No GPU required.

Known Limitations

All rule-based PDF table extractors (including Camelot and Tabula) share these limits:

• Financial statements with visual formatting -- amounts positioned by whitespace rather than cell borders can fragment across columns
• Scanned PDFs / images -- no OCR in base install (use tableshot[ml] or tableshot[ocr])
• Scientific papers with equations -- inline math breaks table boundary detection
• Complex borderless tables -- ambiguous column alignment can cause misdetection

We're honest about these. For edge cases, tableshot[ml] adds Table Transformer support.

Contributing

git clone https://github.com/Bespoke34/tableshot.git
cd tableshot
pip install -e ".[dev]"
pip install fpdf2                 # for generating test fixtures
python tests/generate_fixtures.py # create test PDFs
pytest -m "not slow"              # run 160 tests (skip ML tests)
pytest                            # run all 167 tests (needs [ml] extra)
ruff check src/ tests/            # lint

• 95% test coverage, all tests must pass
• Ruff clean, no lint warnings
• MIT license -- all dependencies must be MIT/Apache-2.0/BSD compatible

License

MIT