Tablestakes

tablestakes

An MCP server that gives LLMs clean, surgical access to tables trapped in messy HTML.

The Problem

Tools like GitBook, Notion exports, and CMS platforms collapse tables into single-line HTML when syncing to Markdown files. The result looks like this in your editor:

<table><thead><tr><th width="520.11">Requirement</th><th width="122.07">Priority</th><th>Priority 1-2-3</th></tr></thead><tbody><tr><td><strong>1.1</strong> Agent sees only their Salesforce-assigned cases <strong>in the currently selected organization</strong> (case is "assigned" when SF <code>Case.OwnerId</code> matches the agent's linked SF user ID)...</td><td>Must</td><td>1</td></tr></tbody></table>

This is unreadable for humans and unreliable for LLMs. Models struggle to parse collapsed HTML tables, frequently hallucinate cell boundaries, and cannot edit them without corrupting the structure.

tablestakes fixes this. It sits between the LLM and the file, converting tables to clean pipe format on read and writing back in the original format on save — preserving GitBook compatibility, HTML attributes, and inline formatting.

What the LLM Sees

Discovery — scan a 26-table document in one call:

26 tables

T0 pipe 5r 3c v:485f65f7b470 [Cross-Domain Dependencies] A:Integration | B:Source | C:Requirements

T2 gitbook 18r 3c v:77a9495fd328 [Case List] A:Requirement | B:Priority | C:Priority 1-2-3

T7 gitbook 3r 4c v:d9a9a45a370f [Attachments] A:Requirement | B:Priority | C:Dependency | D:Priority 1-2-3

Read — collapsed HTML becomes a clean pipe table:

v:d9a9a45a370f gitbook 3r 4c [Attachments]

A:Requirement | B:Priority | C:Dependency | D:Priority 1-2-3

Write — surgical cell edit, version-checked:

v:5749c94ffb1f

14 characters. The file is updated, GitBook HTML format preserved, width attributes intact.

Token Efficiency

Baseline: Claude Code's built-in Read + Edit tools operating on the same file. Measured on a synthetic 18-row, 4-column table with realistic requirement-style content (bold IDs, inline emphasis, mixed-length cells).

Where the savings come from:

• Read (HTML): collapsed HTML tags ( , , , , width="...") are pure overhead. Pipe tables carry the same information without markup. The Read tool also adds cat -n line-number prefixes.
• Read (GFM): modest savings from stripping line-number prefixes and surrounding document context. The table content itself is already clean.
• Write: the Edit tool requires old_string (enough context to be unique in the file) + new_string (the modified version), both generated as output tokens. For GFM, old_string is the entire row line (~190 chars). tablestakes needs only {"row": 0, "column": "B", "value": "Should"} (~18 tokens).
• Discovery: without tablestakes, the LLM reads the entire file to find tables. list_tables returns a compact index — metadata + 1 preview row per table.

• Compact pipe tables with no column padding. Per the ImprovingAgents benchmark, GFM pipe tables achieve the best token-to-accuracy ratio: 1.24x CSV cost at 51.9% QA accuracy, beating JSON (2.08x, 52.3%) and YAML (1.88x, 54.7%).

Experiment details

Tokenizer: tiktoken cl100k_base (GPT-4). Claude uses a different tokenizer, but relative comparisons hold. The benchmark script (script.py) constructs tables programmatically and generates tablestakes output using the actual converter code — no hardcoded strings.

Read baseline: simulate_read_tool() wraps file content in cat -n format (line-number prefix per line), matching what Claude Code's Read tool returns. The full file (document text + table) enters the LLM context.

Write baseline: for each cell edit, the script computes the minimum unique old_string by expanding leftward from the target until the substring is unique in the file. new_string is the same context with the cell value replaced. This is a best-case scenario for the Edit tool — a human might include more context than the minimum.

list_tables baseline: 26 copies of an 18-row GitBook HTML table in a markdown document. Naive = Read the full file (~28k tokens). tablestakes = list_tables output with preview_rows=0..3:

Reproduce: uv run --with tiktoken python scripts/script.py

Quick Start

Claude Code:

claude mcp add tablestakes -- uvx tablestakes

Codex CLI:

codex mcp add tablestakes -- uvx tablestakes

Gemini CLI:

gemini mcp add tablestakes -- uvx tablestakes

Or install from PyPI directly: pip install tablestakes

Other clients (Cursor, Windsurf, Claude Desktop)

Add the following JSON to your client's MCP config file:

{

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

Tools

Discovery & Read

Cell, Row & Column Operations

All write tools require a version hash from read_table — optimistic concurrency that prevents stale overwrites without locks.

Supported Table Formats

While GitBook is the primary motivation, tablestakes works with any Markdown document containing HTML tables — CMS exports, Notion dumps, wiki migrations, or hand-written HTML in .md files.

Column Addressing

Columns can be referenced by:

• Letter: "A", "B", "AA" (bijective base-26, like Excel)
• Name: "Priority" (must be unique)
• Composite: "B:Priority" (for disambiguation)
• Index: "0", "1" (0-based)

Development

make init      # First-time setup: venv + deps + pre-commit hooks

make check # All checks: format + lint + typecheck + test make test # Run tests only make test-cov # Tests with coverage report

Author

Built by Dr. Oliver Borchers, Executive AI and Automation Partner and Head of AI Engineering at IU Group.

License

Apache-2.0

---

mcp-name: io.github.oborchers/tablestakes