Skip to main content
Glama

pf-debate-mcp

A Public Forum debate partner for the AI you already use. Ask it things like:

  • "Make me a neg case on the Sep/Oct topic centered around a nuclear impact."

  • "Cut me a card that data centers raise residential electricity bills."

  • "Analyze this card and give me CX questions." / "Write blocks to their econ contention."

  • "They dropped our link turn. Write my summary."

  • "Scout Lexington AB on the caselist."

It knows PF jargon, speech times, tactics, weighing and the standard impact chains. It never makes up evidence.

Free web app (any age, no account)

https://debate.peshcompsci.org/app runs entirely in your browser:

  • plain-English search of about 266k cut cards

  • finding new papers and news

  • auto-suggested cuts with click-to-adjust highlighting and the verbatim check

  • a doc builder that exports to Word or Google Docs

  • an AI chat, powered by a free shared pool, your own free Puter account, or a one-click hand-off to your ChatGPT/Gemini

Related MCP server: guru-pk-mcp

Pick your app (no terminal needed)

App

What you get

Setup

Claude Desktop (Mac/Windows)

Everything: the card library, verbatim card cutting, caselist scouting, Word speech docs

One-click install

Claude.ai (web/mobile)

PF knowledge + card cutting with Claude's web search

Upload a skill

ChatGPT

PF knowledge + card cutting with ChatGPT's web search

Make a GPT or Project

Gemini

PF knowledge + card cutting with Gemini's web search

Make a Gem

Claude Code, Cursor, other MCP apps

Everything

Developer install

What "full tools" adds. The Claude Desktop and developer installs give your AI a local library of already-cut cards and a card cutter. The cutter checks every word against the real article and rejects anything that isn't verbatim, so fabricated evidence can't slip in. Web chat apps can't run local tools. There, the AI cuts cards from pages it opens, and you should check each card against its link before reading it.

All downloads are on the latest release page.

Claude Desktop (one click)

  1. Install Claude Desktop if you don't have it.

  2. Download pf-debate.mcpb.

  3. Double-click the file (or drag it into Claude Desktop → Settings → Extensions), then click Install.

    • The Tabroom email and password fields are optional. They are only for OpenCaselist scouting, and Claude Desktop stores them securely.

  4. In a new chat, say: "Build the quick card library." This takes a few minutes, runs in the background, and gets about 6k PF cards. Then say "build the full card library" for about 266k cards, including the most-read Policy/LD impact cards (nuke war, econ, heg). That takes a few hours and about 1.2 GB, and it resumes if interrupted.

  5. Start prepping. Speech docs are saved to Documents/pf-debate/.

Claude.ai (web)

  1. Download pf-debate-skill.zip.

  2. On claude.ai: Settings → Capabilities. Turn on Code execution (skills need it) and Web search, then under Skills click Upload skill and choose the zip.

  3. Ask any PF question. Claude loads the skill automatically.

Alternative: create a Project, paste instructions.md into the project instructions, and upload pf-debate-guide.md as project knowledge.

ChatGPT

Download instructions.md and pf-debate-guide.md. Then use either option:

  • Custom GPT (needs a paid plan to create; anyone can use a shared one):

    1. Explore GPTs → Create → Configure.

    2. Paste instructions.md into Instructions and upload pf-debate-guide.md under Knowledge.

    3. Turn on Web Search.

    4. Save it. Share the link with your team so they need no setup.

  • Project (any plan with Projects): New project → Instructions: paste instructions.md → Files: add pf-debate-guide.md.

Gemini

  1. Download the same two files (instructions.md, pf-debate-guide.md).

  2. gemini.google.com → Gems → New Gem. Paste instructions.md into Instructions and add pf-debate-guide.md under Knowledge. Save it.

  3. You can share the Gem with teammates.

Tools (Claude Desktop / MCP installs)

Tool

What it does

pf_guide

PF knowledge: glossary, format, tactics, impacts, evidence ethics, how to build cases/blocks/scouts

search_cards / get_card

Plain-English or keyword search (meaning + keywords) over the local library of about 173k cut cards from OpenCaselist (PF, LD, Policy, camp files; 2014–2022), with "popular" ranking by how many teams read a card

find_sources

New evidence: recent papers (author affiliations for quals, free PDFs) and current news. Free, no API keys

fetch_source

Any article or PDF, turned into clean paragraphs plus citation metadata

suggest_cut

Proposes the best passage and highlights for your claim (exact source text)

cut_card

Cuts a card; the text must match the source verbatim or it's rejected. Full NSDA citations.

export_doc

A Verbatim-compatible .docx (Pocket/Hat/Block/Tag, underline, highlight), or format="gdocs" for a Google-Docs-ready page. version="read" makes the read-ready copy to speak from (tags, short cites, highlighted words only, each section timed)

read_speech

Exact speech timing at the debater's pace: total vs. the speech's limit, time per contention/block, what to trim, and the read-aloud script

caselist_search / caselist_team / caselist_download

OpenCaselist with your Tabroom login: round reports, cites, open-source docs as cards

build_library / library_status

Build the card library in the background and check its progress

Skills: pf-debate, pf-cut-card, pf-analyze, pf-case, pf-blocks, pf-scout (in src/pf_debate_mcp/skills/).

Developer install

Requires uv.

Claude Code (plugin: server + skills):

/plugin marketplace add SujayGG/pf-debate-mcp
/plugin install pf-debate@pf-debate

Any MCP host (Cursor, Codex, Windsurf, and others):

{
  "mcpServers": {
    "pf-debate": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/SujayGG/pf-debate-mcp", "pf-debate-mcp"]
    }
  }
}

Terminal commands (optional; the build_library tool does the same from chat):

uvx --from git+https://github.com/SujayGG/pf-debate-mcp pf-debate-mcp build-library           # full library (a few hours, resumable)
uvx --from git+https://github.com/SujayGG/pf-debate-mcp pf-debate-mcp build-library --quick   # PF cards only, a few minutes
uvx --from git+https://github.com/SujayGG/pf-debate-mcp pf-debate-mcp login                   # OpenCaselist via Tabroom
  • The full build keeps every PF and OpenEv card, plus LD/Policy cards read by 5 or more teams (--min-reads).

  • The login stores only the session cookie, never your password. You can also set TABROOM_USERNAME and TABROOM_PASSWORD.

  • Data lives in ~/.pf-debate/ (PF_DEBATE_HOME). Exports go to ~/Documents/pf-debate/ (PF_DEBATE_EXPORTS).

Evidence ethics

With the tools, cards are never generated: cut_card stores only exact slices of fetched source text. In web chat apps, the instructions forbid writing card text from memory, but nothing checks it automatically. Verify every card against its source. You are responsible for following NSDA evidence rules.

Development

uv run pytest
uv run pf-debate-mcp                         # stdio server
uv run python scripts/package.py             # build the release downloads into dist/
npx @modelcontextprotocol/inspector uv run pf-debate-mcp

Releasing: bump the version in pyproject.toml, manifest.json and .claude-plugin/plugin.json. Then run package.py and gh release create vX.Y.Z dist/*.

Card data: the OpenCaselist dataset (MIT), from the debate community via openCaselist. Code: MIT.

Available Tools

16 tools
auto_cutA

Find evidence for a claim in one step: the best already-cut library cards PLUS new cards auto-cut from fresh papers/news (fetched, passage picked, highlighted, verified verbatim). Tags default to the claim: rewrite them to what each highlight proves, and check quals before using a card.

Inputs: claim (required), max_new. Returns: result

build_libraryA

Install the card library in the background (one time; resumable; progress via library_status). mode "download" (default): the prebuilt library (~170k cards incl. the most-read PF, LD, Policy and camp cards; a few minutes). "quick" / "full": build from the raw dataset instead (PF only in minutes, or everything in a few hours).

Inputs: mode. Returns: result

caselist_downloadA

Download an open-source .docx from OpenCaselist (the 'opensource' path from caselist_team or a search 'download_path') and import its cards as cN ids you can read, analyze, re-cut or export.

Inputs: path (required). Returns: result

caselist_entriesA

Scout a whole tournament field: match every entry to its caselist page in one call. entries: the CSV Tabroom exports from a tournament's Entries/Field page (Institution,Location,Entry,Code,...), or one line per team like "Plano West, Park & Jiang". Returns per team: disclosed or not, round count, recent open-source doc paths (for caselist_download) and cite titles. Then follow pf-scout "Prep out a tournament".

Inputs: entries (required), caselist. Returns: result

caselist_teamC

A team's caselist page: rounds (tournament, side, opponent, round report, open-source file path) and cites (their disclosed cards). Use for scouting and building blocks against their case.

Inputs: team (required), school (required), caselist.

cut_cardA

Cut an evidence card. The body is the exact source text from start_quote through end_quote (a few sentences to a few paragraphs; enough context that the author's meaning is clear). highlight = phrases read aloud; underline = wider phrases that give context (defaults to highlight). Every quote must be copied verbatim from the source (only whitespace, quote marks and dashes are normalized); anything else is rejected with a hint showing where the text stopped matching. source_id: sN from fetch_source, or a card id (cN, lib:N) to re-cut an existing card. paragraph: the [N] number from fetch_source where the card starts; required when start_quote appears more than once in the source.

Inputs: tag (required), url (required), date (required), quals (required), +9 more. Returns: result

export_docA

Write a speech doc / case / block file. items in order, each one of: {"pocket": "..."} {"hat": "..."} {"block": "..."} (Verbatim headings: Pocket > Hat > Block), {"tag": "..."} (analytic tag, no card), {"text": "..."} (speech prose), {"card": "c12" | "lib:345"} (optionally {"card": id, "tag": "new tag"} to retag a card in the export). format "docx": Verbatim-compatible Word file. format "gdocs": an .html file for Google Docs users (open it, select all, copy, paste into a Google Doc; or upload it to Google Drive and open with Google Docs). version "full" (default): complete cards, the doc you send in the email chain / keep as the file. version "read": read-ready copy to speak from (Rhetorify): each card cut to tag, short cite and only its highlighted words, each hat/block heading marked with its time at wpm, total time on top. Returns the file path.

Inputs: items (required), title (required), wpm, format, +2 more. Returns: result

fetch_sourceA

Fetch an article/PDF (or re-open a fetched source by its id) as clean numbered paragraphs + citation metadata. Long sources are paged: call again with start_paragraph to continue. Copy quotes for cut_card exactly from this text.

Inputs: url_or_id (required), start_paragraph. Returns: result

find_sourcesA

Find NEW evidence to cut: recent scholarly papers (with author affiliations for quals and free PDF links) and current news articles. kinds: ["papers", "news"] (default both). Free and keyless; if a service is busy it is listed under "skipped". Then fetch_source a result's url and cut it.

Inputs: query (required), kinds, limit.

get_cardA

Card by id (lib:N or cN). view="read" (default, compact): tag, cite, and only the underlined/highlighted text (==highlighted== is read aloud, underlined is context, ... marks skipped text). view="full": the whole body, needed to judge context, find indicts in unhighlighted text, or recut.

Inputs: card_id (required), view. Returns: result

library_statusB

Whether the local card library is built, how many cards per event, years covered, and build progress.

Inputs: no named parameters.

pf_guideA

PF debate knowledge. Read "pf-debate" first. Task guides: pf-case, pf-cut-card, pf-analyze, pf-blocks, pf-blockfile (team block files), pf-scout, pf-practice. References: glossary, format, tactics, impacts, evidence-ethics.

Inputs: name. Returns: result

read_speechA

Time a speech exactly and return its read-ready script (what is actually said aloud). Use it to check a case fits before exporting, and to see what to trim. Counting is exact, never estimate word counts yourself. items: the same list as export_doc (headings, {"tag"}, {"text"}, {"card": id}). A card counts its tag, short cite and highlighted words only. For a case pasted as plain text, pass the words the debater actually reads as {"text": ...} items (highlighting in pasted text can't be seen). speech: constructive (4:00), rebuttal (4:00), summary (3:00), final focus (2:00). wpm: the debater's pace. Lay about 160, fast about 200, circuit about 230+. Ask their pace if it matters.

Inputs: items (required), wpm, speech. Returns: result

search_cardsA

Search already-cut debate cards. Plain English works ("cards saying data centers raise power bills"), and so does debate shorthand ("econ decline war", "heg", "prolif").

mode: "hybrid" (default: meaning + keywords) or "keyword" (exact words only).
scope: "library" = OpenCaselist corpus (PF, LD, Policy, camp OpenEv files; 2014-2022);
       "mine" = cards the user cut or imported (local installs).
sort: "relevance" or "popular" (most-read by teams first; a strong quality signal for impact cards).
side: "A"/"N" (aff/neg). event: pf | ld | cx | openev. Try a couple of phrasings for important searches.

Inputs: query (required), mode, side, sort, +4 more. Returns: result

suggest_cutA

Suggest a card for claim from a fetched source (sN / s_… from fetch_source): the best-matching passage and highlight phrases, all exact source text. Review and adjust, then pass the returned start_quote, end_quote, paragraph, highlight and underline to cut_card with a tag and cite.

Inputs: claim (required), source_id (required). Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.8/5.0

    Scored across 16 tools

    Disambiguation5/5

    Each tool targets a distinct resource or action: guide, library install/status, card search/retrieval, source fetching/cutting, caselist scouting/download/search, document export, and speech timing. Even the similar cut tools (suggest_cut vs. auto_cut) are clearly differentiated by whether they operate on a provided source or perform an end-to-end evidence search.

    Naming Consistency4/5

    Most tools use a verb_noun pattern (build_library, export_doc, search_cards, get_card, cut_card, fetch_source, suggest_cut, find_sources, read_speech), but a few deviate with noun-first caselist_* names (caselist_team, caselist_entries, caselist_download, caselist_search), noun_noun names (pf_guide, library_status), and the compound auto_cut. This is readable and groups related tools, but the mixed conventions are a minor inconsistency.

    Tool Count4/5

    At 16 tools, the set is slightly above the 3-15 'well-scoped' range but not excessive for a debate prep server covering research, card cutting, caselist scouting, document export, and speech timing. Each tool serves a distinct workflow step and none feel redundant, so the count is justifiable.

    Completeness4/5

    The tool set spans the full debate prep pipeline: finding/fetching sources, cutting and searching cards, scouting/downloading opponent cases, building exports, and timing speeches. Minor gaps include no card update/delete operations and no direct import from arbitrary pasted text, but these can be worked around through re-cutting or using read_speech.

    Maintenance

    ActivitySlowing
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers