Ipsissima-MCP

Ask your assistant for an argument map of a paper, and get one that cites the text it is a reading of. Ask for a map of a debate — the arguments for and against a conclusion, belonging to no single text — and there is nothing to cite: the assistant writes the map directly, and the checks still hold it to parsing and structure.

Ipsissima-MCP is an MCP server. It does the mechanical half of building a reconstruction — getting a document into structured Markdown with its paragraphs and page numbers intact, and checking a finished reconstruction against its sources word for word. The judgement half, turning a paper into an argument, is done by the assistant you are already talking to; this server hands it the instructions and then checks its work.

The result opens in Ipsissima, which shows the map, the passage each claim was drawn from, and how far each claim stands from the source's own words.

Install

The easy way: the Claude Desktop bundle

Download the .mcpb bundle from the releases page and double-click it. Claude Desktop installs it, provisions Python and the dependencies itself, and there is nothing to configure — no terminal, no virtual environment, no path typed into a JSON file.

You still need Node on the machine. The Argdown parser is JavaScript and is carried inside the bundle, so there is nothing to install for it, but something has to run it. Nothing else is required: Claude ships a Node runtime for its own extensions but does not expose it, and Python it does not ship at all — which is why the bundle asks the host to provide one rather than carrying its own.

If the checker later reports that it cannot find Node, install it and restart Claude Desktop. It looks in the usual places as well as on the PATH, because an application launched from the Dock does not get the PATH your terminal has.

The terminal's way: from PyPI

Already in a terminal, with uv on the machine? The whole install is one line:

claude mcp add ipsissima -- uvx ipsissima-mcp

uvx makes an environment, installs the released ipsissima-mcp into it and throws it away afterwards, so there is no virtual environment to make, remember, or type the path of. Any MCP client can run the same command — uvx ipsissima-mcp is the server. Node still has to be on the machine, as above, and pandoc is still worth having. The source checkout below remains the way to work on Ipsissima-MCP itself.

The developer's way: from source

You need Python 3.11+, Node (any current version — it reads the Argdown), and ideally pandoc.

git clone <this repository> && cd ipsissima
python3 -m venv .venv && .venv/bin/pip install -e ipsissima-mcp

A virtual environment is not fussiness: a Homebrew or system Python will refuse pip install outright (PEP 668).

With uv instead, which is the same thing and faster:

uv venv && uv pip install -e ipsissima-mcp

A uv-made .venv has no pip in it. That is uv working as designed, not a broken environment — but it means .venv/bin/pip install … fails with no such file or directory, and the obvious reading of that message is that the virtual environment did not get made. Use uv pip install for a uv venv, or python3 -m venv if you want pip inside it. The entry points below are installed either way; ls .venv/bin/ipsissima-mcp is the quick way to tell whether the package is there.

Node is needed, npm install is not. The checker uses the Argdown parser as ground truth for whether a file is valid, and a copy of it is bundled into this package as a single file that node runs — so there is nothing to install for it and nothing to keep in step. It used to reach into app/node_modules instead, which meant the server could only run from a source checkout and could not run on Windows at all, npm writing no shim of that name there. If node is missing the checker says so and names the download page, rather than reporting a missing CLI.

rapidocr installs with it and is not optional. Without an OCR backend, a scanned paper converts silently to its cover page and nothing reports it — measured at 345 words of a 1,220-word article, no error, a perfectly well-formed Markdown file. See eval/CONVERTER-FINDINGS.md.

Your assistant needs to be able to read and write files

This is the requirement most likely to catch you out, and it is not obvious from the tool list. Ipsissima-MCP does not carry the whole job. Look at where the work actually happens:

extract_text writes source/ and returns a report of what it wrote rather than the text: the reply carries a short head of each file and no more, because a book-length source would otherwise arrive in the conversation twice over. The head is there to show that the conversion worked — headings intact, page markers present — and is deliberately far too little to reconstruct from. And no tool here writes the reconstruction — that is the model's judgement, and the model saves it with its own file-writing tool.

So an assistant that cannot touch the filesystem gets as far as a folder full of Markdown it cannot read, and has nowhere to put the map it then cannot write.

And it has to be the same filesystem, which is easier to lose than it sounds. The table asks whether each party can reach a disk. What it does not ask is whether they reach the same one. An assistant that talks to this server across a device bridge — a hosted session reaching your Mac, and any arrangement like it — runs the server on your machine and the model in a sandbox somewhere else. Every tool then works perfectly and writes to a disk the model cannot open: extract_text reports a folder of Markdown that genuinely exists, at paths that resolve to nothing where the model is standing.

Measured once, on a 321-page book. The model could not open the paths, and went back to reading the original PDF itself — producing a competent Argdown file with no chapter on any claim, which is to say a map that cites no text and that check_reconstruction cannot verify one word of. Nothing errored and nothing said anything was wrong, which is why it is worth naming here: a map that cannot be checked against its source is the one thing this project exists to prevent.

Connect the folder before you start — hosted clients ask for one, and every bridge has an equivalent — and the arrangement works exactly as the table says. If you cannot, run the server from something sitting on the same machine as your files.

What works:

• Claude Code — has its own file tools. Everything below works unmodified.
• Anything agentic with a workspace (Claude's Cowork, an IDE assistant like VS Code or Cursor) — same, provided it can reach the folder you are working in.
• Claude Desktop — the .mcpb installs and every tool runs, but on its own the assistant has no file access. Add a filesystem MCP server alongside this one, pointed at the folder you keep reconstructions in, and the whole workflow works.

How to check yours in one line, rather than trusting a list that will date: ask it to write a file called test.md in this folder, then read it back to me. If it can do both, it can drive Ipsissima-MCP end to end. If it cannot, you can still use the server for everything up to the reconstruction — extraction, page images, repair, the Zotero lookup — and then paste the map into a file yourself before asking for check_reconstruction. That works; it is just handwork the other clients do for you.

Which assistants can use this

Two questions that get confused, and they have different answers.

The server works with any MCP client. It is a plain MCP server — the official Python SDK, protocol 2024-11-05, spoken over stdin and stdout — and there is nothing Anthropic-specific in it. Claude Code, Claude Desktop, VS Code, Cursor and anything else that talks MCP can run it. Point the client at the ipsissima-mcp command, in whatever way that client is configured.

The .mcpb bundle is Claude Desktop's install format, not a protocol. No other client opens one. Elsewhere, install from source and configure the command yourself, as below. Nothing is lost by doing that; the bundle is a convenience, not a capability.

What genuinely varies is prompts and resources. MCP has three kinds of thing, and clients support them unevenly: tools are universal, prompts and resources much less so. And the unevenness is not the whole problem. The protocol makes prompts user-invoked and resources application-attached, so even a client that surfaces both beautifully gives the model no way to pull them mid-task — and it is the model, halfway into "make me an argument map", that needs the method: the Assertibility Question, linked versus convergent support, what fidelity levels mean, what to do about what an author did not say. There is very little Argdown in the world, and a model guessing at it confidently writes files that do not parse.

So the method is served on every channel, and the tool channel is the one that cannot go missing. The reconstruct_argument prompt and the ipsissima:// resources carry it for clients that surface them. For everything else — measured on a hosted session reaching this server across a device bridge that forwarded tools alone — extract_text returns a compressed conventions digest with the extracted text, and the reconstruction_method tool serves the full documents as ordinary tool results, one call each. A model that can call tools has the whole method; no client configuration and no user instruction is required. (The documents also sit in src/ipsissima_mcp/docs/ for anyone who wants to paste them by hand, but nothing now depends on that.)

A fourth thing can go missing, and it is the quietest of them. An MCP server also ships instructions — prose the client is meant to put in front of the model beside the tool list. This server's says in its opening lines that "make an Argdown", "map the argument" and "reconstruct this paper" are one request, and then gives the order of work. A client that drops it leaves the model holding a list of bare tool names with no reason to connect any of them to what was asked.

Observed, in a hosted session reaching this server across a device bridge: the instructions of every connected server were dropped, the tools arrived as bare names in one alphabetised list with no descriptions attached, and a request opening with the words "Make an argdown file" never called a single tool of the server built to answer it. Nothing errored. If you are running through a bridge or a hosted client and maps come back with no provenance, check this before you look anywhere else — and the fallback is the one above: hand the model extraction-prompt.md yourself.

The model is a separate question from the client. Ipsissima-MCP does not reconstruct arguments; it prepares sources and checks results, and the reading itself is the model's judgement. A weaker model behind a fully capable client will produce a weaker reconstruction, and check_reconstruction will report the difference rather than repair it.

Tell your assistant about it

Claude Code

claude mcp add ipsissima -- /absolute/path/to/ipsissima/.venv/bin/ipsissima-mcp

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "ipsissima": {
      "command": "/absolute/path/to/ipsissima/.venv/bin/ipsissima-mcp"
    }
  }
}

Use the absolute path. The server finds the Argdown parser and its own documents relative to where it is installed, not to where the client happens to be running.

claude mcp add is a command of the Claude Code CLI, so run it in a terminal where claude is on the PATH. It is not available from inside every client that can talk to an MCP server, and command not found: claude there means you are in the wrong window rather than that anything is missing.

Updating and removing it

The bundle. Claude Desktop's own extension settings install, update and remove it — a newer .mcpb from the releases page replaces the one you have. Removing it there is the whole uninstall: the dependencies live in an environment the host made and manages, not in your Python installation, so nothing is left behind in yours.

From source. Update with git pull and re-run the install command; an editable install (pip install -e) needs nothing more than the pull, which is what editable means. To remove it:

claude mcp remove ipsissima      # or delete the entry from the client's config
rm -rf .venv                     # the environment, and everything installed into it

Nothing is installed outside that virtual environment and the repository, and neither holds your reconstructions — those are ordinary files wherever you chose to keep them.

Nothing checks for updates on its own — but you can ask. The server makes no network request of its own accord, ever: no startup check, no timer. What it has is a check_for_updates tool, the server's equivalent of the application's Help ▸ Check for Updates menu item — ask your assistant "is there a new version of Ipsissima?" and one request goes to the GitHub releases API, only then. It sends nothing about you or your documents, downloads nothing, installs nothing: it names both versions and says how to update by the route you actually installed. For bundle users it matters more than the menu item does in the app, because the .mcpb format has no update feed of its own — this tool, or watching the releases page, is how you hear a new one exists. It also tells you that the desktop application is a separate program on its own schedule: a new Ipsissima-MCP does not mean a new Ipsissima.

Check it before you rely on it

The server speaks over stdin and stdout, so "it started" and "it works" are different claims, and a client that finds no tools will usually say nothing about why. This asks it directly:

.venv/bin/ipsissima-mcp <<'EOF'
{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"probe","version":"0"}}}
EOF

One line of JSON comes back naming the server and its version. A working install answers 11 tools (10 without a Zotero library on the machine), 2 prompts and 8 resources; if the count is lower than that, the client is not the problem. (That sentence is held against the live server by the test suite, so it cannot quietly drift the way it did twice: it said 9 for five days after check_for_updates made it 10, and the corrected 10 then promised the Zotero tool to machines with no Zotero — which the first public CI run caught within the hour.)

Note that pymupdf prints a deprecation notice on startup. It goes to stderr, where it is harmless — the protocol is on stdout — and it is not a sign of a bad install.

Using it

It takes a path, not an attachment. The server reads files from the disk it is running on: sources is a list of paths, and anything else comes back as no such file or directory. A document attached or pasted into the conversation is text in the conversation — the server never sees it and cannot be given it. Save it to a folder first and pass the path, which is one more instruction to the same assistant.

Within that, ask in your own words. All of these work, and all mean the same thing:

Use Ipsissima to map the argument in ~/papers/rawls-1971.pdf. Make an Ipsissima diagram of ~/papers/rawls-1971.pdf. Make an Argdown from ~/papers/rawls-1971.pdf.

Say "Ipsissima" if nothing seems to happen. The other phrasings are meant to work without it: the server ships instructions saying that all of them are one request, and a client that delivers those will route any of them here. Not every client does. When one drops them the model is left holding bare tool names with nothing to connect them to what you asked — and the literal word Ipsissima in your sentence is then the only thing that still matches, because it is in the tool names themselves. It costs nothing, so the first example says it.

One article. The assistant extracts the text, reconstructs the argument, and checks it until the checker is satisfied. You get a folder with source/the-paper.md and a .argdown beside it. Open the folder in Ipsissima.

Several articles, separate maps.

Make separate Argdown maps for each article in ~/papers/to-read. Put them in a new folder, with the extracted texts and the Argdowns side by side.

A book.

My manuscript is in ~/book. Create a single Argdown map using the latest draft of each chapter.

Text only, no reconstruction.

Just extract the markdown from these PDFs — don't make a map.

A debate, not a document.

Map the arguments for and against lowering the voting age.

No source text, so no extraction and no fidelity borders — the map is the assistant's survey of the debate, each side's best case laid out step by step, and it stands on being checkable on screen rather than checkable against a text. The checker still holds it to parsing and structure, and greets the genre as one: a sourceless map is not asked for provenance it cannot have.

An overview written first, then mapped.

Write me a report on the arguments about nuclear power, then map it with Ipsissima.

The report is a manuscript like any other — gold-route Markdown — so the map of it gets the full treatment: quotations verified, departures marked, the text beside the map. Two honest labels ride with it: the map's front matter declares text-provenance: generated, which Ipsissima shows beside the title, and what the checks establish is that the map is faithful to the report — not that the report is true.

It will ask before it spends

A reconstruction is the expensive step, and "a map of each of these articles" and "one map from these chapters" look identical from the files alone. So the server reports what is ambiguous rather than guessing, and refuses to convert several sources until the question is answered. If you are asked which of two drafts is current, or whether you want one map or six, that is the server declining to spend your money on a coin toss.

What a reconstruction costs

Measured over nine runs, not estimated. Roughly:

tokens  ≈  147,000  +  0.9 × bytes of map  +  2.3 × words of source

Three things follow, and they are the whole of what you need to know before pressing go.

Most of it is fixed. About 147,000 tokens go on the procedure itself — reading the instructions, reading the source, writing, checking, fixing — whatever the paper. That is 92% of the cost of mapping a 266-word passage and still 60% of mapping a long one. Reconstructing a short paper is not cheap, and there is no setting that makes it so.

A long source costs even when the map is small. This is the part that surprises. Two maps of the same 68,695-word book — one of 88,029 bytes, one of 39,708 — came out only 58,000 tokens apart while both sat about 90,000 above what their map size alone predicted. The reading has to happen before there is anything to be brief about. Budget on the length of the text, not on the size of the map you want.

In practice, at max effort: a journal article runs 200,000–250,000 tokens, a short passage about 160,000, and a book-length manuscript 280,000–350,000 — the last two figures being a whole book mapped once, not per chapter. Doing a book chapter by chapter costs the fixed 147,000 every time, so eleven chapters is over 1.5 million tokens before a single claim is written. Map the book in one pass; go back for the chapters that turn out to matter.

Two levers, and they are not equal:

• Effort is the largest and the one with a real cost. high came in 16% cheaper than max on the same paper — and made eight mechanical faults where max made none. Good for a draft you will check yourself; poor for anything you intend to publish or show its author. See eval/effort-testing/, which keeps both arms side by side.
• Round trips are worth about 47,000 tokens and cost nothing at all. They are already applied: the instructions carry four rules about them, and the same passage went from 41 tool calls to 6.

A caveat this file would rather state than bury. The source term rests on two runs at one source size, so do not extrapolate it much past 70,000 words. The workings, the residuals and what would falsify the model are in eval/COST-2026-08-27.md; eval/run_cost.py measures any run of your own from its transcript.

Which file to give it

This is the single most useful thing on this page. The same paper can arrive in three very different states, and the difference is permanent — a manuscript converted badly stays badly converted for every claim that ever cites it.

1. Markdown is gold

Already structured; nothing is recovered because nothing was lost. If you draft in Markdown, Ipsissima reads your live file, so the Manuscript view is always your current draft.

2. Anything pandoc reads is silver

.docx, .odt, .html, .epub, .tex. The document's own structure survives: a heading is a heading because the document says so, not because something guessed from the type size.

If you are not writing in Markdown, the best workflow is to export to Markdown yourself with pandoc and keep that file as the manuscript:

pandoc --wrap=none -o chapter-3.md chapter-3.docx

No pandoc installed? The web version at https://pandoc.org/app/ does the same job in a browser. When you want Ipsissima to show a newer draft, export again and save over the same Markdown file — the map is watching it, and the Manuscript view will update by itself.

3. PDF is bronze

A PDF records where ink sat on a page. Paragraphs, headings, columns, footnotes and reading order all have to be inferred from geometry, and inference has a hit rate.

So: do not give it a PDF of your own draft. You have the .docx. If a folder holds chapter-3.pdf beside chapter-3.docx, the server will say so before converting anything.

For published articles you do not have in any other form, PDFs are handled, and handled carefully:

• Where the article has a DOI, the server checks whether a machine-readable version exists (Europe PMC and other open-access routes) and prefers it. That turns a bronze source into a gold one for the cost of one lookup.
• Otherwise it converts, and assesses the difficulty first. A clean text layer is converted mechanically. A damaged one is OCRed, and every route tried is scored and reported so the choice is visible.
• Where neither works — an old scan, a strange typeface, essentially photographs of text — it says so, renders the damaged passages as images, and asks the assistant to read them. You get a usable result either way; what changes is how much it costs.

Age is a good predictor. A PDF made in the last decade is usually fine. One from twenty-odd years ago, or a scan of a photocopy, is where the difficult cases live.

Page numbers survive. Where the source is paginated, page breaks are kept in the Markdown as comments, and Ipsissima's Manuscript view shows the page number beside the text.

Zotero makes it better, and is not required

Everything above works with no Zotero installed. If you do have it, the server notices, reads it read-only (never the live database — a copy, because Zotero holds the file open), and can:

• build maps from items in your library, found by DOI, item key or title;
• use an item's HTML snapshot for the text and its PDF for the page numbers — the best combination there is, and the reason a snapshot is worth keeping. Each has exactly what the other lacks: a web page has no pages, and a PDF has no idea what a heading is. The page number is read off each sheet, not counted from one, because a paper whose front matter says it starts at 511 can print 514 on its first sheet;
• find snapshots you had forgotten were there, which read far better than the PDF beside them.

Nothing is ever written into your Zotero storage.

What the server offers

Prompts — reconstruct_argument (the full instructions, read off disk every run) and extract_text_only. Resources — the Argdown reference, the map semantics, the order views, the viewer guide, served on demand rather than loaded into every conversation.

The command line is still there

The server did not replace it, and for a one-off it is often quicker:

.venv/bin/ipsissima-ingest paper.pdf --out "samples/Author 2026 - Title"
.venv/bin/ipsissima-check reconstruction.argdown --source-root .
.venv/bin/ipsissima-split book.md --out chapters/

ipsissima-check has two modes worth knowing. Bare, it prints the full report — the apex, the sections, the tag census, the debt. --only-problems prints just what is wrong, and --format json prints the same as data. The short forms exist because a check-and-fix loop reads this file every round: on the Darwin sample the full report is 687 words and 10.5 seconds, and the short one is 221 words and 2.4.

The pieces

One rule that is not obvious

check_argdown.py and app/src/argdown-positions.js implement the same rule in two languages, because the viewer needs it in the browser and the checker needs it here. app/test_argdown_positions.mjs exists solely to police the drift between them, and it is the reason the viewer's build shells out to Python rather than reimplementing near-matching in JavaScript.

Tests

node app/run_all_tests.mjs        # everything, and the suite that matters

The MCP server's own tests drive it over a real stdio session, because the thing most likely to break is the contract with the client rather than the logic behind it.

Evaluation

eval/ holds the harness and the findings from choosing a converter. CONVERTER-FINDINGS.md is worth reading before replacing anything in the PDF route: marker, docling and a local llama.cpp were all measured and removed — about 3.9 GB between them — because pymupdf4llm with rapidocr beat them all on the labelled test set at a fraction of the cost.