jstage-mcp

A FastMCP stdio server exposing the J-STAGE WebAPI as three tools for use with Claude Desktop.

What this is for

J-STAGE holds the full text of journals published by Japanese learned societies, and this searches inside the articles rather than across a catalogue. A term that no cataloguer chose as a keyword is still findable if an author used it in an argument, which makes this the route for concepts that circulate before they are named.

Resolve a J-STAGE DOI straight to its record, or walk a journal's volume and issue spine to see a run whole.

Run a term here and on cinii-mcp and read the gap: a wide divergence tells you whether your vocabulary belongs to catalogue description or to the prose of the field, which is a finding about the literature before it is a finding in it.

Related MCP server: Japan Data MCP

What the receipts are for

A search you cannot re-run is a claim you cannot check. When a footnote rests on a database query, say that no article in this index uses a term before a certain year, the reader is asked to take the search on trust: which term, in which script, on what date, against which index and which version of it, and how far down the results the author went. Ordinary searching leaves none of that behind. This server leaves all of it. Every query-answering tool returns its envelope through the ledger, which appends one line to an append-only file: the term actually sent and its script, how the source matched it, how many records existed and how many came back, the diagnostics, the tool and its parameters, the server version, a timestamp, and the hash of the previous line. The hash makes the file a chain: a line cannot be altered, removed or reordered afterwards without the verifier saying so.

What that gives a researcher:

• A citable search. Name the receipt in the footnote (session slug, server, date, line hash) and a reader can see exactly what was asked and run it again against the same version.

• Negative findings that carry weight. "Not found" is evidence only if the search that produced it is on record, with its term, its script and its breadth.

• A method section that writes itself. jstage-mcp-ledger manifest <folder> summarises every query a project made, by server, script and session: the disclosure a journal, a data-availability statement or a research-integrity review asks for.

• A record of AI-mediated research. When a model chose the term, the receipt shows the term it chose and what came back, which is the thing to disclose about work done with an assistant.

• Nothing interpreted. The receipt is the source's own answer with credentials removed. The server does not summarise, rank or paraphrase, so the record is of the source, not of the tool.

Receipts are off until you name a folder (MCP_RECEIPT_DIR); each server then writes its own <server>.jsonl inside it, and MCP_RECEIPT_SESSION stamps a project or article slug on every line so one folder can serve several projects. jstage-mcp-ledger verify-dir <folder> checks the chains. The mechanics, the variables and what the envelope says when nothing is deposited are in the receipts section below.

Tools

All tools return one typed JSON response envelope with bilingual (English / Japanese) titles, authors, and journal names where J-STAGE provides them — see Response format below. The JST attribution requirement is met by the envelope's attribution field, present in every response.

Response format

Every tool returns one JSON response envelope, built by mediation.py and defined in response-schema.json. Schema version 2.3.0. The same module and schema are vendored byte-identically across the server family, so an envelope from one server can be read by a consumer written for another.

The envelope reports how the search was made, not only what it found:

• searched_for — on search operations, the term actually sent, its detected script, and the matching mode, hoisted to the top of the envelope so a relaying client cannot drop it. Fetch operations (jstage_get_article_by_doi, jstage_list_issues) omit it: they were handed an identifier and chose no term.

• query — input_terms as supplied, normalized as sent, and the detected script. This pair is the record of any rendering performed between the caller's language and the corpus.

• matching_mode — full_text_broad for this server. It tells you how to read result.total.

• result.breadth — none, narrow (1–50), broad (51–1000), very_broad (>1000). Thresholds are low on purpose: a few hundred hits that look like a literature are marked rather than passed through clean.

• items[].matched_in — which field the match was made in, per record.

• receipt — an ISO 8601 timestamp, a SHA-256 taken over the normalised query and its parameters, and the identifiers returned. The hash verifies a term you already hold; it cannot be inverted to produce one, so the unit of deposit is the envelope, not the receipt.

• attribution — the required credit line, in every response.

Diagnostic codes

Typed and closed. A diagnostic is never prose the client has to parse.

Query receipts

Every envelope can be deposited to an append-only, hash-chained JSONL log by ledger.py. It is off unless MCP_RECEIPT_DIR (or the legacy MCP_RECEIPT_LOG) is set, and a logging failure is swallowed rather than raised — a search matters more than the record of it. Secrets are redacted before a line is composed.

Since schema 2.3.0 the envelope says so. When a response is not deposited, emit() appends RECEIPT_NOT_DEPOSITED if the variable is unset, or RECEIPT_WRITE_FAILED if it is set and the write did not land. The gap is then visible in the artefact that becomes the record, rather than only in a configuration file. mediation.deposit_enabled() reports the same fact on demand.

MCP_RECEIPT_DIR=C:\path\to\receipts        # a folder, not a file
MCP_RECEIPT_SESSION=project-or-article-slug
MCP_RECEIPT_STRICT=1                         # optional: make logging failure raise
MCP_RECEIPT_LOG=C:\path\to\receipts.jsonl  # legacy single file; ignored when _DIR is set

A folder, and one file per server. MCP_RECEIPT_DIR points at a directory and each server writes its own <server>.jsonl inside it. That is not tidiness. Appending is read-the-last-hash-then-write, and the lock around it is a threading lock, which holds within one process and not between several — six servers are six processes, and two answering at the same moment will both read the same predecessor and both claim it. Measured, not theorised: six processes writing 150 lines to one file produced fourteen forks. MCP_RECEIPT_LOG still works and is still correct for a single server; it is the wrong shape for a family.

install.ps1 sets this up for all six and writes a README into the folder.

Verify one chain, or the whole folder:

jstage-mcp-ledger verify      receipts/jstage.jsonl
jstage-mcp-ledger verify-dir  receipts
jstage-mcp-ledger manifest    receipts        # writes receipts/manifest.json

verify exits non-zero on failure and says which kind it found: a fork (concurrent writers — a configuration fault, and every line is still there), a missing line, a reordering, or tamper (a line that does not hash to its own content). Only the last is a claim about honesty, and reporting them alike would invite a reader to mistake one for the other. The manifest is the object to cite: one description of the whole deposit — per-file line counts, first and last timestamps, terminal hashes, and combined totals by server, script and session.

Install

Three routes. All three give you the same server; pick by how much you want to see of it.

Python. The pip and source routes need Python 3.10 or later; 3.10, 3.12, 3.13 and 3.14 are tested in CI on Windows, macOS and Linux. The Claude Desktop bundle uses whichever of these is already installed, and has uv download one only if none is.

Getting Python

Every route needs Python 3.10 to 3.14. The Claude Desktop bundle uses one already on the machine and has uv download one only if none is; the other routes also need the venv module, which the official installers include.

• Windows. Download the 64-bit installer from python.org/downloads and run it; tick "Add python.exe to PATH" on the first screen. Afterwards py --version (the launcher the installer adds) or python --version in a new terminal should print 3.1x. If typing python opens the Microsoft Store instead, Windows has no Python yet: that Store page is a stub, and it is also what "'python' is not recognized" usually means.

• macOS. The python.org installer, or brew install python@3.13 with Homebrew. The /usr/bin/python3 that Xcode's command-line tools provide may be older than 3.10; python3 --version says.

• Linux. Your distribution's package: sudo apt install python3 python3-venv on Debian and Ubuntu, sudo dnf install python3 on Fedora. Or let uv provide one (next line).

• Any platform, with uv. uv installs Python itself: uv python install 3.13, then uv venv or the uvx route below.

One click: the Claude Desktop bundle

Download jstage-mcp-3.1.2.mcpb from the latest release and open it; Claude Desktop installs it. One bundle serves Windows, macOS (Apple Silicon and Intel) and Linux. Claude Desktop asks only for a receipts folder at install time.

The bundle carries the server's source and a lock file, nothing compiled, and needs no Python of its own. Claude Desktop builds the bundle's environment when you install it, with uv, a copy already on your PATH if there is one and otherwise one the app downloads for itself: uv takes a Python 3.10 or later already on the machine, downloads one only if there is none, and installs the locked libraries, roughly 40 MB, behind the install progress bar. Every launch then reuses that environment and takes under a second. Bundles 3.1.0 and 3.1.1 kept pyproject.toml one folder down, which made Claude Desktop skip that install step and download everything during the first connection attempt instead; on a slow or filtered network that attempt never completed and the app reported it could not connect to the extension server. Bundles before 3.1.0 vendored libraries compiled for CPython 3.12 only and failed on every other interpreter. See Troubleshooting.

From GitHub, pinned to a release

pip install "git+https://github.com/ckgerteis/jstage-mcp@v3.1.2"
# or, without an environment of your own:
uvx --from "git+https://github.com/ckgerteis/jstage-mcp@v3.1.2" jstage-mcp

installs the jstage-mcp console script and jstage-mcp-ledger. The tag is the thing to cite; @main gets whatever is current. Then register it in Claude Desktop (below), or let install.py do that.

The whole family

pip install "git+https://github.com/ckgerteis/bibliograph-mcp@v1.0.4" && bibliograph install

installs all six servers and registers them together — one receipts folder, credentials asked for once. See bibliograph-mcp. From a checkout of this repository, python install.py does the same for this server alone, python install.py --all for the six, on Windows, macOS and Linux; install.ps1 remains for Windows.

From source

python3 -m venv .venv
.venv/bin/pip install .

On Windows:

py -3.11 -m venv .venv
.venv\Scripts\pip.exe install .

Or straight from the repository, without cloning:

uvx --from "git+https://github.com/ckgerteis/jstage-mcp" jstage-mcp

Verify the install:

.venv/bin/python -c "import jstage_mcp; print(jstage_mcp.__version__)"

That fails loudly if the package or one of its vendored modules is missing. Do not use jstage-mcp --help as the check: unknown arguments are ignored, the server starts, reads end-of-input and exits 0, so it reports success whatever the state of the code.

Installing more than this one

Six independent packages. None imports another, none depends on another, and each installs and answers on its own — pip install . in this directory is a complete install of this server and nothing else.

They do share three things: a response envelope, a query ledger, and — if you run more than one — a receipts folder. install.ps1 is vendored byte-identical into all six and handles that on Windows; install.py is its cross-platform port. Both install this server by default, because cloning one repository is not a request for five more.

.\install.ps1                        # this server
.\install.ps1 -All                   # all six
.\install.ps1 -Servers jstage,cinii        # a chosen subset

Nothing about where things go is decided for you. The script asks where to install (the virtual environment Claude Desktop will be pointed at), which folder receives the receipts, and which session slug to stamp on them, offering a neutral suggestion for each that Enter accepts; run without a terminal it does not guess, and stops unless --venv and --receipts-dir (or --no-receipts; -VenvDir and -ReceiptsDir for install.ps1) say so. Whatever subset you name is registered against one receipts folder, asked for once. The script prefers a sibling checkout to the network, carries across credentials already registered rather than asking again, leaves servers it was not asked about alone, and stops rather than guessing where the servers already registered disagree about the folder or the session slug. It also asserts that ledger.py and mediation.py are byte-identical across everything it installed, so two envelope versions cannot end up in one environment unnoticed.

Any other MCP client

Nothing here is specific to Claude. The server speaks the Model Context Protocol over stdio and nothing else: any client that can start a process and talk JSON-RPC to it (Claude Code, Cursor, VS Code and Continue, Zed, LibreChat, a script of your own using an MCP SDK) can use it. The Claude Desktop bundle and the installers are conveniences for one client; the server underneath is the same console script. Register it anywhere by giving the client the absolute path of the console script and, optionally, the environment:

{
  "mcpServers": {
    "jstage": {
      "command": "/absolute/path/to/.venv/bin/jstage-mcp",
      "env": {
        "MCP_RECEIPT_DIR": "/absolute/path/to/receipts",
        "MCP_RECEIPT_SESSION": "project-or-article-slug"
      }
    }
  }
}

Claude Code takes the same thing on the command line:

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

On Windows the path ends in \.venv\Scripts\jstage-mcp.exe. MCP_RECEIPT_DIR and MCP_RECEIPT_SESSION are optional; without them the server runs and every envelope says RECEIPT_NOT_DEPOSITED. The stdio transport is the only one: there is no HTTP endpoint to expose, and nothing to host.

Troubleshooting

"Server disconnected" is all Claude Desktop says when the server process exited before or during the handshake, whatever the reason. The reason is in the log:

• Windows: %LOCALAPPDATA%\Claude\Logs\mcp-server-<name>.log (builds before August 2026: %APPDATA%\Claude\logs) (the extension's display name, or the key under mcpServers), with mcp.log beside it for the app's side of the conversation.

• macOS: ~/Library/Logs/Claude/mcp-server-<name>.log and mcp.log.

• Linux: ~/.config/Claude/logs/.

Read the last launch from the bottom up. Three shapes account for nearly every report:

• A Python traceback ending in ImportError or ModuleNotFoundError (for example No module named 'pydantic_core._pydantic_core'). The interpreter started, the code was found, and a compiled library did not match that interpreter. This is what every bundle before 3.1.0 did on any Python other than 3.12. Install the current bundle, or use the pip route, which resolves wheels for the interpreter you install into.

• 'python' is not recognized, spawn python ENOENT, or a line from the Microsoft Store: no interpreter was found on the PATH Claude Desktop constructs. Nothing of this server ran. The current bundle does not launch python at all; for the pip route, register the console script by absolute path as shown above.

• A line from uv (error: ..., or a download that never finished), or Unable to connect to extension server with nothing from the server in the log: the environment was not built. From 3.1.2 the app builds it at install time; look in main.log beside the server log for lines tagged [UV Runtime], which record the download and the uv sync, and for missing pyproject.toml, which means a bundle older than 3.1.2. If the install-time build failed (no network, a proxy that blocks github.com or pypi.org), reinstall the bundle once the network is back; the launch also rebuilds the environment itself, so a second launch on a working network completes.

The bundle's own entry point writes one line naming the interpreter, its path and the supported range before re-raising an import failure, so a log from 3.1.0 onwards says which of these it is.

Claude Desktop configuration

Add an entry to %APPDATA%\Claude\claude_desktop_config.json under mcpServers, pointing at the console script in the environment you installed into. On macOS or Linux use the absolute path to .venv/bin/jstage-mcp.

{
  "mcpServers": {
    "jstage": {
      "command": "C:\\path\\to\\.venv\\Scripts\\jstage-mcp.exe"
    }
  }
}

Changed in 3.0.0. Earlier versions were registered by path — "command": "…\\python.exe", "args": ["…\\server.py"]. That entry will not start this version, because server.py is now a module inside a package rather than a script beside its imports. Replace it with the console script above.

Restart Claude Desktop. The three tools should appear under "jstage" in the tool list.

Rate limiting

The server enforces a one-second minimum interval between outbound requests in line with JST's prohibition on bulk downloads. The limit is per-process; if you run multiple Claude Desktop sessions concurrently you may exceed it, so don't.

Limitations

• There is no journal-search tool. jstage_search_journals existed in v1.x and was removed in v2.0.0. J-STAGE announced a journal-search endpoint (service=4) on 26 March 2026 and the public API still rejects that service code with ERR_004; a tool that silently falls back to volume search is not a journal search, and this server would rather not offer one. Until JST activates service=4, use jstage_list_issues against a known title, ISSN or cdjournal.

• jstage_get_article_by_doi requires J-STAGE-issued DOIs. The WebAPI does not expose a doi= query parameter. The tool decomposes DOIs that follow J-STAGE's pattern (10.<registrant>/<cdjournal>.<vol>.<no>_<page>) into cdjournal+vol and matches the result against the response. For DOIs outside that pattern the tool returns the doi.org resolution URL with a note.

• Commercial use requires registration. Per the JST Terms of Use, commercial use needs an application form sent to contact@jstage.jst.go.jp. Research and teaching use does not.

API notes

Endpoint: https://api.jstage.jst.go.jp/searchapi/do

Service codes used:

• service=2 — Volumes/issues

• service=3 — Article search

• service=4 — Journal search (documented, rejected with ERR_004 as of 23 August 2026; not used by any tool)

Valid article-search query parameters confirmed against the live API: material, article, author, affil, keyword, abst, text, issn, cdjournal, vol, no, pubyearfrom, pubyearto, start, count.

Attribution

Powered by J-STAGE

This string is included in every tool response.

Citation

If this software supports your research, please cite it. See CITATION.cff, or use the "Cite this repository" button on GitHub.

Tests

.venv/bin/python tests/smoke_stdio.py

Starts the installed console script over stdio, performs the MCP handshake, and checks tools/list against the tool table above; exits non-zero on a mismatch. RUN_LIVE=1 … <tool> '<json params>' adds one live call and reports the envelope's diagnostic codes.

License

MIT © 2026 Christopher Gerteis.

This license covers the server code only. It grants no rights over J-STAGE content or the J-STAGE WebAPI, which remain governed by JST's Terms of Use.

Disclaimer

A research tool, maintained on a best-effort basis and provided "as is", without warranty. Not affiliated with or endorsed by the Japan Science and Technology Agency. JST does not provide support for the WebAPI.

Author

Dr Christopher Gerteis, SOAS University of London.