Skip to main content
Glama

ipa-mcp: the department's imaging site for AI assistants (draft)

Lets an AI assistant answer questions about the UNIGE Biochemistry imaging site, biochem-ipa: the microscopes, hoods and lab VMs, their protocols, and the measurements made on them.

Is there a water objective on the Nikon scope? Has the laser power of the 3i spinning disk dropped since last year? Get me the PSF of the 60x. Is napari installed on the VMs? Open the booking page for the spinning disc.

When the answer is on the site, the assistant gives it with the page to check. When it is not, it says so and gives the way forward: the protocol, and that the facility can measure it with you.

It is an MCP server, so it is not tied to one assistant: any application that can start a local MCP server, with any model it offers. It reads your own synced copy of the site's library, read-only, with no credentials: what the assistant sees is what your account can already read. What you then send to a model is your decision, as with any file you open.

Install

  1. Sync the library once. Open the site's library and click Sync. OneDrive creates a folder named biochem-ipa - Documents. Large raw files stay placeholders until something reads them.

  2. Install uv and git (uv uses it to fetch the server).

  3. Install the server in your application.

Windows: double-click

Download the repository (on GitHub: Code, Download ZIP) and unzip it. In its install folder, double-click the script for your application:

Script

Installs in

Then

claude-code.bat

Claude Code, for every project

start a new session

claude-desktop.bat

Claude Desktop

quit it (also from the tray icon) and open it again

vscode.bat

VS Code, your user profile

MCP: List Servers shows ipa

check.bat

nothing: says whether the server starts and finds the library

Each script checks that uv and git are there, fetches the server and starts it once (so the application does not wait for the first download), says whether it found the library, then registers it. Running it again updates the server; -Remove (for example claude-code.bat -Remove in a terminal) takes it out again. The unzipped folder is not needed afterwards.

By hand (any system, any application)

The server runs straight from the repository with uvx --from git+https://github.com/unige-biochem/ipa-mcp ipa-mcp. Run this once first, with --status at the end: the first run downloads and builds it, which can take longer than an application waits for a server to start, and it shows whether the library was found. Then register it:

  • VS Code: Install in VS Code

  • Claude Code, for every project, then start a new session: claude mcp add -s user ipa -- uvx --from git+https://github.com/unige-biochem/ipa-mcp ipa-mcp

  • Claude Desktop: Settings, Developer, Edit config, under mcpServers: "ipa": {"command": "uvx", "args": ["--from", "git+https://github.com/unige-biochem/ipa-mcp", "ipa-mcp"]}, then quit and reopen it.

  • Anything else: command uvx, arguments --from git+https://github.com/unige-biochem/ipa-mcp ipa-mcp.

The server finds the synced folder by itself (OneDrive's registry on Windows, the usual places on macOS). If it does not, set IPA_LIBRARY to the folder, or tell the assistant where it is. The first time it matters, the assistant asks who you are and which lab you are in, and remembers it in ~/.ipa-mcp/config.json.

It is not on PyPI, and will not be: it only makes sense for the department's site, so it is installed from git. uvx ipa-mcp alone does not work. To stay on one version, add it to the URL (git+https://github.com/unige-biochem/ipa-mcp@v0.1.0).

If the tools do not show up

  • Restart. Applications load their servers when a session starts: a new Claude Code session, Claude Desktop quit and reopened, or MCP: List Servers, ipa, Start in VS Code.

  • See what it says. In Claude Code, /mcp lists the servers and their state, and claude mcp get ipa shows where ipa is registered.

  • First start too slow. Run uvx --from git+https://github.com/unige-biochem/ipa-mcp ipa-mcp --status once (or check.bat), then restart.

  • "MCP server ipa already exists". It is already registered: nothing to do. To replace it, claude mcp remove ipa first, or run claude-code.bat again.

  • Works in one folder only (Claude Code). claude mcp add without -s user registers it for the current folder only, and that entry wins there. In that folder, claude mcp remove -s local ipa, then add it with -s user.

  • Library not found. --status gives the setup steps; or set IPA_LIBRARY.

  • Newer version. uvx keeps what it built: run the --status command above with --refresh after uvx (or the .bat again), then restart.

Testing from a checkout

Clone the repository and run the same scripts with -Checkout (for example install\claude-code.bat -Checkout): they register the checkout, with uv run --directory <checkout> ipa-mcp, instead of the GitHub repository. For the Claude Desktop bundle, see DEVELOPMENT.md.

Related MCP server: knowledgehub

What it knows

From

Every page of the site: hardware, software, start-up, quirks, protocols, VMs

site/*.md

The equipment registry: objectives, light paths, laser lines, rooms

site/registry.json

Laser power at the sample over time

data/laser-power/laser-power.csv

PSFs per objective and light path, with the PSF images

data/psf/psf.csv, data/psf/<session>/

Courses, presentations, viewers

builds/*/*/build.yaml

The copy is as fresh as OneDrive's last sync. The site is the reference, and the assistant cites its pages. The library's private/ folder is never read.

Development

See DEVELOPMENT.md.


Created by: Nicolas Chiaruttini. GitHub (code, issues, releases): https://github.com/unige-biochem/ipa-mcp. The facility's project record stays on GitLab: https://gitlab.unige.ch/unige-biochem/common/ipa-mcp

Data folder (Windows): I:\common\ipa-mcp. macOS (Finder, Go, Connect to Server): smb://isis.unige.ch/nasac/s/biochem/biochem-ipa/ipa_biochemistry/common/ipa-mcp.

Confidentiality level for this project: open: code, scientific question and sample images may be shown to AI tools.

Available Tools

13 tools
ipa_bookingC

Where to book an item: its page (the calendar is embedded there), the SharePoint list behind it, and the booking directory of the whole site. Booking is per lab: confirm the user's lab (ipa_status) before opening another lab's scope.

Inputs: id_or_name (required).

ipa_buildsC

The courses, presentations and viewers on the site (interactive HTML in the library), with title, summary, URL and source repository. kind: course | presentation | viewer.

Inputs: kind. Returns: result

ipa_equipmentA

Every item on the site - microscopes (kind 'scope'), hoods, VM desktops - with id, lab, aliases, room, page, objectives, light paths and laser lines. Filter by lab and/or kind.

Inputs: lab, kind. Returns: result

ipa_filesA

List the files of a library folder - 'data/equipment/', 'data/psf//raw', 'data/laser-power' - with size, URL and whether OneDrive has downloaded each one. Raw acquisitions are large: list before reading, and read only what is needed.

Inputs: folder (required), max_files. Returns: result

ipa_findA

Which item does the user mean? 'the spinning disc', 'the Nikon scope', 'Dobby'... Ranked candidates; the user's lab (from ipa_status) breaks ties. If several labs have a match, ask the user which lab rather than guessing.

Inputs: query (required).

ipa_itemA

Everything the site knows about one item, structured: identity, objectives (with immersion and NA parsed from the tag), light paths, laser lines, the page's outline, where to book it, and which measurements exist (PSF per objective, laser power per line). Read a page section with ipa_page for the prose.

Inputs: id_or_name (required).

ipa_laser_powerA

Laser power at the sample, per scope, laser line and light path: every measurement by date with the change since the first and since the previous one (per cent). Scope is an id from ipa_find. With nothing measured for the scope, says so and gives the protocol and whom to ask. The protocol compares runs within one light path only.

Inputs: path, scope, line_nm.

ipa_openB

Open a page of the site, a booking list or a library file in the user's browser. Only URLs of the site are accepted.

Inputs: url (required).

ipa_pageB

Read a page of the site (Markdown): by id ('dudin_scope_sdc-nikon', 'lab-vms', 'protocol-psf', 'image-analysis') and optionally one section by title ('Software', 'Known quirks', 'What is installed'). A long page without a section gives its outline and the first part; ask for sections to read the rest. Cite the page's URL in answers.

Inputs: name (required), section, max_chars.

ipa_psfB

PSF measurements: per session (scope, objective tag such as '100x-146-oil', light path, date), the FWHM in x, y, z per channel against the expected value, the chromatic shift, and the local path of each channel's PSF image (psf_.tif, for deconvolution). Newest first. With nothing measured, says so and gives the protocol and whom to ask.

Inputs: path, scope, objective.

ipa_set_userA

Remember who the user is (name, lab - Aumeier, Dudin, Kaksonen, Kruse, Roux or other) and, if auto-detection failed, where their synced library is. Ask the user once, then call this.

Inputs: lab, name, library.

ipa_statusA

Is the synced library here, how fresh is it, and who is the user? Call first when anything fails; it gives the setup steps when the library is missing.

Inputs: no named parameters.

Tool Schema Changelog

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

Loading tool schema changelog…

    TDQS

    A3.5/5.0

    Scored across 13 tools

    Disambiguation5/5

    Each tool targets a clearly distinct resource or action: setup/status, inventory listing, item resolution, item detail, page reading, global search, laser-power measurements, PSF measurements, builds, booking pointers, opening URLs, and file listing. The only minor overlap is that ipa_item mentions booking and ipa_booking elaborates on it, but the descriptions make the boundary clear.

    Naming Consistency4/5

    All tools consistently use the ipa_ prefix and lowercase snake_case, making the set predictable to scan. The names mix nouns (ipa_item, ipa_page) and verbs (ipa_find, ipa_search, ipa_open), so it is not a strict verb_noun convention, but the style is still coherent.

    Tool Count5/5

    The 13 tools fit comfortably in the well-scoped range and each appears to earn its place across status, inventory, search, measurements, bookings, and file access. The surface is broad but not bloated.

    Completeness4/5

    Core workflows are covered: setup and user identity, equipment discovery, item and page lookup, search, laser/PSF measurements, builds, booking pointers, and library file listing/opening. Minor gaps remain around direct booking availability/calendar access and reading file contents rather than only listing/opening them, though browser/page workflows provide workarounds.

    Maintenance

    ActivityMaintained
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers