Exegete

A qualitative analysis application you use in conversation with an AI assistant, compatible with QualCoder.

Analyse your interviews by asking, in your own words. Your AI assistant reads, searches and suggests; you decide. Every suggested coding quotes the text word for word and waits for your approval, and your project stays one that QualCoder opens. QualCoder is free software for qualitative analysis, of the same kind as NVivo, ATLAS.ti and MAXQDA.

You: Bring this into the project Practice as "Interview 3". [text]
Assistant: First: which name should my work be stored under?
You: "AI assistant".
Assistant: Done. Practice now holds Interview 3.
You: Add a code "coping"; suggest codings in Interview 3.
Assistant: What should I look for, how long should a passage be, and may a passage carry more than one code?
You: Ways of managing alone; a sentence or two; yes.
Assistant: Three suggestions, each quoting the text word for word:

1. "I managed on my own: I made lists." (explicit)
2. "Nobody rang that winter, so I walked." (interpretive: rests on "so I walked"; coping implied, not said)
3. "My sister came at weekends." (interpretive)

You: Approve 1 and 2. Reject 3: that is support, not coping.
Assistant: 2 approved, 1 rejected, 0 pending. Apply them?
You: Yes.
Assistant: Backup taken; 2 codings written under "AI assistant".

An illustration, shortened, with made-up practice text; your assistant's words will differ.

Exegete (formerly qualcoder-mcp) is a qualitative analysis application in its own right, with no window of its own: your assistant, such as Claude Desktop, starts it, and you work in the conversation, from starting a project to exporting your codebook and coding. It is built to stay interoperable with QualCoder: your project is a folder on your computer, kept in QualCoder's format and conventions, so you can work on it in either program, one at a time. You do not need QualCoder to start, or running while you work, though a few things are not in Exegete yet (listed below).

• New here? Start here: Claude Desktop, one click, then a practice project.
• Interviews or other participants' data? First read Where your data goes: the AI behind your assistant runs on its maker's computers, and what it reads goes there; some assistants also open files by themselves.
• Advanced user, developer or IT? For advanced users: an MCP server over stdio, two tool sets, every tool, pipx install exegete.

Exegete is free, open-source software, under QualCoder's own licence (LGPL-3.0-or-later), and costs nothing (your assistant may: see "Start here"). It is tested on Windows, macOS and Linux with every change. It is an early version (an alpha) built by one researcher, independently of QualCoder's developers; parts marked Experimental have had little or no use yet. Questions and ideas go to GitHub Issues, not email; never put participant data in an issue.

What you can do

• Start a project in QualCoder's format, or work on a copy of yours.
• Bring in transcripts as documents from your computer (Word, OpenDocument, RTF, text, Markdown, web pages, subtitles; PDF and EPUB with the optional part, which the extension has), read by Exegete on your computer, their text kept off the conversation, or through the conversation; sort them into cases with attributes such as age, role or site.
• Code the files you choose, with suggestions you approve: each quotes the text word for word (Exegete checks the words are in the file), with the assistant's reading: explicit (the passage states what the code names) or interpretive (the code rests on what it implies). Whether a code fits the words, and what a coding means, stays your judgement.
• Grow the codebook: new codes the assistant proposes from your files, created once you approve them; rename, recolour, move, merge and delete, with a preview before merging or deleting.
• Explore: search texts, codings and memos; frequencies; codes that occur together; a case-by-code matrix; cases and files by attribute; a whole transcript with its codings, read by the assistant, or by you on a page in your browser.
• Write memos, annotations and a research journal.
• Coder comparison: per-code agreement between two coders, with QualCoder's own coefficient and Cohen's kappa side by side.
• Replace participants' names in text already coded, every coding moving with the words, with a count of listed names that remain (other identifying details are not looked for).
• Export the codebook, a coding report, frequencies and the case-by-code matrix as CSV, text or Markdown.
• Go back: a backup before each change, by default, and a way to restore one.

Not in Exegete yet, and done in QualCoder for now: bringing in images, audio and video; coding by selecting text yourself; coding images, audio, video or an area of a PDF page; graphs; and QualCoder's Reports menu. If your study needs any of these now, you may want QualCoder from the start.

The aim is the whole life of a project in Exegete, from its creation to the finished analysis, without needing QualCoder for any of it, while every project stays one that QualCoder opens.

How it works

Exegete runs on your computer. It has no AI of its own. The AI model behind your assistant, which reads and suggests, runs on its maker's computers unless it is a local one.

┌─ Your computer ──────────────────────────────────┐
│  You ─ ask ─► Assistant app ◄────────────────────┼─┐
│               (Claude Desktop, ChatGPT...)       │ │
│                    │ uses Exegete's tools        │ │
│                    │ (MCP, over stdio)           │ │
│                    ▼                             │ │
│               Exegete (no AI of its own)         │ │
│                    │ reads (read-only) and       │ │
│                    │ writes (after a backup)     │ │
│                    ▼                             │ │
│               Your project, in QualCoder's       │ │
│               format                             │ │
│                    ▲ one program at a time       │ │
│               QualCoder (optional)               │ │
└──────────────────────────────────────────────────┘ │
  The AI model, on its maker's computers ◄───────────┘
  (or yours, if local): what the assistant reads
  through Exegete goes there. Some assistants also
  open files by themselves: see below.

When anything is written. Reading changes nothing in your project. Suggested codings and proposed codes wait in a review list outside the project until you approve them and the assistant writes them. Other changes, such as making a code or writing a memo, are made when the tool runs; the larger ones (merging or deleting codes and categories, replacing names, restoring backups, bringing documents in, correcting your own coder name) show a preview first. A suggested coding's path:

You ask        "Suggest codings for coping in Interview 3"
  │
  ▼
The assistant  reads the file through Exegete and suggests;
  │            Exegete checks each quote is the file's own words
  ▼
Review list    outside your project, not yet written;
  │            the assistant is told to show each passage with
  │            its reading and its reason
  ▼
You decide     approve, reject or reopen, in the conversation;
  │            the assistant passes it on: check the counts
  ▼
Apply          Exegete checks again, takes a backup, then
  │            writes every approved coding, or none
  ▼
Your project   the codings, under the AI coder name you chose

Your approval, and its limit. Exegete hears only from the assistant, so it records the approval the assistant reports: it cannot tell whether you gave it. Keep the assistant asking for the steps that record your decisions and write them (INSTALL.md names them): "Allow once" in Claude, each prompt answered in Codex. Before any coding is applied, check that the counts (approved, rejected, pending) match what you said; if not, say so.

Where your data goes

In short. Exegete has no online service. All it sends itself is a request for a file that says whether a newer version exists, with nothing from your projects: at most once a week, and when you ask, while checking is on, as it is in the Claude Desktop extension unless you switch it off ("Updating", below). What the assistant reads through it (passages, codes, memos, names) goes to the maker of the AI behind your assistant: Anthropic for Claude's apps, OpenAI for ChatGPT's desktop app and Codex, no one with a local model. Text you paste or attach goes in full; a document in the project, brought in through Exegete or imported in QualCoder, only as far as the assistant reads it: Exegete's preview of an import gives the assistant the documents' names, sizes, lengths and warnings, never their text (PRIVACY.md, "Bringing documents in").

Assistants that open files by themselves. Exegete's answers hold some things back from the assistant, such as the private part of a memo (the ##### mark below). Some assistants can also open files on your computer by themselves, outside Exegete. What they read that way goes to their AI's maker in full: Exegete cannot see such a read or stop it, and its protections (that mark, your approval before codings are written, the backups) do not apply to it. So, with participants' data, it matters which assistant you use and how it is set up:

Exegete's own answers also tell the assistant where your project is, so one that opens files by itself can find it (PRIVACY.md, "Assistants that open files by themselves"). Which terms apply is set by your account, not by Exegete; institutions should prefer organisational accounts (INSTALL.md, "Choosing your AI host").

Before you use participants' data with Claude Desktop's chat (also where chat and Cowork are one conversation, below):

1. Keep computer use off (Settings, General): it lets Claude see your screen and use other apps.
2. Keep folders that hold your projects or transcripts unconnected (your home folder, Documents or a whole drive included; connected folders may be listed under "Trusted folders"), and add no other extension that reads files: Claude reads a connected folder by itself, and such an extension can reach your project too.
3. Switch training off before participants' data: while it is on, Anthropic may use your conversations to train its models. On a personal plan (Free, Pro or Max) it is the Model Improvement setting, at https://claude.ai/settings/data-privacy-controls. Rating a reply (thumbs up or down) can still let Anthropic train on that conversation (PRIVACY.md quotes the terms).
4. On an account your university or employer provides, ask whoever manages it which terms apply.
5. Take PRIVACY.md's checklist to your ethics committee or data protection officer.

Private notes and names. Exegete never passes the part of a memo from a ##### mark onward (QualCoder's mark for a private note) to the assistant, whichever program made the project. The mark works in memos, annotations and journal entries, not in the text of a transcript, and exported files keep the whole memo, private part included. Replacing names reduces the risk; it does not make anyone anonymous. Exegete can replace them, with the project's names list as documents come in or with pseudonymise_source afterwards. The list can be one you fill in on your own computer, in a file Exegete prepares, so the names do not pass through the conversation. Replacing names helps you manage what you share with other people, though the project's folder still holds the originals and any names list. For what the assistant reads, the replacement should not be trusted blindly: Exegete's tools return the pseudonyms where names were replaced (PRIVACY.md says what is not replaced), but an assistant that opens files by itself can reach the originals, and the names file, outside Exegete. To keep real names from the assistant altogether, a safer way is to replace them before the documents reach the computer where the assistant works (PRIVACY.md, "What pseudonymisation is for").

What stays on your computer, unless a sync service such as iCloud copies its folder (PRIVACY.md says where):

• your project, backups, exports and pending suggestions;
• the pages and copies Exegete opens for you to read a file (an EPUB opened in Apple Books joins its library, and iCloud if Books syncs);
• the texts an import's preview read, kept in Exegete's own folder until the import (PRIVACY.md says for how long);
• the files you fill in for a project, such as the names file (Names to replace.txt, in the project's Exegete folder), which Exegete never shows to the assistant;
• Claude Desktop's log of the extension, with every request and answer;
• Codex's session files, with what Codex read by itself.

Start here

Claude Desktop, with one click, is the easiest start, and the one this project suggests for participants' data. Came straight here? First read Where your data goes.

What it costs. Exegete costs nothing; your assistant may. On 6 October 2026 the makers' pages listed, in US dollars, Claude's Free plan, Pro at $20 a month, Max from $100, Team and Enterprise by the seat (Enterprise also by use); and ChatGPT's desktop app for local chats, and Codex's command line, from Plus ($20 a month), with only the desktop app mentioned for Free and Go, "subject to rollout". Plans have usage limits: Claude's reset every five hours, and its paid plans add weekly limits, which longer work can reach; OpenAI's may also be weekly. At a limit you wait, move up a plan or, on a paid plan, pay for extra use. Longer conversations and more tool use count for more, so coding many transcripts uses far more than practice; OpenAI adds that every MCP server "uses more of your limit". LM Studio is free with a local model; its pages recommend 16 GB of memory or more.

Claude Desktop, with one click

1. Get Claude Desktop for macOS or Windows (not Claude in a browser): https://claude.ai/download. Sign in.
2. Download the extension: on the Releases page, take the file whose name starts with exegete- and ends in .mcpb from the newest release that has one under its Assets.
3. Install it: double-click the file, click Install, then Install again to fetch what it needs. Leave its "Tool set" setting as it comes (full): core, the smaller set for local models, cannot create a project (INSTALL.md).

To check, click "+" in a new conversation, then Connectors: Exegete is listed. Then:

• When Claude asks to use a tool, choose "Allow once": it keeps Claude asking.
• If your message box offers no choice between "Chat" and "Cowork", the two are one conversation (Pro and Max plans; Anthropic's page, read on 1 October 2026, says this is reaching accounts gradually). Keep its permission setting on Manual, its default: on Auto, Claude does not ask.
• Before participants' data, go through the five checks: Manual keeps Claude asking; the checks keep your files out of its reach and switch training off.

QualCoder, if you want it. Exegete opens a file for you to read with its coding highlighted; QualCoder shows it too, and does what Exegete does not do yet: download 3.8.2 or 4.0. Exegete works with both, but can tell that QualCoder has a project open only with 3.8.2 (below). 3.8.2 is listed just below 4.0, the release marked "Latest" when this was checked, on 6 October 2026.

ChatGPT's desktop app and Codex (OpenAI)

These can start Exegete too (ChatGPT in a web browser cannot). Codex reads files by itself, so for participants' data this project suggests an assistant with no file access of its own, such as Claude Desktop's chat, until a setting that stops Codex's reads has been tested with Exegete. Experimental: written from OpenAI's documentation, read on 30 September 2026, and not yet tried by this project.

1. Switch training off before participants' data: while it is on, OpenAI may use your conversations to train its models. On a personal plan (Free, Go, Plus or Pro) it is "Improve the model for everyone" in ChatGPT's Settings, Data controls, and Codex's separate "Include environments". Rating a reply (thumbs up or down) can still let OpenAI train on that conversation.
2. Then follow INSTALL.md's steps, by the Terminal route (a few typed commands; Python 3.10 or newer): they make the app ask before every change Exegete makes.
3. Give Codex a folder of its own, as those steps do: it keeps your study's files out of the place Codex works in, but does not stop Codex reading them.

A first session

Practise on text that is not from a participant, such as a page you write. Codex and Claude Code can open files on your computer by themselves, so a real study kept on the same computer is within their reach even while you practise, and Exegete's list of projects tells them where it is. If that matters for a study, you could keep practice projects in a folder of their own, or work on that study with Claude Desktop's chat. A folder of their own keeps practice projects apart but does not put the study out of reach, and what they open goes to their maker, which may train on it while training is on.

Ask the assistant to "Create a new project in Exegete called Practice", then bring in your page: paste it, as in the example, or save it as a Word or text file and give the assistant its place, which keeps its text off the conversation (on a Mac, hold Option, right-click it and choose Copy as Pathname; on Windows 11, Copy as path). With the extension or OpenAI's steps, the project is made, by default, in "Exegete projects", in your home folder (the assistant's answer gives the full path): ask to open a file for reading, or open the project in QualCoder (Project, Open Project), to see your coding in the text. More requests to try.

Two coder names. When it makes the project, the assistant asks for your own coder name, unless you have filled it in Exegete's settings; the project keeps it so that your work and the AI's stay apart (QualCoder, if you use it, records your coding under it). Not using QualCoder? Any name you choose will do. Do not know yours? Say so. Before its first write, it asks for the AI coder name, kept per project, under which Exegete writes what is done through the conversation, unless the project was made while it was filled in Exegete's settings. Both names, and whether you also use QualCoder on this computer, are set in Claude Desktop's Settings, Extensions, Exegete, or elsewhere as INSTALL.md's Environment variables the server reads says.

A project you already have. Ask the assistant to copy it into its folder for projects and to work on the copy. The original is not touched, though an assistant that opens files by itself, such as Codex, can read it from the path you give. A project from before QualCoder 3.8 must first be opened once in 3.8 or newer; without QualCoder, someone who has it and may see the project's data can do this and send you the updated copy. Without QualCoder at all, a new project made in Exegete from the original documents is a way on, without the old project's coding.

One program at a time. If you also use QualCoder, close the project there before the assistant changes it. With QualCoder 3.8.2, an open project is detected and the change refused. In QualCoder 4.0 it cannot be detected, so there only you can make sure; an open 4.0 window shows the changes only once the project is opened again.

Other assistants, and updates

Other assistants. Claude Code (not on Claude's Free plan), LM Studio, other MCP hosts and Claude Desktop set up by hand take the Terminal route (INSTALL.md), whose default tool set, full, can create a project too. There, new projects and copies go to ~/Documents/Exegete projects, which iCloud or OneDrive may sync, unless EXEGETE_WORKSPACE names another folder. Claude Code opens files by itself ("Where your data goes", above), so for participants' data this project suggests Claude Desktop's chat.

Updating. Updates are never installed by themselves and never touch your projects. In the extension, Exegete tells you when a new version is out: at most once a week, and when you ask "Is Exegete up to date?", it fetches a small file from its website, which GitHub hosts. Nothing from your projects is sent; GitHub records your computer's internet address. It tells you about the check first, through Claude, and waits a week before checking; to switch it off, open Claude Desktop's Settings, Extensions, Exegete, "Tell me when a new version is out" (PRIVACY.md says what is sent). Once it has checked, ask "How do I update Exegete?" for the steps on your computer, or install the newer .mcpb the same way. On the Terminal route the check is off unless you set EXEGETE_UPDATE_CHECK=on; INSTALL.md has the steps. Coming from qualcoder-mcp? INSTALL.md says how.

Three commitments

Compatibility with QualCoder. Exegete reads and writes projects in the formats of QualCoder 3.8.2 to 4.0, reading what each supports from the project itself, and does not write to a project in a newer format until it has been checked against it. Where QualCoder has a rule, Exegete follows it, and names any departure with its reason (TOOLS.md). It is not QualCoder, nor an add-on or a remote control for it, and it is not made or endorsed by QualCoder's developers (QualCoder is free software by Colin Curtain and contributors). It is a separate program that reads and writes projects in QualCoder's file format; it contains a small number of routines and values taken from QualCoder so that its results match QualCoder's exactly, and NOTICE lists them, with where each comes from.

Symmetry: the same work in either place, as a commitment. The aim is that the analytic work you can do in QualCoder, you can do from the conversation. It is not yet a fact: today the two differ in both directions. QualCoder 3.8.2 and 4.0 as checked on 10 October 2026, beside Exegete 0.15.0:

Interoperability. Work on a project in QualCoder and from the conversation, one program at a time (above); QualCoder's own export (Project, Export, REFI-QDA Project export) writes the exchange file other packages read. Before you edit transcripts in QualCoder 3.8.2's coding view, read its edit-mode caution.

QualCoder's own MCP server (checked 6 October 2026). QualCoder 4.0's assistant works through an MCP server built into QualCoder. Pull request #1571, by kaixxx, merged on 10 September 2026, added a setting, off by default, that opens it to MCP hosts on the same computer while QualCoder runs, for its open project. It is in QualCoder 4.0, released on 2 October 2026, as "allow external MCP access" in QualCoder's settings, called experimental there. The server calls itself "qualcoder-mcp", Exegete's former name, and the extension QualCoder's source can build is named "qualcoder", a name an earlier Exegete setup may also use (INSTALL.md says how to keep the two apart). QualCoder's AI permission setting (Read-only, for example) governs its own assistant and that server, not Exegete. TOOLS.md gives the commits these facts were read at.

This project welcomes QualCoder's own server, and is ready to cooperate with QualCoder's developers. Exegete has an aim of its own: that you can run a whole project, from its creation to the finished analysis, from the conversation, with QualCoder able to open the same project at any stage, one program at a time. That is a direction, not yet a fact: today QualCoder is still needed for several things (above). On the way there, the commitments above hold: every project keeps QualCoder's format and conventions, so you can work on it in either program; Exegete follows QualCoder's rules and names any departure with its reason; and you work on a project in one program at a time.

For advanced users

Exegete is a Model Context Protocol (MCP) server, which is why more than one assistant can use it. It runs locally over standard input and output, in Python 3.10 or newer, with no online service and no telemetry; its one request of its own is the check for new versions ("Updating", above). What the assistant reads through it goes to the maker of the AI behind it (Where your data goes). It reads a project's SQLite database read-only; each tool that writes opens its own connection, after a backup (by default), and refuses while the project's lock file is active (QualCoder 3.8.2 has the project open, or another conversation using Exegete is writing to it). Each backup copies the whole project folder, with any media stored in it, next to the project; prune_backups clears old ones. Install it with pipx install exegete; INSTALL.md covers every host and setting, TOOLS.md every tool.

That is how much of a model's context a host uses when it sends every tool's definition (name, description and arguments) with each request: measured on Python 3.13, at four characters a token; about five per cent more on 3.10 to 3.12. lifecycle, the earlier name of full in the extension, is still accepted and means full.

Beyond the basics:

• Coding with your approval, the path drawn above: start_coding_session, record_suggestions, review_suggestions, edit_suggestion, update_suggestion_status, apply_codings; codes the assistant proposes: propose_codes, update_proposal, create_proposed_codes.
• Large projects, larger than a model's context: paging cursors, orderings and a character budget on get_coded_segments; cursors and exclude_code_ids (text the codes you name have not reached) on search_files and search_coded_text.
• QualCoder's conventions: per-coder visibility (reads hide what QualCoder hides; a coder argument reads one coder in full), the ##### private part, an AI coder name per project (set_project_ai_coder_name).
• Analysis: find_cooccurring_codes with a character window, query_by_attribute with comparisons, get_case_code_matrix, compare_coders.
• Guarded changes, previewed, then made on a token bound to the preview: merge_codes, merge_category, delete_code, delete_category, pseudonymise_source (these five also say whose work is affected), restore_backup, prune_backups, set_project_coder_name (your own coder name, while nothing is recorded under it).
• Documents in, whole files read: import_documents (a preview, then the import on its token, after a backup; PDF and EPUB need exegete[pdf-epub]) and open_file_for_reading (on the researcher's screen, never in the conversation).
• The brief (provisional), how the assistant is to work with you (read_brief). Claude Code keeps the first 16,384 characters of a tool's description when it finds the tool through its tool search, as it does by default, and the first 4,096 when it loads the tool at the start (2,048 before its version 2.1.296). So the rules that matter come first in each description, within 2,048 characters; the rest reach the model through read_brief or the answers. One description is longer than 16,384 characters, pseudonymise_source's (17,791 on Python 3.13, 18,919 on 3.10 to 3.12): INSTALL.md says how to let it through whole.
• Resources (exegete://...), prompts, and exegete --check-transition for a move from qualcoder-mcp.

Tested. More than 8,000 automated tests run on Windows, macOS and Linux, with Python 3.10 and 3.13, on every change pushed; they test Exegete, not a researcher's use of it (CONTRIBUTING.md).

The repository, each document linked from this page:

github.com/nicotem/exegete
├── README.md          this page
├── PRIVACY.md         where data goes, in full; ethics checklist
├── INSTALL.md         every assistant and setting; problems
├── TOOLS.md           every tool: what it reads and writes
├── AI_CODING_GUIDE.md the coding loop, with examples
├── CHANGELOG.md       every release, in detail
├── CONTRIBUTING.md    reporting, building and testing
├── NOTICE             what comes from QualCoder, and where
└── src/exegete/
    ├── server.py           the tools, resources and brief
    ├── database.py         QualCoder's format, read and written
    ├── sessions.py         the review list of suggestions
    ├── pseudonymise.py     replacing names, keeping the coding
    ├── coder_comparison.py agreement, both coefficients
    ├── memo_privacy.py     the '#####' private-note rule
    └── preview_tokens.py   a preview before larger changes

AI_CODING_GUIDE.md · CHANGELOG.md

What comes next

Plans, not promises: the order may change with what testers report.

• v0.15.1, the safety net: undo everything a session did; when replacing names, choose which mentions to keep; old tools marked as going are retired (TOOLS.md names each)
• v0.15.2, method: how the assistant supports a study's method, reviewed; shorter tool descriptions; the earlier names of the two renamed tools retired
• v0.16, chat-first: more of the analysis shown in the conversation, as tables and graphs (codes that occur together, the code tree, counts by attribute); PDFs that say where their text came from; fuller counts of where names remain after replacing them
• v0.17: a user manual, and codings placed by the passage they quote, even when a quote does not match exactly
• Later: coding images, audio, video and areas of PDF pages; more QualCoder 4.0 conventions; and what testers ask for (file yours)

Disclaimer

This software is provided "as is", without warranty of any kind, express or implied. The authors accept no responsibility or liability for any damage, data loss, or other issues arising from the use of this software. Users are solely responsible for ensuring the integrity and backup of their projects. Always work on copies of your data, not originals.

Licence

Exegete is licensed under the GNU Lesser General Public License, version 3 or (at your option) any later version (LGPL-3.0-or-later), QualCoder's own licence: COPYING.LESSER, with the GNU General Public License it incorporates, legal/GPL-3.0.txt. NOTICE lists the code derived from QualCoder, with its file and lines, and the terms of earlier releases. The conditions apply only to someone who distributes Exegete, in practice a modified version, whose source must then be available under the same licence.

Acknowledgements

• QualCoder by Dr Colin Curtain and Dr Kai Dröge (homepage)
• Model Context Protocol by Anthropic
• Claude Desktop by Anthropic