Pangur docsModels, assistants and who pays
Some of Pangur's work is done by a language, vision or speech model, and each such action counts against someone's allowance. This page lists which actions use a model, whose allowance they use, which limits apply, and why work done by your own assistant over MCP costs nothing on your Pangur plan.
Some of Pangur's work is done by a model: writing an answer, summarising a book, reading a letter book into its people and places, reading a manuscript leaf, or transcribing a recording. Pangur sends that work to a model chosen by its administrators, and each such action counts against an allowance belonging to the person who started it. Most other work, including searching, importing, reading files into the index and exploring an analysis, uses no model at all. When you connect your own assistant, such as Claude or ChatGPT, the assistant's model does the reading and writing, and none of that counts against your Pangur plan.
Pangur's models and your assistant's model
Two different models can be involved in your research, and they are paid for differently.
Pangur's models run when you act in the web client, the Mac or iPhone app, or the command line. Pangur's administrators choose which model handles each task (answers, analysis, page reading, speech), and you do not choose. An Ask answer names its model under the question, after "Answered by".
Your assistant's model runs when an assistant connected to Pangur over MCP uses Pangur's tools. The tools find and return the material, and the assistant's own model reads it and writes the result. Pangur does not run one of its own models for an MCP session, so the cost belongs to whoever pays for the assistant.
Note that the command line counts as one of your own surfaces rather than as an assistant. pangur ask and pangur manuscript transcribe run Pangur's models and count against your allowances, just as the web client does.
Actions that run Pangur's model
The table lists the actions that use one of Pangur's models and whose allowance each one uses. "You" means the person who starts the action.
| Action | Model | Whose allowance |
|---|---|---|
| An Ask answer, including Ask beside the reader | language model | yours: one Ask question, plus model use today |
| A summary of a whole work, or an answer that reads whole works | language model | yours: model use today |
| Expand — full write-up of an answer, or a full-depth summary | long-form model | yours: one full write-up, plus long-form writing today |
| An analysis reading of a source or a corpus | language model | yours: model use today |
| Reading one manuscript leaf, from Ask or Read this leaf with the vision model | vision model | yours: one manuscript reading |
| Transcribe with AI on a manuscript leaf (two passes) | vision model | yours: one manuscript transcription |
| Transcribing a recording | speech model | yours: transcription minutes this month |
| Library map entries | language model | the project owner's model use today |
| A corpus digest, written a few minutes after a corpus changes | language model | the project owner's model use today |
| Testing a concept you suggest in a corpus | language model | the project owner's model use today |
Some smaller steps also use the language model and count towards your model use for the day: naming an Ask conversation, interpreting a follow-up question, an optional pass that extracts references from messy pasted notes, Write's Read more widely and summaries, and composing or generating an exhibition.
The library map, corpus digests and concept tests run in the background on behalf of the project as a whole, which is why they are charged to the project's owner rather than to the person whose action prompted them. When the owner's allowance for the day is spent, library map entries and corpus digests wait for the next day's allowance instead of failing.
Actions that use no model
None of the following spends model allowance:
- Searching your library and outside catalogues, adding works, and finding PDFs.
- Reading files into the index, including OCR of scans and the embeddings used for search, which run on Pangur's own server.
- Reading in the readers and the reading room, and Find in text.
- Line detection on a manuscript leaf (Detect lines), which uses a line-detection program on Pangur's server rather than a language model.
- Previewing how analysis will divide a source, and the estimate shown before a reading.
- Exploring a corpus that has been read, through its records, timeline, map, network, trends and Words and phrases.
- The starter questions on an empty Ask page.
- Ask when no model is switched on, which shows matching passages only.
pangur ask --materials, which returns passages without an answer.
The limits that apply
Pangur applies two kinds of limit. Plan allowances depend on your plan. Daily guards apply on every plan and exist to stop a runaway loop from spending without limit. The figures below are the defaults, which administrators can change, and the Plans, limits and usage page is the reference.
| Limit | Default | Resets |
|---|---|---|
| Ask questions, Free plan | 25 | each question counts for 30 days |
| Ask questions, Researcher and Institution | 200 | each question counts for 24 hours |
| Ask questions per hour, every plan | 20 | after the hour |
| Model use today | 400,000 tokens | midnight UTC |
| Long-form writing today | 800,000 tokens | midnight UTC |
| Full write-ups today | 4 | midnight UTC |
| Manuscript readings | 40 | 24 hours after the first |
| Manuscript transcriptions | 20 | 24 hours after the first |
| Recording transcription, Free plan | 60 minutes | each calendar month (UTC) |
| Recording transcription, Researcher and Institution | 1,200 minutes | each calendar month (UTC) |
The Usage page shows how much of each plan limit and daily guard you have used and when it resets, recording minutes included. The transcript pane of a recording also says how many minutes are left this month.
When an action would pass a limit, Pangur refuses it with a sentence that names the limit, for example "You have used this library's model allowance for today. It resets at midnight UTC." A long job that runs out part-way keeps what it has done, so an analysis reading that runs out of allowance is saved as partial, with a note of how many batches were read.
Shared projects
In an organisation project, model use stays with the person who acts. Each member's Ask questions and model use count against their own plan and daily guards, even though the library is shared. The exceptions are the background jobs listed above (the library map, corpus digests and concept tests), which are charged to the project's owner. Storage in an organisation project counts against the organisation's pooled storage.
Deployments without a model
A deployment may run without a hosted model, or with some tasks switched off. Pangur then does what it can without one.
- The home page's search offers Find passages in place of Ask, and a question returns the matching passages with a note that the model is not switched on.
- Analysis offers no model or cost estimate, and a reading fails because no model tier is configured.
- The library map is not written.
- Expand — full write-up is not offered unless long-form writing is switched on.
- The transcription buttons say that reading with AI is not switched on for this server.
Searching, reading, highlighting, transcribing by hand and writing all work as usual.
Work done by your assistant
When an assistant connected over MCP asks Pangur for something that would need a model, Pangur's tools take a route that needs none and hand the rest to the assistant.
ask_libraryreturns the ranked passages and the library map, and the assistant writes the answer. If the assistant then records its answer in one of your Ask conversations, the turn is labelled with the assistant's name and does not count as an Ask question.summarise_itemandcompose_exhibitionreturn instructions and materials for the assistant to work from. A summary with checked quotations made by Pangur itself is made in the Pangur app, where it counts against your plan.transcribe_manuscript_leafdetects the lines and returns them with their image regions, and the assistant reads each line and saves the transcription.analysis_materialsandsubmit_analysislet the assistant analyse sources with its own model and submit the result for Pangur to check and store.
Over MCP, Pangur still runs its search, its local embeddings, line detection and lookups in outside catalogues, and none of these spend model allowance. What assistants can and cannot do is set out in Assistants and Pangur.
What is sent to a model
When an action uses one of Pangur's models, Pangur sends the model only what the task needs. For an answer that is the question, the recent conversation, the library map and up to 12 passages from your library. For a summary or a library map entry it is the whole text of the work, for analysis the source's text unit by unit, for a manuscript leaf its page images, and for a recording its audio. Apart from the background jobs described above, nothing is sent to a model until you take an action that needs one. The outside services, and what each one receives, are listed in Security and privacy.
How Pangur reads your sources
Pangur extracts and indexes the text of your files in the background, and Ask, analysis and passage search work from that text. This page explains which formats it reads, how to tell whether a source has been read, what the library map is, and what it means when a work has no full text.
Find and add a work
The search bar finds a book, article or early printed edition in your library and in outside catalogues at the same time, and saves it as an item. This page explains how the search works, how Pangur handles identifiers and duplicates, and how to add a reference by hand.