DunneWrite · AI and ingestion
AI providers and costs
Which parts of DunneWrite use an AI, how to connect Claude, Gemini or a local Ollama model, and who pays for what.
Most of DunneWrite never touches an AI. Writing, editing, History, searching, the maps, the graph, the timeline, and importing and exporting stories are all free and work with no key.
A small number of features send your text to an AI to read it. For those you bring your own: an API key for Claude or Gemini, or a model running on your own computer through Ollama. DunneWrite has no AI account of its own and sells no AI credit.
What sends text to an AI#
Seven buttons read your material with an AI. Each is charged to your own provider account (or runs on your own machine, with Ollama).
| What you press | Where | What it does |
|---|---|---|
| Ingest a file — see the cost first… | Ingestion Review | Reads your notes and pulls out characters, places and props. |
| Re-extract with Bible — see the cost first | Ingestion Review | Reads the same material again, knowing what your bible already holds. |
| Consolidate Duplicates | Ingestion Review | Groups names that look like the same person or thing and asks a model to judge them. |
| Infer Character Depth | Ingestion Review | Reasons across all your material to fill in what no single passage states. |
| Derive map / Re-derive | Globe | Works out where your places sit in relation to each other. |
| Propose alignments | Terrain Import | Matches names on an imported map to the places in your bible. |
| Inaccurate — rescan with another model | Quotes to check | Asks a different model what a quoted passage actually says. |
Chapter chat in the Writer's Room is also a conversation with an AI. It runs on whichever provider you have chosen, Claude, Gemini or a local Ollama model, and asks before it sends anything.
Every one asks first#
Nothing is sent until you say so. Each of these opens a confirmation that tells you who is being billed, which model will run, and how much text it is about to send: an exact amount where the app can count it, and an honest "at least" or "no more than" where it can only estimate. Cancel is the default: pressing Enter or closing the box sends nothing.
While a run is going you can watch its progress and cancel it. Cancelling stops the paid work.
DunneWrite never shows prices#
The confirmation tells you the amount of work, never a price. That is deliberate: a price list built into an app goes stale, and your provider's rates are theirs to change. Check your provider's own pricing page for current rates.
As a very rough guide only, reading a large pile of notes has cost in the region of US$20–40 in provider charges, and much less for a small one. That money goes to your provider, not to DunneWrite, and it varies with the provider, the models you choose and your material.
A subscription is not a key#
Paying for a consumer chat subscription from an AI company does not pay for this. An API key is free to create, but the credit it spends is bought separately on the provider's developer website. A key with no credit behind it is rejected with a message about billing, not about the key.
Connecting a provider#
Open Settings at the bottom of the left-hand menu.
- Under AI keys, paste your Claude or Gemini key into its box and press Save. You can save both.
- Press Check this key (free). It tells you straight away whether the key works, and it spends nothing.
- Under Which one reads your notes, choose the provider that should do the reading. You are only offered a provider the app has actually checked.
Your keys are stored encrypted on this computer and are sent only to the provider they belong to. A key never travels inside a story bundle.
Choosing Gemini models#
If you choose Gemini, you are asked to pick three models before it can run:
- First pass (cheap) runs on every chunk of text, so it decides most of the bill.
- Main reading does the real work of pulling out characters, places and props.
- Second look (only when unsure) re-reads only the chunks the main pass was unsure about.
The list comes from your own key, so it is always current, and choosing is free. If you have no strong opinion, start with lower-cost models and work up.
Gemini runs these jobs through Google's batch queue, which only promises to finish within 24 hours. In practice it is usually far quicker. You can close the page and come back: the run keeps going, and the progress banner finds it again.
Running locally with Ollama#
If you can run a model on your own computer with Ollama, choose it under Which one reads your notes, and use Check the connection (free) to confirm DunneWrite can reach it. Reading your notes, merging duplicate names, filling gaps in the bible and placing your world on the map all run on it.
- Your text never leaves your computer.
- Nothing is billed, because there is no account. It is not free in every sense: it uses your own hardware and electricity, and it may take longer.
- How well a local model does these jobs depends on the model you pulled and on your machine. DunneWrite checks that each job will fit before it starts. Try it on a few chapters before committing a whole manuscript.
The Writer's Room's Chapter chat runs on Ollama too, so a whole conversation about a chapter can stay on your computer.
Choosing a provider for one run#
The cost screen for an ingest also lets you choose who reads this run, and on which models, separately from Settings. The choice applies to that run only. See Ingesting your notes for what happens when two different readers read the same notes.
What needs no key#
- Opening a bundle someone sent you.
- The chapter brief: DunneWrite writes it, and you use it with whichever AI you like, through your own account. See The chapter brief loop.
- Everything else not listed in the table above.
Related#
- Tutorial: Connect an AI provider
- Known limitations
Something unclear or out of date? Open a support ticket.