How the AI works

You bring the key. You see what is sent. You accept every change.

Quarivell has AI. Quarivell is not an AI tool. An accomplice, not an author — it works for the work, never in place of the one who writes it.

Five rules that don't break

These are engineering invariants — written in the code, not only the intent. They are not a privacy mode you switch on; they are how the app behaves.

  1. 01

    The AI acts only when you ask

    No background requests, no hover suggestions, no preloading. No byte of your work reaches a model without an action of yours.

  2. 02

    You always see what is sent

    The send preview shows what leaves your machine before the request — document by document, with token counts, and the option to remove anything.

  3. 03

    Your keys live only in the keychain

    Never in a config file, an environment variable, a log, or on screen. Quarivell asks the operating system to use the key; it never reads it.

  4. 04

    We never proxy your AI traffic

    You talk to the provider you chose, at cost. No middle server of ours, no collection, no training of models on your text.

  5. 05

    Local and private by default

    If a model runs on your machine, that is the recommended path. Outbound connections are limited to a reviewed allowlist.

The AI, level by level

Three levels, around your writing.

From the sentence to the whole work to your sources — the words stay yours at every level.

01

Beside the sentence

Rewrite a line in your voice and accept it as a diff — your original is never replaced without your yes. Ask about what you are writing, scoped to a selection, a chapter, or the whole project, with answers cited by source pills. An offline dictionary and thesaurus, and multilingual spellcheck, round it out.

02

Across the work

The AI reads your prose and proposes cards for the characters, places, and themes it finds — you review and confirm each. It proposes typed relations between them, draws a force graph, and drafts a synopsis for a scene or a whole folder. Run any of it over a tree as a durable batch job that survives a restart.

03

Through your sources

Read and annotate PDF and EPUB beside the draft, turn a source into a citation in one click, and ask a question about your own materials to get an answer cited to the file and the page — never an answer from nowhere.

Your key, or your machine.

Bring your own key

Connect Anthropic, OpenAI, or Google with your own keys. The AI is billed by the provider directly, at cost, with no middleman. Switch provider or model whenever you want.

Or your own machine

Run open models locally through Ollama, LM Studio, llama.cpp, Jan, or vLLM. When one is detected, it becomes the default — the recommended path for sensitive work, because nothing leaves the machine.

Cost you can read

A live cost line, spend caps per provider and per day, and a usage dashboard with a simulator that shows what the same work would have cost on a paid model. The currency is yours to set.

No silent switches

Prompt caching reuses the context prefix to save tokens across a conversation, and the saving is measured and shown. There is no automatic failover between providers — your choice of model holds.

You see what is sent

Nothing leaves without you.

Before any request, open the bundle that would go out — document by document, with token counts and source pages — and drag to remove anything you don't want to send. A local pass first flags secrets, like an API key, before they ever leave.

Send preview ≈ 0.4¢ · claude

↑ Sent to provider

Selected passage (1 paragraph) + your instruction: “tighten this without losing the rhythm.”

⦸ Held back on your machine

the rest of the manuscript · your name · file paths

Open to agents

Your project, open to the agents you trust.

Connect Claude Desktop, Cursor, Zed, or the Claude Code CLI to your project — without copying or uploading the manuscript anywhere. Reading is always available; writing is off by default, and you turn it on per project. Every change an agent makes is marked and restorable.

Write toolSafeguard
Create documentSafe path, never overwrites, atomic write
Append textMarkdown only, bounded per call, atomic
Edit by exact matchExact text required; a snapshot is saved before any change, restorable
Rewrite, metadata, synopsesThrough the typed core stores; every replaced value logged first

Private by default.

Your work never leaves your disk without you sending it — and never trains a model. For sensitive work, run a local model and keep everything on the machine. That path is first-class, not a degraded mode.

Getting a key, step by step

If you have never used an API key, this can read like a wall. It is three steps, and we walk you through each one.

  1. Step 1

    Create an account with a provider you trust.

  2. Step 2

    Copy the key they give you — once.

  3. Step 3

    Paste it into Quarivell. It stays on your machine.

On the way

Designed, and being built.

Honesty about state is part of the house voice. These are designed and not yet here — we don't sell them as present.