Appearance
The assistant
The assistant reads the workspace and works on it for whoever asks. It runs inside Notesflow — the searching, reading and writing are Notesflow's, checked against what that person may do — and the model that does the thinking is one you choose.
Where its writing goes
When somebody asks the assistant something, what it reads to answer — pages, search results, the page they have open — is sent to the model they are using. Whoever decides which models are here decides where this workspace's writing may go. That is why adding a model for the workspace takes its own capability, manage_assistant, and why members cannot add their own unless the workspace says they may.
There are three kinds of model:
| Who pays the provider | Who decides | |
|---|---|---|
| The workspace's models | The workspace, with its own key | Somebody with manage_assistant |
| A member's own model | The member | The member, if the policy allows it |
| The platform's model | Nobody here: it is charged in credits | Allowed or refused by the workspace |
Adding a model
In Settings → Workspace → Assistant, choose a provider, the model's id as that provider names it, and a key. Try and save sends one short question that can only be answered by calling a tool, and saves the model only if it answers that way. A wrong key, a wrong model name, an address that is not a model at all, and a model that cannot call tools are each refused there, with the provider's own words.
The named providers — OpenAI, Anthropic, Azure OpenAI, Google Gemini, Mistral, Groq and OpenRouter — answer at the provider's own address, which nobody can change. Another OpenAI-compatible endpoint takes an address: your own vLLM, SGLang, Ollama or LM Studio server, or a gateway. It must be https, and may not be on a private network; see below. An Azure address must be your resource's: https://<resource>.openai.azure.com/openai/deployments/<deployment>.
The key is kept sealed with the deployment's key and is never shown again — the console says only that there is one. Rotating NOTESFLOW_SECRET makes every saved key unreadable; each model then says so when tried, and has to be given its key again. A model that stops answering is never replaced by another without the person choosing it: the run stops and says why.
The policy
| Setting | What it decides |
|---|---|
| People here may use the assistant | Off turns it off for everybody here. Conversations and models are kept. |
| Allow the platform's model | Whether anybody here may use the deployment's own model. A company whose contracts forbid a third party reading its writing turns this off. |
| Members' own models | Not allowed (the default), named providers only, or named providers and addresses of their own. Narrowing it keeps what members added and stops it being used. |
| Most a member may let it write | The ceiling on writing: without asking, ask me first, or never. |
| Most a member may let it remove | Ask me first or never. Removing is never done unasked. |
| Default model | What a new conversation uses when the person has not chosen. |
| Keep conversations for | How long a conversation is kept, in days. |
Every change to the policy and to the workspace's models is in the audit log, with what it was before. A model's key never is.
What it may do without asking
Each tool the assistant has is one of four kinds. Reading and suggesting are always allowed; writing (changing a page's text, adding to one, making, renaming or moving a page, setting a property) and removing (moving a page to the trash) are what the ceilings decide:
| Ceiling | Writing | Removing |
|---|---|---|
| Without asking | a member may let it write unasked, or ask, or never | — |
| Ask me first | it asks every time, or never | it asks every time, or never |
| Never | the tools are not offered to it | the tools are not offered to it |
An answer to may it do this? comes from the person's own signed-in session, never from anything in the conversation, so nothing a page says can approve a write. Every write goes through the same routes and the same three-way merge as a person's, and is recorded as the person's, with the assistant's reason on every op, so it is in the page's history, revertible, and marked in the timeline.
What a member chooses
In Settings → Account → Assistant, each person chooses their model and what the assistant may do without asking them, up to the workspace's ceiling. A ceiling lowered later caps what they chose; raised again, it gives it back.
The assistant can never do more than the person asking could. Somebody who may only read a page has an assistant that may only read it, whatever the settings say.
The platform's model
A deployment may run a model of its own and sell it in credits. The operator sets it in /admin → Assistant — the address, the model, its key, the protocol, what a credit is, and what a credit past a package's costs — and it is tried with a tool call before it is saved. What is set there wins over the environment, which may also set it, with NOTESFLOW_ASSISTANT_URL (a Chat Completions address, up to and including /v1), NOTESFLOW_ASSISTANT_MODEL (by default Qwen/Qwen3.8-27B), NOTESFLOW_ASSISTANT_KEY, NOTESFLOW_ASSISTANT_PROTOCOL (chat, or messages for Anthropic's) and NOTESFLOW_ASSISTANT_LABEL (what workspaces are told it is called); Use the environment's in the console goes back to it. With neither, there is no platform model and the assistant runs only on models workspaces bring. The console's key is sealed with NOTESFLOW_SECRET; rotate that and the model stops, saying why, until the key is given again.
The platform's model is the operator's, so it may be on the deployment's own network. A workspace's models may not.
Credits
On the platform's model, what a run used is counted in credits: so many tokens read, or so many written, make one. The operator sets the rule in /admin → Assistant — by default 10,000 tokens read or 2,500 written — and a package's Assistant credits a month is what it includes. Every model's runs are counted, so an admin sees them in the console; only the platform's are charged.
| The operator sells credits past a package's | When a workspace has used its month's |
|---|---|
| No (no price set) | Its next question is refused until the 1st. One already being answered may finish its answer — nobody knows an answer's size before it is written — and stops before another. |
| Yes, at a price per credit | It goes on. The whole credits used past the package's that month are one line on the next monthly invoice. |
Private networks
A model's address is typed in a workspace and connected to from inside the deployment. So it must be https, and an address on a private network — loopback, 10.0.0.0/8, 192.168.0.0/16, link-local and the cloud metadata address among them, written any way — is refused when the model is added and again whenever it is used, after the name is resolved. Redirects are not followed. A deployment whose every workspace is its own can allow private addresses with NOTESFLOW_OUTBOUND_ALLOW_PRIVATE=1; webhooks follow the same rule.
In the package
Two features: Assistant (the assistant on models the workspace brings) and Assistant on our model (the platform's model). When the first lapses, nobody can ask the assistant and everything is kept; when the second lapses, the platform's model cannot be asked and the workspace's own keep working.
The assistant's index
In Settings → Workspace → Assistant → The assistant's index, someone who manages the assistant can have it read the workspace in the background. It turns each page into short notes: decisions, commitments, definitions, open questions, and claims the team relies on. Each note cites the blocks it came from and when they were written. When the assistant is asked "what did we decide about…", its recall tool answers from these notes and cites them.
- What is read. Every page except pages kept away from the assistant (
assistant: false, and everything under them), templates, and archived or trashed pages. A page is read once nobody has changed it for ten minutes, and read again only after it changes. Notes about a page that is kept away are deleted at the next read-through. - With which model. One of the workspace's own models, or the platform's. Never a member's own model, because nobody asked for this reading.
- What it costs. Before switching it on, the console shows how many tokens reading everything once would take, and how many background credits that is on the platform's model. Background credits are a monthly limit of their own (Background credits a month). Background reading never uses the credits people's questions use. When they run out, the reader stops, says so here, and starts again next month. Use on the workspace's own models is counted but not charged.
- Never secret. On any page, What the assistant knows about this page (in the palette) shows what was noted, and what the page said before it changed. Anybody who can read the page can mark a note This is wrong; it is not used again, and the next reading is told about the correction.
- Who sees what. Notes are kept per page and put together only when someone asks, from the pages that person may read. Nothing noted from a page reaches someone who cannot open it.
- A page that gives instructions. The model is told the page is data, and every note it returns must cite blocks of that same page. So a page that tries to steer the model can at most produce notes about itself, seen by its own readers.
It is a package feature, Assistant in the background. When it lapses, the reader stops. What it noted is kept and still used, and Suggestions and reminders go on.
Writing up a draft
On any page you may edit, Write this page up with the assistant (in the palette) treats the page as a draft. The assistant works out where the writing belongs: a new page, perhaps under a related one, or the end of a page you started. It writes it in the workspace's terms. Your draft stays as you wrote it, and gains a link to what it became. Anything added to a page somebody else started is a suggested edit for them to accept, never a change.
How much it does on its own is yours to choose, in Settings → Assistant → Writing up a draft:
- Ask me first. It always shows you the plan, and writes only when you say so.
- Ask when unsure (the default). It goes ahead when everything is a new page or a page you started and nothing is left to ask. Otherwise it shows you the plan.
- Go ahead. It goes ahead whenever it has no question. A question it asks gets the recommended answer after four hours if you have not answered.
A workspace that has the assistant ask before writing offers the first two only.
You are told in your inbox when it has written or has a question. The page's header shows the draft's card, with the plan, the questions, and Undo all.
How we write here, and How I write
A workspace's way of writing — tone, terms, headings, where things go — lives in an ordinary page. Somebody who manages the assistant opens it and chooses Use this page as How we write here in the palette. From then on, the assistant follows it when it writes up anybody's draft. Anybody may do the same with a page of their own, Use this page as How I write, which the assistant follows for their drafts only, after the workspace's.
The assistant takes instructions from these two pages and no others, so it reads each one only while it is still theirs:
- How we write here is followed only while everybody who has changed it since it was chosen manages the assistant. If somebody else edits it, it is not followed until somebody who manages the assistant reads the change and chooses the page again.
- How I write is followed only while nobody but you has changed it.
Settings → Assistant says whether each is followed, and who changed it if not. When a draft's writing follows a habit the guide does not mention, the draft's card offers it under Not yet in "How we write here". Suggest adding it makes a suggested edit on that page, for the people who write it to accept or not.