AI in flowss

flowss uses AI to do the slow parts of making a figure: drafting a diagram from a sentence, rebuilding a whiteboard photo as editable source, converting between notations, repairing a diagram that will not render, extracting quote-backed claims from papers, tutoring you through a course, and running each studio's own named agent, which build, check and refine your work. AI is never required to draw: every engine, template, renderer and export works without it. This page is the map of AI across all nine studios: who can use it, how it is paid for, where it appears, what it is sent, the safeguards around what it changes, what works when AI is unavailable, and what to do when a request is refused. The details of each topic live on their own pages: AI points and limits, Bring your own AI key, The flowss Studio Agent: prompt and quick help, The flowss Studio Agent and its modes and The flowss MCP server.

At a glance

On Windows and Linux, read ⌘ as Ctrl throughout this page.

Who can use AI

AI access depends on your plan. Every AI feature on every studio follows the same rule.

A few rules that apply across plans:

• The allowance is monthly on every plan. A year paid up front still resets on the 1st of each month (UTC); it is not delivered as a year of points at once. Unused monthly points do not carry over.
• Bought points are spent last and never expire. They are used only once the month's allowance is gone. If your plan later ends or moves to Free or Starter, points you bought stay on your balance and hosted AI keeps drawing on them until they run out; you just cannot buy more.
• One feature is a plan feature, not just an AI feature. The Studio's advanced Photo → Engine (clean-up modes, a forced target engine, alternates and render repair) needs Plus or above even when you use your own key. Ordinary photo import works on Starter with your key. See Importing and converting.
• You must be signed in. Hosted AI needs a signed-in account with points, and your own key is checked against your plan, so it is refused while you are signed out.

Plans and prices are compared on Plans and what they include and the pricing page.

Two ways to pay for AI

Every AI request is paid for in exactly one of two ways, and flowss picks for you:

If a key is available, it is always used, and no points are spent even if you have points. To spend points instead, remove the key. Hosted AI is described on AI points and limits; keys on Bring your own AI key.

Note: Calls made through the REST API or the MCP server work slightly differently: a key in your browser or saved to your account is not used for them. A REST call can carry a provider key in its own request header; MCP calls always use hosted points. See The flowss MCP server and The REST API.

Where AI appears

Every studio has AI, and every studio's AI has a name: each studio has its own flowss agent, and most studios reach it through the same two doors: the prompt at the bottom of the screen with its Quick and Agent modes, and the agent's pane for multi-step work.

The Library's cross-studio rewrites are done by the flowss Library Agent, the Teams workspace planner is the flowss Teams Agent, and help on the support page comes from the flowss Support Agent. In flowss Academy, questions about the step you are on go to the flowss Academy Agent (Plus, Ultra and Enterprise at 1 AI point per answer; Starter with your own AI key); it is closed while an exam is open.

The tables below list every AI feature by studio with its hosted price. On your own key the price is always zero points.

Studio

Details: The flowss Studio Agent: prompt and quick help, The flowss Studio Agent and its modes, Importing and converting, Data, timeline, TikZ and the diagram tools.

Study

Details: Study.

Weave

Details: Weave.

Lab

Details: Lab, The FlowScript language.

BPMN

Details: BPMN.

Science, Figure and Sketch

The agent panes of the flowss Weave, Lab, BPMN, Science, Figure and Sketch Agents accept an attached or pasted image; an image adds 2 points to the run.

Details: Science, Figure, Sketch.

Evidence

Details: Evidence, Screening, extraction and synthesis.

Elsewhere in flowss

How an AI request works

There are two kinds of AI request in flowss.

One-shot requests ask the AI once and come back with one result: a Quick change, a photo reading, a conversion, a tutor answer, a set of flashcards. Diagram generation and refinement include one automatic retry when the AI's reply cannot be read, at no extra cost. Most one-shot requests finish in seconds; the largest (a whole course) are allowed up to two minutes.

Agent runs are multi-step. An agent plans, uses tools (reading the figure, editing it, rendering it, checking it), verifies its own work and reports. A run can take up to 300 seconds and a fixed number of steps for its mode, and you can stop it at any time. Agents are described on The flowss Studio Agent and its modes.

What happens to the result depends on the feature:

Tip: Before a large AI change in a studio whose AI replaces the whole document (BPMN in particular), save a version first. See Version history and undo.

Which AI models do the work

Hosted AI runs on models from Anthropic or OpenAI, depending on how the service is configured. flowss chooses the model by the kind of work rather than using one model for everything:

The three classes apply when the work runs on Anthropic's Claude models; when it runs on OpenAI, one model serves all three. The split is why points differ: one point is one ordinary request's worth of work, and deep runs cost more because they reason on the most capable model across many steps.

With your own key, requests go to the provider you chose. The full multi-step agent runs on Anthropic keys, and hosted agent runs are sent to Anthropic whenever the service has an Anthropic key; on the other seven providers an agent run is a single pass without tools: one edit to the open figure, then the platform's own verification and at most one repair. Image features need a provider with a vision model, which DeepSeek does not have. See Bring your own AI key.

What the AI is sent, and what happens to it

Nothing is sent to an AI provider until you do something that asks for AI (send a prompt, press an AI button, run an agent, or paste an image or code into an import), with the one exception of automatic repair. What a request carries depends on the feature: typically your instruction and the open figure's engine and source; for an agent run, up to 40 figures and the last eight messages of the conversation; for a photo, the image; for Evidence, the text of your sources; for a tutor, the course or lesson content it is grounded in. The Studio's exact list is on The flowss Studio Agent: prompt and quick help.

• No training. flowss does not train AI models on your diagrams.
• Not retained. flowss does not keep your prompt or the AI's answer after returning it to you, with one exception: each agent run keeps a short record (the first 280 characters of what you typed, with anything that looks like a credential removed, plus the tools it called, how it ended and the points it cost) for 90 days, so your account page can show it under Agent runs. It never holds the model's reasoning, the conversation, the diagram or your key.
• The provider's policy applies to its leg. On hosted AI the provider is Anthropic or OpenAI; on your own key it is the provider you chose.
• Your own key is never logged, and a key saved to your account is encrypted. See Bring your own AI key.

The full account is on Privacy and your data and the privacy policy.

Automatic repair

When a Studio diagram fails to render because of its source (not because a rendering service is down), flowss waits until you stop typing and then sends the source and the error for a repair, up to two attempts per failure. While it works the canvas shows "Auto-fix attempt 1/2…". The repair is applied only if the figure is still open and unchanged, in the same engine, and ⌘Z undoes it. If both attempts fail, the error box says "Auto-fix tried 2 times and stopped, so it is not still working on this. Edit the source, or use “Explain & fix” to try once more." A failure caused by a rendering service being unavailable is never sent for repair, and BPMN diagrams are never auto-repaired.

To switch it off, open Studio Settings (⌘,) and use Turn auto-repair off under Repair broken diagrams automatically. The setting is kept in this browser. Each repair attempt costs 1 point on hosted AI.

Safeguards around AI changes

AI output in flowss is a proposal you can inspect and take back, and several mechanisms check it before and after it lands:

• Text the AI did not get from you is material, not instructions. Words inside a fetched file, a pasted document, an image, a renderer's error message or a diagram that say "ignore your instructions" are diagrammed if they are meaningful and otherwise ignored.
• Approvals. An agent cannot, without your approval, delete a figure that existed when the run began, overwrite a figure other than the one you have open, or remove three or more elements from a board in one run. It proposes the change instead, and the pane shows Approve / Reject.
• Verification. When an agent finishes, every figure it touched is checked: does it parse, is it complete, do its references resolve, does the engine's own analysis flag anything blocking. A mode that can edit is sent back to fix what is blocking, at most twice, and findings appear in the run log.
• Illustrative values are marked. When an agent has to invent a value to show the shape of a figure, it lists it under "Illustrative values — not data" and marks the figure itself, so the value cannot pass as data.
• Your data stays verbatim. Dataset charts are built by deterministic generators from your values; a model never retypes a number.
• Grounded answers are checked mechanically. Evidence keeps only claims backed by a verbatim quote from the source, and Weave's Deep analysis strikes any sentence naming something that is not on your canvas, or rejects the whole reply.
• Your edits win. If you edit while an agent works, your version is kept and the agent's is offered beside it.

What works without AI

AI is an accelerator, not a dependency. With no AI at all (on Free, or when hosted AI is unavailable), all of this keeps working:

When AI is unavailable

Several different things can stop an AI request, and each says what it is:

Every refusal is explained, with its fix, on AI points and limits and Bring your own AI key.

The flowss Support Agent on the support page

The support page matches your question to a help topic in your browser first. When the flowss Support Agent is available, you can also tick "Use the flowss Support Agent to select a help topic. I agree to send this question to Anthropic when I choose Find help." and press Find help:

• You must be signed in ("Sign in to use the flowss Support Agent. Help topics and human support remain available."), and you consent each time.
• The question is up to 500 characters. Keep it general: "Do not include passwords, API keys, payment details or private project content."
• It spends no AI points and never uses your own key. It is limited to 20 questions a day per account and 60 a day per network address; past that you see "The flowss Support Agent has reached its daily limit. Use the help topics or contact a person."
• The flowss Support Agent only picks a topic, and can pick the wrong one; it has no tools to change your account. Questions are not saved as a conversation. Contact a person is always offered.

When it is off, the page says "The flowss Support Agent is currently off. Topic matching runs in your browser and does not send your question to a model."

AI from outside flowss

• The REST API. POST /api/v1/diagram generates a diagram from a prompt with your flowss API key, charged to your account's points. See The REST API.
• The MCP server. Connect Claude, Cursor, VS Code or any MCP client to https://www.flowss.ai/api/mcp and your assistant can list engines and templates, verify and critique diagrams, and, with your API key, generate, convert, render and diff diagrams. See The flowss MCP server.

Tips

• Use Quick for one clear change and Agent for anything that needs several steps or several figures. A Quick change costs 1 point; an agent run costs 8.
• Keep an eye on your avatar. It gains an amber ring when your balance runs low and a red ring at zero, and the AI points row in its menu takes you to where you can top up or add a key.
• If you already pay an AI provider, Starter plus your own key is the cheapest way to AI in every studio, except the advanced Photo → Engine.
• Deep specialist runs (47 points) are worth it for structural judgement on finished work. For routine edits, a standard mode does the job at 8.
• Before letting AI rewrite something you care about, save a version. Undo covers most AI changes, but a saved version covers all of them.
• If a diagram keeps failing to render and you do not want AI touching it, switch automatic repair off in Studio Settings.

Limits and known constraints

• Free has no AI at all: no hosted points, and no own key.
• Hosted points are a fixed monthly allowance. They are never unlimited, and unused monthly points do not roll over.
• Your own key is unmetered but not unlimited: 1,000 AI requests a day per account, at most 300 of them agent runs.
• Weave's Generate on a mind-map branch is limited to 60 requests a day per network address, whichever key pays.
• TikZ / LaTeX export on hosted AI has its own daily cap by plan (5 a day on Free and Starter, 25 on Plus, 100 on Ultra, 200 on Enterprise), on top of its 6-point price. On your own key it is not capped beyond the fair-use ceiling.
• A typed request is limited to 6,000 characters in the Studio prompt and agent; the agent sees at most 40 figures and the last eight messages.
• Hosted AI is not available while the Studio shows AI is off here, and in that state the Studio prompt is disabled even for your own key; use the agent pane or another AI door instead.
• The full multi-step agent needs hosted AI or an Anthropic key; other providers get a single pass.
• DeepSeek keys cannot read images: photo, sketch and handwriting features need another provider.
• The advanced Photo → Engine needs Plus or above, even with your own key.
• AI answers can be wrong. The checks above catch structural problems, not every factual one; read what the AI produced before you publish it.

Troubleshooting

More messages and fixes: AI points and limits, Bring your own AI key, Troubleshooting and FAQ.

Related pages

• AI points and limits
• Bring your own AI key
• The flowss MCP server
• The flowss Studio Agent: prompt and quick help
• The flowss Studio Agent and its modes
• Plans and what they include
• Privacy and your data
• The REST API