SchemAgentic
AI
AI appears in five places in SchemAgentic, and they are not the same thing wearing five hats. This page is what each one is, what it costs, and what leaves your machine.
Where the AI is
| Where | What it does | Runs |
|---|---|---|
| The map chat | Builds the schema: nodes, content, agents, services and connections, from a description — and the scene behind it. | When you ask |
| The agent chat | A conversation with one agent about its own work. It can propose changes to itself; you press. | When you ask |
| Agents | Standing instructions carried out on a schedule, on our servers, with the document closed. | On their own |
| Drafting a service | Describe an API and a service definition is drafted for you to check before it is saved. | When you ask |
| A scene background | Describe a background and a gentle scene is designed to sit behind the schema. | When you ask |
There is a sixth, in the other direction: the MCP server lets an external AI
agent read your maps (in depth too: one node at a time, the tasks across your maps, or a
whole map as Markdown, OPML, CSV or JSON), add to them and edit their nodes (rename,
restyle and move them; delete or replace only with your approval in the app), add content to them (text, links,
pictures, videos, check boxes, lists, tables, charts, places, calendar cards and emoji, or text at the end of a
note), link their nodes and restyle their lines,
format them (fonts, emphasis, shapes, numbering and grouping; the words
never change), tick their check boxes, fold their branches or focus one,
copy branches, refine their
style (spacing, the style sheet's defaults or the theme's, the glow and flowing lines) and, from the next
SchemAgentic update, ask you to approve deleting or replacing parts of them or
taking back an earlier request (what they ask happens only once you approve it in the map's Agent Requests) and
leave you messages, and, from the next SchemAgentic update,
build new maps from text, rename, move and copy their files, make and rename
folders, link one map to another, and ask you to approve moving a map to the trash; and, from the next SchemAgentic update, read who is in the maps shared with you, their conversation, recent changes and version history, and ask you to approve restoring an earlier version; and, from the next SchemAgentic update, export them as FreeMind, a web page, a JSON tree or CSV, list their dates as a calendar, read their tables, lists and links and a small picture of each, fill in their tables (and ask you to approve changing them), and save an export as a file on your device; and, from the next SchemAgentic update, find your maps by their words (word search, not semantic search: no AI model on our side), read and add the items inside notes, add drawings made from SVG, reorder content, set a saved or new scene and turn a title into a short list, with review mode if you want to approve every change first — Claude
Code, Cursor, VS Code, or any MCP client that can send an access token; and claude.ai, Claude Desktop's connectors
and ChatGPT, which sign in with your SchemAgentic account instead of a token (signing
in is being switched on in stages; until then Claude Desktop connects with a token through the
mcp-remote bridge). You can disconnect a connected app, or revoke a token, at any time, and either can
be read-only. From the next SchemAgentic update, a request is applied at once while the map is open.
No agent deletes a file or a folder: a map leaves only through the trash, when you approve it, and nothing is
ever written over a file or folder that is there. The recipes
are under Clients. With its switch on, the app publishes each map's node titles and notes for
those agents, and from the next SchemAgentic update a fuller copy with their
colours, check boxes, links and cross-links; the server calls no model and uses none of your allowance. All
sixty-two tools are described there; many of them are switched on on the server in
stages, and until one is, it answers that it is not available at the moment.
Scene backgrounds
A schema can wear a soft vector scene behind it — aurora curtains, a slow drift through stars, sun rays under water, clouds and rain, mountains, rivers and lakes, the seasons, gentle abstract shapes. Open a node's Style sheet (the gear), go to Theme, and under Canvas turn on Scene background. 38 scenes are built in, in 7 groups (sky and space, clouds and weather, seasons, rivers and water, under the sea, mountains and desert, abstract); most move slowly, many in depth — the nearer layers a little faster than the farther ones — and some are still. None of that uses AI.
- Animate holds a moving scene still when it is off. Exports and thumbnails always show the still frame.
- Strength is how far the scene stands out from the canvas colour; Speed is how fast a moving one moves.
- 3D depth is how far the scene's layers follow you as you pan and zoom, the nearer ones more than the farther ones; at 0 the scene stays put. Exports always show it at rest.
- Scenes are made to stay behind the work: slow, sparse, quieter in the middle, in colours that follow the theme - low in contrast on a dark theme, daylight colours (a sky, land and water) on a light one; Northern Lights alone fills a dark sky with vivid aurora. A scene replaces the gradient background, is saved with the schema and is undone like any other style change.
Keeping a scene behind a schema needs a subscription of your own, which is why the Scene background switch carries a Premium tag. Without one you can still try any scene in the style sheet, built-in or your own: it is taken back when you close the sheet or tap outside it, and the app tells you why. A scene that comes with a schema stays: one shared with you, or made by an agent over MCP, shows its scene whatever your plan.
Generate with AI
Above the scene cards, Generate with AI designs a new scene from a description. Say what you want, choose Animated or still, and press Generate. A generated scene can use what the built-in ones use — layers at different depths, mountain ranges, sea swell, palms, rivers with a current, falling leaves, raindrop rings — and is held to the same limits: the built-in scenes' colour budget, slow motion, no text and nothing that could pass for a node or a connection. It is kept under Yours, shown first among the scene cards, with your account, so it is on every device you sign in on (the 24 most recent); the trash on each of your own scenes deletes it after asking, and schemas that already use it keep it. It needs a subscription of your own and uses your own allowance.
Asking the map chat
Ask the map chat for a background and it puts the built-in scene that fits best behind the schema, then asks: Keep it, Try another (its next choice, the reply showing which one you are on, such as 2/4), or Create a custom one, which designs a new scene from what you asked for. Ask for something specific that no built-in scene shows and it designs one straight away. Designing works as Generate with AI does: the scene goes into Yours, it needs a subscription of your own and uses your own allowance, and the reply offers Keep it and Generate another. The chat can also take the scene away or make it slower, calmer, fainter or still, and every change can be undone. Trying the chat without a subscription, you get a suggestion instead: the scene's name and an Upgrade button, with nothing changed.
An external agent can choose a built-in scene on the MCP server: with
create_map's background for a new map, or with set_background for one
you already have, which the app applies the next time you open it. A scene put behind a map you already have
needs a subscription of your own, as it does in the style sheet: without one, the app leaves the map as it is and
the agent is told why. Taking a scene away never needs one. An agent cannot design a scene.
Choosing a model
The chat has a model picker. The default suits almost everything; a heavier model is worth it for a large restructure or a subtle question, and costs more of your allowance per turn.
Agents have their own model setting, separate from the chat's. An agent that runs hourly on a cheap check should not be on the most expensive model.
What it costs
AI is measured in units — tokens weighted by what the model costs, so one unit means the same amount of money whichever model produced it. Your plan includes a monthly allowance; the editor shows what this schema has used and the file list shows the account total.
In a shared schema, the owner pays. Every turn, every agent run, every service call is billed to whoever owns the schema — not to whoever triggered it.
But doing those things needs a subscription of your own as well, even in somebody else's schema: running an agent, talking to one, calling a service. Reading a schema and everything its agents have written needs nothing at all.
An agent can also be given its own monthly budget, under the account's. That is what stops one misconfigured agent spending an entire allowance before anybody notices. Talking to an agent does not draw on that budget — it is for its runs, and a conversation should not be able to stop it working.
Your data
The AI is Anthropic's Claude, reached through our own server rather than from the app. The app never talks to Anthropic directly, and no API key of yours is involved.
What is sent, and only when you ask for it:
- The map chat: an outline of the schema — each node's id and title — plus your message, anything you attached, and which scene is behind the schema along with the names of your own scenes in Yours.
- The agent chat: that agent's instructions, schedule, settings, budget and its recent runs, plus your message.
- An agent run: its instructions, its node's text, and the results of the service calls it makes.
- A scene background: your description of it, whether it should move, and whether the canvas is dark or light. One made together with a new schema (the generate form's Generate Background Scene) sends that schema's title and the name of its theme in place of a description.
What is never sent:
- Service credentials. Keys live in Secret Manager and are attached to a request on the server, after the model has decided what to call. An op carrying one is dropped rather than forwarded — see the chat page.
- Other people's schemas, or anything outside the one you have open.
What an agent reads is not trusted. When an agent tells you what it found, that text came from Stripe, Slack or whatever else it reached — written by other people. It reaches the model fenced and marked as data, and no reply can change a running agent on its own: it proposes, and you press. Why that matters.
The MCP server is separate from all of this: nothing goes to a model. While its switch is on, the app publishes each map's node titles and notes to our server for the agents you connect. From the next SchemAgentic update it also publishes a fuller copy of each map: titles up to 1,000 characters and notes up to 4,000, each node's colours and shape, check boxes with their labels, the text of other content, cross-links, and the web addresses (http or https only) of links and media. Neither copy ever holds a file path on your device, an attachment or a file (from the next SchemAgentic update, the small picture of a map described below shows its pictures at low resolution), or a schema's agents, services, credentials or connections. From the next SchemAgentic update, once we switch them on, what an agent asks you to approve, and the messages agents leave, wait on our server until you answer or read them (at most 14 and 30 days) and are deleted at your app's first sync 30 or more days after they end. From the next SchemAgentic update, once our server shares them, the copies also say where each map's file is (the drive, the folder and the file name, never the full path on your device) and which maps are open, so that agents' file requests wait until you close a map (see where your files are); file requests wait on our server until your app does them (at most 7 days, 30 for a new map) and are deleted 30 days after they end, and the record of a map you let an agent move to the trash 90 days after it was trashed, or when it is recovered. From the next SchemAgentic update, once we switch it on, an agent also reads, on your behalf, what you can see of a shared map your app published, with less about people (display names, never email addresses; the days people were last there, never where in the map they are), and the saved versions of your maps; versions saved with sync on also hold the copy agents read, and the owner of a shared map can hide it from every member’s agent (see maps shared with you). From the next SchemAgentic update, once we switch them on, the copy agents read can also hold the cells of tables, list items, chart and calendar card titles and times, which of your maps a map links to and the size and type of its pictures and other media (never their names or files, and nothing about places), and a small picture of each map as you last closed it, left out as soon as you change the map; they are deleted with the copy. An export an agent asks for is saved as a new file in SchemAgentic's exports folder on your device, never on a cloud drive, and its content never passes through our server (see rich content and exports). From the next SchemAgentic update, once we switch them on, our server also keeps with each copy a list of the words in it, in a simplified form that ignores capitals, accents and word endings, so agents can find maps by their words; it is made from the copy, never shown as such, deleted with the copy, and no AI model reads or processes it (see word lists and drawings).
Conversations are kept on your device and in your schema so you can come back to them. Deleting a schema deletes its chats with it. The scenes you generate are kept in your account (the 24 most recent), so Yours is the same on every device you sign in on.
When you run out
- At 80% of the monthly allowance the account carries a warning.
- At 100%, new agent runs and new turns are refused with a message saying so. A run already in progress finishes its current step rather than stopping halfway through something.
- The allowance returns on the 1st.
Agents on a schema whose owner is no longer subscribed are paused, never deleted, and start again by themselves when the subscription does.
Not yet
Two things you may expect and will not find:
- Generating a whole schema from a description. Building a mind map from a topic is one problem; building a schema — nodes, services, agents, connections, credentials — is another, and doing it badly would produce something that looks finished and is wired wrong. The chat builds a schema piece by piece instead, which you can follow.
- Using your own Anthropic API key. Not available in this app.