Key concepts

Springbase reuses a handful of words everywhere, and a few of them look like each other. Read this once and the rest of the docs will stop being ambiguous. Every entry links to the page that explains it properly.

How a turn works

Auto

The one mode. There is no Fast / Deep / Research menu on this build: you type, and the model is handed the tools and decides what each step of the turn needs — search, read your documents, write files, generate an image, call a connected app. Nothing classifies your message in advance and routes it down a fixed path. A chip under the answer names the model or models that actually ran. See How Auto works.

Intent

A posture for one message, not a mode you switch into. The Auto menu lists them under "For this message": Image, Video, Music, Research, Build, Deck. Arming one nudges the turn towards a kind of craft; it does not lock the output — "make an image and write the caption" is still one send. Only one can be armed at a time, and sending clears it. Typing /image a heron at dusk arms Image and sends the rest. Video needs Max and Music and Research need a paid plan; Build and Deck are open to everyone. See Chatting with Springbase.

Effort

How much to spend on the models a turn reaches for. Four positions:

Saver, Balanced and Max are ready-made setups chosen by Springbase for your plan; they resolve to the same models for everyone on that plan and you cannot edit them. Custom is your own bench, and it is the only position where model pins you set actually apply — Configure models appears inside the Custom row and nowhere else. Free accounts get Saver only; a paid plan adds Balanced and Max; Custom needs Pro or Max. See Models and effort.

Credits

The single meter for AI work. A turn is charged from its measured cost — the models it used, the tools it called, the reading it did — not a flat per-message price, so a hard turn costs more than an easy one. Certain models are covered by your plan for ordinary chat and image generation; "covered" means the model's own thinking, so web search, document extraction, video, music and audio still draw down credits on top. Fair-use caps on covered models never block you — the turn starts costing credits again. Your balance is on the sidebar account button and in Profile → Usage. See Plans and credits.

Your knowledge

Context, and its @handle

A named, reusable folder of knowledge — documents, notes, pasted text, live sources — that you pull into a chat instead of re-uploading files. Every context has a handle generated from its name, shown as a click-to-copy button on the context page: type @my-context in the composer and that context is attached for the turn. Four handles are reserved and never resolve to a context: meeting, meetings, context and contexts — typing @contexts means "search across all of mine". See Contexts.

Document, and indexing

Anything inside a context: an uploaded file, a pasted note, a page pulled in by a live source, an answer you saved out of a chat. On arrival it is split into passages and embedded so it can be searched by meaning. Until that finishes, the document shows Processing and is not searchable — an answer will tell you documents were still indexing rather than silently skipping them. Failed documents show Failed with a reason, and either a Retry or a Remove. See Adding documents.

Citation

A numbered chip in an answer, pointing at the passage the claim came from. Three kinds exist — a web page, one of your documents, or a meeting — and documents also get a collapsible Sources list under the reply that deep-links to the exact passage inside the context. If retrieval was thin, you get an explicit low-confidence note instead of a confident-sounding guess. See Asking your knowledge.

Live Data

A tab on a context, and the feature that keeps it fed. You point it at a web page, sitemap, RSS feed or YouTube video, or connect an account like Drive, Notion, Slack, Gmail, Outlook, GitHub or Linear, and new items arrive on a cadence you choose (hourly, every six hours, daily, weekly). The cadence is approximate — a due source is picked up within roughly fifteen minutes. Crawls consume credits, a source that errors five times in a row disables itself, and the feature needs Pro or Max. See Live Data.

Reusable work

Skill

Reusable instructions the assistant loads when a task matches — your house style, your review checklist, the way your team writes tickets. A skill is guidance, never authority: it cannot grant a tool, approve an action or widen what a turn is allowed to do. Its short description is always visible to the model so it knows when the skill is relevant; the body only loads when it is needed. You can arm up to three per message from the / launcher, and sending clears them. See Skills.

Recipe

A saved prompt with named inputs — a form you fill in instead of retyping instructions. A recipe can carry contexts, produce a chat answer or media, be shared by link, and be the thing an Automation runs. Anyone signed in can create and run one. See Recipes.

Automation

A trigger attached to a recipe. Three triggers: Time-based (a schedule), Post Meeting (when a meeting finishes), App Events (when a connected app emits an event). Every run produces a chat you can open. Automations need Pro or Max, and each family is capped at ten. The honest limit worth knowing: an unattended run cannot write to your connected apps — it will not send the email or post to Slack, and says which tools were withheld. See Automations.

Connected tool, and approval

An app you have connected — Gmail, Slack, Notion, Linear, Drive and many more — that the assistant can act in during a turn. A connection is in one of four states: connected, pending (sign-in started, not finished), needs reconnect (expired or revoked), or not connected. Approval is the card that blocks a turn to ask you first — but it does not appear for everything. A typed chat can write to a connected app with no card at all, so read a connection as "Springbase may change things here". The card, when it does appear, waits about two minutes and then denies, so a turn you walked away from fails safe. In a live voice session the rule is stricter: every external write asks. See Connected tools.

What it produces

Artifact, and Canvas

An artifact is something a turn makes that you open rather than read — a document, a deck, a web page, a small app. Canvas is the editor beside the chat where you preview, edit, browse version history and export it (PDF, DOCX, PPTX, or the raw file). Publishing an artifact defaults to unlisted. See Canvas and artifacts.

Library

Your file and media shelf, in the sidebar under Library. It holds what you uploaded in chat, what the assistant generated for you, and anything you upload directly, with a storage meter that depends on your plan. You reuse an item with Choose from Library in the composer. It is storage, not knowledge — nothing here is searched when you ask a question. See Library.

Meeting

A captured call — from a notetaker bot, a browser or Mac-app recording, an audio upload, or a pasted transcript. What you get back is a transcript with named speakers plus a summary, action items, decisions and key topics, on two tabs: Summary and Transcript. Meetings can be attached to a chat the same way a context can. See Capturing a meeting.

Sharing

Visibility

Who can reach a thing. The values are Private (only you, plus anyone you invite by email), Unlisted (anyone with the link — the link carries a token you can regenerate), Team (one team in your organisation) and Organization (everyone in it). Contexts also have a second, independent axis: RAG-only, where recipients can query the knowledge and get answers from it but cannot see the source documents. Publishing a context publicly is not self-serve — share by link, team or organisation instead. See Sharing a context.

Words that get confused

• Library vs Context. The Library stores files; a Context is searched. A PDF sitting in your Library is not part of any answer until you either attach it to a message or add it to a context.
• Recipe vs Automation. A Recipe is what runs. An Automation is when it runs. An Automation always points at a Recipe — you cannot schedule a bare prompt.
• Skill vs Recipe. A Skill changes how the assistant works and loads itself when relevant; you never "run" one. A Recipe is a job you run deliberately, with inputs you fill in.
• Intent vs mode. An intent lasts exactly one message and disappears when you send. There is no mode to be in and no mode to get back out of — the composer always returns to plain Auto.