You've installed the Host. You've connected the Desktop. You've maybe paired the iOS app. The engine is sitting on your hardware, wired into your Tailnet, waiting for you to give it something to work with.
This is where most knowledge tools lose people. The app is open. The blank screen is judging you. You don't know what to do first. You think I'll come back to this when I have more time, and you never come back.
Carabase is built to avoid that trap — the engine does most of its work in the background whether you touch it or not — but there are a handful of habits that take it from "interesting piece of software" to "thing I cannot work without." This guide is the short version of those habits. Read it once. Refer back to it in a week.
The setup day
Before anything else, spend twenty minutes on the ingestion controller. This is the page where you decide what the engine gets to see. The temptation is to connect everything at once and sort it out later. Resist it. The engine is only as useful as the signal-to-noise ratio of what you feed it, and the ratio is easiest to protect on day one.
Recommended first pass:
- Calendar — connect the calendar you actually use. If you have a separate personal calendar, decide now whether you want the engine to see it. Most people should say yes, filtered by event type.
- Email — do not connect all of inbox. Connect specific labels. "Sent" is usually high-signal. "Clients" or a project-specific label is high-signal. The newsletter folder is not.
- GitHub — connect the repos you work on, scoped to PRs and issues. Skip the dependency bumps.
- Meetings — connect Granola (or whatever you use) to capture transcripts. This is the single highest-signal source most people have available.
- iMessage — consider carefully. For most people, this is a line too far. Therapy, family, your running inside-joke with your partner, and the board drama all live in the same pipe. If you still want to connect it, use thread-level filters. Aggressively.
The ingestion controller is also where you set the Host's dream cycle — the nightly pass where the engine synthesises folio summaries, extracts new entities, runs memory distillation, and updates the knowledge graph. Leave it on. The engine gets smarter while you sleep — a sentence I did not think I would write in a product manual, but here we are.
Workflow 1 — Capture on the move
The iOS client is the one you will use more than you expect.
You're walking, listening to a podcast. Someone says something that lands. Pull out your phone, screenshot the player, swipe to Carabase, hit the capture button, dictate: what Jobs said here about saying no to a thousand things — drop in projects/product, tag it philosophy. The capture lands in your Host over Tailscale. When you open the Desktop later, the screenshot is waiting, tagged, filed, linked to the project folio, entity-extracted ("Steve Jobs" is now a node in your graph if it wasn't already).
Three months later, in a board meeting, you want to make the "saying no" point. You hit Cmd+Shift+Space, type that thing Jobs said about saying no, and Carabase returns the screenshot, the transcript around it, and your one-line reaction. In under 100ms. You paid no Librarian Tax on the way in. You paid no Retrieval Tax on the way out.
This is the core move. Learn it on day one.
The same pattern works for:
- A Siri Shortcut on your Lock Screen that starts a dictation capture and files it straight into today's log
- The Share Extension, for any webpage or snippet you want to save without leaving what you were doing
- Voice memos from a run, which the engine transcribes and tags overnight
Capture is the cheapest input. Build the habit of reaching for the iOS client whenever a thought would otherwise be lost.
Workflow 2 — The five-minute morning log
Open the Desktop. Cmd+1 focuses the sidebar, Cmd+2 focuses the middle pane. You are now in today's daily log.
Spend five minutes typing what is on your mind. Forget tags. Forget folios. The engine handles those. You handle the thinking. If you want to be explicit — if you know this note is about a specific project — use a slash command (/folio:q3-roadmap) or an @ mention (@david-chen) and the engine will file it there. Otherwise, just write.
The five-minute log is the highest-leverage daily habit in Carabase. It gives the engine raw material to reason over for the rest of the day, and it gives you a place to think without filing anything anywhere. When you ask Carabase something an hour later, the answer will be better because of the five minutes.
A pattern that works: open the log first thing, type what you did yesterday and what you intend to do today, close the log, go do the thing. The engine will pull tasks out of it automatically and surface them in the tasks panel.
Workflow 3 — The summon, not the chat
Most of your interaction with the AI side of Carabase should be through the summon panel (Cmd+3, or tab to it). Not the chat window. The summon panel is the same interface, but the framing is different: you are asking an oracle a question, not making small talk.
Good summons look like this:
- What did I commit to David this week?
- Summarise the three most recent Acme conversations, with citations.
- Projects blocked on me right now.
- Who introduced me to the lead engineer at the company we almost acquired in February?
- Verify the claim that we agreed to a 24-month indemnity cap.
Bad summons look like every conversation you have ever had with ChatGPT:
- Hey can you help me think through something?
- I've been thinking about my Q3 strategy, what do you think?
- Act as a strategic advisor and...
The distinction is that Carabase has access to your actual context. Asking it to generate strategic advice from nothing is asking the weaker part (the LLM) to do the work the stronger part (your retrieval-indexed life) is built for. Ask for retrieval. Ask for verification. Ask the specific thing you would ask a very attentive colleague who had been sitting next to you for five years.
The engine will sometimes return short answers (milliseconds, no LLM involved — pure retrieval with citations) and sometimes longer reasoning (a few seconds, LLM summarisation on top of retrieval). You will learn to recognise which questions get which kind of answer. Both are useful. The former is more common than you'd think.
Workflow 4 — Verify before you trust
Carabase ships a hypothesis verifier. Use it.
When you find yourself about to repeat a fact you sort-of-remember — "we told Acme the price was firm," "the deadline is the 15th," "Sarah said she was fine with the new scope" — type the claim into the verifier before you act on it. The engine partitions the evidence into corroborated, contradicted, mixed, and inconclusive, with citations to specific artifacts. You will discover, quickly, that a meaningful fraction of what you think you remember is wrong.
This is especially valuable before a meeting where you're about to make a commitment, before a negotiation where you're about to cite a precedent, and before any email where you're about to assert a fact whose source you can't quite remember.
A rule worth adopting: if it's going into a document, a contract, or a senior-stakeholder email, verify it.
Workflow 5 — Plug in the other AIs
The quiet flex is the MCP server.
Carabase's Host exposes your knowledge base as a Model Context Protocol server at /mcp/sse. Any MCP-compliant AI tool can query it. This means when you use Claude Desktop on the same Tailnet, Claude has access to your entire context — not because you pasted it in, not because Claude remembers, but because Claude is querying your substrate directly every turn.
The practical consequence: you can keep using whatever coding assistant, research assistant, or draft-writing assistant you like. They all get to see your context. They all stay as good as they already were. The intelligence lives in the substrate, and the agents rent it.
Set this up on day three, once the ingestion has caught up. In Claude Desktop's settings, add the Host URL. You're done.
A few things to avoid
- Do not try to pre-organise. The old Notion instinct — set up a perfect hierarchy before you start — is the Librarian Tax trying to kill Carabase on day one. Resist it. The engine's job is to organise. Your job is to think.
- Do not chat your way through problems. The summon panel is not a therapist. When you find yourself typing a long prompt explaining context, stop. Carabase already has the context. Ask the actual question.
- Do not connect every integration at once. Signal-to-noise matters more than volume. A thoughtful ingestion is worth ten firehoses.
- Do not share your Host URL. It's supposed to be on your Tailnet only. The security model assumes the transport is private. Don't test the assumption.
The thirty-day mark
Give Carabase thirty days of reasonable input. A daily log most mornings. A handful of summons a day. Captures on the move as the habit builds. Ingestion tuned after the first week.
Somewhere in week three, the engine will answer a question with a specific fact you had forgotten — a name, a date, a phrase, a commitment — and you will feel the shape of what you've been missing from every other tool. That moment is the product. Everything in Parts 1 and 2 is just an argument for why it should exist.
You are now running an engine that remembers.