Commands
A command is a speed-dial button. Pressing one doesn't add a phone line — the line was already there. It just dials a number you would otherwise have to remember, in the right order, without a typo.
That is the whole distinction on this page. Commands are what you type; actions are what your assistant calls on its own. A command adds no capability — it tells the assistant which actions to run, in what sequence, and which link to hand back to you when it's done.
Type /nirvai: in your assistant and the fourteen appear in its command menu.
The fourteen
| Command | Type it when you want to… | What it kicks off |
|---|---|---|
/nirvai:connect | Introduce a fresh assistant to Nirvai | Learns the connection's rules, lists what you already have, and reconciles what it durably knows about you into your Nirvai memory. Run this first. |
/nirvai:context | See your whole account at a glance | Lists your agents, tools, connections, memories, databases and live apps — counts, names, and a link into each. |
/nirvai:remember text | Keep a fact for good | Picks the right memory store (or offers to make one), extracts the people and things involved, writes the note, hands back the link. |
/nirvai:recall query | Look something up you told it before | Searches your memory and summarises the hits with links — or says plainly that it found nothing. |
/nirvai:new_credential provider | Connect a paid service or account | Researches how that provider's login works, proposes the field names, and hands you a review page where you type the secret. |
/nirvai:new_tool description | Turn an API into something reusable | Authors the tool spec, binds it to one of your connections, hands you the review page — then can test it once it exists. |
/nirvai:new_agent description | Draft an AI employee | Lists what you already own, assembles a draft with your tools, databases and memories attached, and stages it for review. Creates nothing. |
/nirvai:new_database description | Set up a place for records | Designs the space, tables, columns and views, and stages them. Schema first — no data at setup. |
/nirvai:new_memory name | Start a new shared memory | Creates the store, optionally seeds a few facts, returns its link. |
/nirvai:run_agent agent | Put one of your agents to work | Resolves the agent by name, runs it in the cloud, waits for the reply, and returns the chat link. |
/nirvai:session name | Keep a long task in one thread | Checks whether that named session already exists, then reuses the name for everything that follows so it lands as one run. |
/nirvai:activity | Ask "what has happened?" | Summarises the recent run — its steps, its files — and points you at the Activity Feed. |
/nirvai:help topic | Get a real answer about Nirvai | Searches and reads the official documentation, then answers from the page and cites it. Never from memory. |
/nirvai:build goal | Build the whole thing, end to end | The onboarding command. See below. |
The two that matter most on day one
/nirvai:connect is the first thing to run in a fresh assistant. It teaches the assistant the
connection ritual and folds what it already knows about you into Nirvai, so your agents inherit it
too. Skipping it means an assistant that guesses at your setup instead of reading it.
/nirvai:build is the onboarding command — the one to reach for when you know the outcome you
want but not the pieces. Given a goal, it:
- Reads Nirvai's recipe book for the closest known shape.
- Grounds the uncertain parts in the official documentation rather than guessing.
- Lists what you already have, so it reuses instead of quietly building a second "Leads" table.
- Plans the steps in your chat — ordered, one line each, saying whose step is whose.
- Works them one at a time, waiting for your confirmation between steps.
- Proves it at the end by running what you built and showing you the reply.
There is no saved Nirvai plan document. Nirvai keeps the record — every action lands in the session's activity run — while the plan stays in the conversation you're already having. A stored checklist would drift the moment a step went differently, and a stale plan claiming work that never happened is worse than no plan at all.
Worked examples
A fact you'll want next month. You type:
/nirvai:remember Acme renewed their contract in March on a 24-month term
Your assistant picks the store, names Acme as the entity the fact is about, writes it, and returns the link. Your Nirvai agents can now recall it too — see Memories.
A connection you can't finish yourself. You type:
/nirvai:new_credential stripe
It looks up how that login works, proposes the fields by name only, and stops. You get a review page, you type the key there, you press Connect. Then it verifies the connection is live and carries on.
A question with a real answer. You type:
/nirvai:help how do I put an agent on WhatsApp
It searches the documentation with short keywords, reads the best page, answers from it, and links you to the page it used — see Web, media & docs for the machinery.
Limits & guarantees
Some steps are only ever yours. Good commands say so before they start, not at the end:
| Only you can… | Why |
|---|---|
| Type a secret | It never passes through your assistant. You enter it on Nirvai's own page. |
| Connect a channel | WhatsApp, Instagram, Messenger, Telegram, Slack, Discord, Teams and the web chat widget have their own setup in Nirvai. No command can do it. |
| Press Save / Create | Every new_* command ends at a review page with a real button. Nothing exists until you click it. |
| Activate an automation | Drafts land switched off deliberately. You turn them on. |
Beyond that:
- A command is a suggestion your assistant follows, not a script Nirvai runs. It expands into instructions in your conversation; the actual work is the same actions documented across these pages, under the same scopes.
- Nothing is deleted. No command can remove anything — deletion isn't grantable over the connection at all.
- One command, one run. Multi-step commands pass a single session name to everything they do, so the whole job is one navigable entry in your feed rather than fifteen scattered ones.
- Your assistant should offer them. When you ask for something a command covers, a well-behaved assistant tells you the shortcut and does the work anyway.