Connect your agent
Think of Nirvai as a pair of hands your favourite AI assistant doesn't have yet. You already chat with a coding or writing assistant — Claude Code, Codex, Cursor, Gemini — and it's great at thinking. What it can't do is remember across chats, keep your data in one place, run your real tools, or build something that outlives the conversation. Connect your agent gives it all of that: one connection, and your assistant can use everything in your Nirvai account.
How it works
Your assistant keeps doing the thinking. Nirvai does the doing and the remembering — every action runs safely in your Nirvai account, never on your assistant's side. The two talk over a common standard (MCP, the "Model Context Protocol" — a shared plug every modern AI assistant understands), so the same connection works no matter which assistant you use.
Shared projects
The same projects your Nirvai agents use — so facts learned in one place are known everywhere.
Your real tools & data
Databases, hundreds of pre-built connections, and your own tools — reused, never re-pasted.
Agents & live apps
Build, run and edit Nirvai agents and interactive dashboards straight from your assistant.
Step 1 — Mint a connection token
A connection token is like a key card for a trusted colleague: a private code that lets your assistant act in your Nirvai account (within the permissions you choose), and nothing else. You create it once in Settings → Connections:
Claude CodeAnthropic's terminal agent. Add Nirvai with claude mcp add.Pick a name so you recognise it later, the provider (your assistant — this tags every action in the feed), and the scopes — what it may do. Full access is recommended; read + run never need confirmation, while write + create let the agent change and build things. Copy the token now — it's shown once.
Anyone with the token can act in your account within the scopes you gave it. Don't paste it into a public chat or commit it to a repo. Lost or shared it? Revoke it on the same page and mint a new one.
Step 2 — Add Nirvai to your assistant
Run one command with your token. The exact command per client — Claude Code, Cursor, any MCP client — is on the next page.
Step 3 — Teach your assistant what Nirvai can do
A fresh chat doesn't know what Nirvai is for yet — when to reach for a database, a project, or an automation. Drop one starter file into your project and your assistant learns every capability without you spelling it out each time. Pick your assistant, then download the file (or copy it in) — it's yours to edit afterwards.
Tells the agent that Nirvai is connected and — the part a fresh session is missing — when to reach for each capability. Copy it in and edit it freely.
Working with Nirvai (connected over MCP)
Nirvai is your backend and native UI: shared memory, structured databases, hundreds
of credentialed tools, cloud skills, agents, live apps, and automations — with an
activity feed where every action you take is visible and deep-linked.
The shape of every action
- Discover, don't guess:
nirvai_list_resources(kind)thennirvai_describe(ref). Address resources by their uuid, never by name. - One thing per call:
nirvai_execute(alias, args, description, session).descriptionis a REQUIRED one-line "why" — it is what the user reads in their feed. - Keep a session:
nirvai_create_session(name)once, pass its uuid assessionon every related call so the work threads together. - Every call returns a link. Hand it to the user; never paste raw data dumps.
First move in a new session
nirvai_download_skill_package("_root", session)and read it.nirvai_list_resourcesfor what already exists — reuse by uuid before creating.nirvai_describethe thing you will touch, download its entity pack, then act.
Where each kind of data belongs
- Databases — anything that is a LIST OF RECORDS THAT NEEDS ORDER: a CRM of contacts, a task tracker, leads, inventory, invoices. Spaces hold Databases (tables) with typed columns (text, status, select, date, person, relation, file) and views (grid, kanban, calendar, form). Verbs:
db.create_space,db.create_database,db.create_view,db.insert,db.update,db.select. Hand back the table link. - Projects — long-term KNOWLEDGE ABOUT SPECIFIC THINGS you must remember across conversations: a client's preferences, an ongoing investigation, decisions and their reasons. A knowledge graph plus a file appendix. Verbs:
projects.recall,projects.add,projects.store_file. Reach for a Project when the fact matters next week, not just this turn — andprojects.recallBEFORE you answer from memory. - Live apps — an interactive dashboard or app built on the user's real tools and data, not a static report. Verbs under
live_apps.*. - Agents — build and run AI agents on the user's account:
agents.run,agents.create. Delegate a real task, then show the run transcript link. - Automations — a workflow on a trigger (manual, schedule, webhook, or an app event) with agent steps and free deterministic steps. Flow:
automations.templatesthenautomations.validate(spec)thenautomations.create(spec); hand off the review link — the user tests and activates. - Tools & credentials — connect a provider and run its credentialed tools in the cloud. Call
credentials.native_requirementsandtools.search_templatesFIRST, thentools.use. Built-ins:web.search,web.fetch,media.image. - Skills — run one of the user's cloud skills end to end:
skills.run. - Help —
docs.search/docs.readfor product documentation.
Non-negotiables
- You NEVER hold a secret. Propose a credential by field NAME; the user types the value on the link Nirvai returns. Never ask the user to paste an API key into chat.
- Reuse before you create. A retry with the same
idempotency_keywill not duplicate. - One plain-language
descriptionper action. It is the spine of the activity feed.
Master tools
nirvai_list_resources · nirvai_describe · nirvai_execute · nirvai_create_session ·
nirvai_list_sessions · nirvai_get_session · nirvai_sync_session · nirvai_transfer_file ·
nirvai_download_skill_package
Recipes
Stand up a CRM
db.create_space { name: "Sales" }db.create_database { database_session, name: "Contacts", columns: [...] }db.insert/db.select— hand back the table link.
Remember a client across sessions
projects.create_store { name, description }(description = when to use it)projects.addto store observations;projects.recallnext time before answering.
Run a task with an agent
nirvai_list_resources("agents")then pick by uuidagents.run { agent_uuid, message }then show the run link.
Ship an automation
automations.templatesthen pick a triggerautomations.validate(spec)then fix the reported issuesautomations.create(spec)then send the review link; the user tests + activates.
What happens next: it proposes, you approve
Your assistant never changes anything behind your back. It works in its own chat, and when it's ready it hands you a bookmark into Nirvai — a review page with the details and a single Create button. You get a preview of what it built before you save a thing. It's the same flow whether it's a live app, an agent, a tool or a login:
Done — I drafted your Sales pipeline live app on your HubSpot deals. Review and save it:
Live appSales pipelineOpen review in NirvaiI set up a Support Agent with your order-lookup tool and help docs. Review it:
AgentSupport AgentOpen review in NirvaiSecrets are always typed by you, on Nirvai's secure page — never through the assistant. Nothing is saved until you click Create.