nao
nao is an open source analytics agent builder for data teams. You engineer the agent's context like a file system, unit-test its reliability, then deploy a chat interface so anyone in the company can query your warehouse.
What is nao?
nao is an open source framework for building and deploying analytics agents, published by nao Labs, a company started in 2024 that went through Y Combinator in 2025 and works between Paris and San Francisco. Its argument is unusual for the category: an agent's reliability depends on its context, not on the model behind it, so the context is what deserves engineering. The whole product follows from that. The tool comes in two surfaces. nao Agent is the framework and the chat interface; nao IDE is a desktop data editor for macOS and Windows. The agent's life cycle runs through four commands. nao init creates the context as a real file system, where you drop data, metadata, rules, docs, tools and MCP servers without limit, described in a nao_config.yaml and a RULES.md. nao sync pulls context automatically from existing sources: databases, with per-table accessors for columns, description, preview and profiling, plus git repositories such as dbt or Looker, and Notion pages. nao test runs question-to-SQL unit tests and returns hard numbers — pass rate, answer rate, average time, tokens consumed. nao chat then deploys a UI where business users ask questions in plain English and see the agent's reasoning and sources. Eight warehouses are supported: BigQuery, Snowflake, Postgres, Databricks, DuckDB, MotherDuck, Amazon Athena and Amazon Redshift. Four LLM providers are wired in — Anthropic, OpenAI, Gemini and Amazon Bedrock — with Mistral cited in the privacy policy, and the integrations page lists twenty-four connections in all, eleven of them through MCP. Deployment is self-hosted with Docker or Kubernetes, on nao Cloud, or under an Enterprise licence. Security is presented as the first feature rather than an afterthought: the application runs locally and connects straight to your warehouse, credentials never reach nao's servers, only metadata goes to the LLM unless you say otherwise, and the company holds a SOC 2 Type II certification with a Zero Data Retention policy. Named users include LiveKit, Everyday, Finfrog and Qare.
What it does
- Build an analytics agent's context like a file system: data, metadata, rules, docs, tools and MCPs, with no limit
- Deploy a chat interface where anyone in the company asks questions in plain English and gets charts back
- Measure agent reliability with question-to-SQL unit tests before anything reaches production
- Synchronise context automatically from warehouses, git repositories and Notion
- Answer from Slack, Teams, WhatsApp, Telegram, or any MCP client such as Claude, Cursor or Codex
- Self-host the whole stack on your own infrastructure with your own LLM keys
- Write data stories, schedule recurring agent tasks, and cap LLM spend per user
When to use nao / When not to
A quick filter to help you decide if nao is the right fit.
When to use nao
- Data teams that want to own and version their agent's context instead of trusting a black box
- Companies that must keep warehouse data inside their own infrastructure, using their own LLM keys
- Organisations already running a cloud warehouse alongside dbt, git repositories and Notion
- Business users who would rather ask a question in Slack, Teams, WhatsApp or Telegram than open a dashboard
- Engineering-led companies such as LiveKit, Everyday, Finfrog and Qare, which nao names as users
When not to use nao
- Anyone expecting a mobile app: nao IDE ships for macOS and Windows only
- Teams with no data warehouse, and no appetite for building a context layer through a CLI and a YAML file
- Organisations that need turnkey hosting today, since the free tier is self-hosted Docker or Kubernetes and nao Cloud is still early access
- Buyers who need Microsoft SSO, Redshift row-level security or white-label branding without paying for an Enterprise licence
- Users under 18, and teams that need a documented interface language other than English
How to use nao
A typical end-to-end flow, from setup to results.
- Install nao and run $ nao init, which creates the file-system context for your agent
- Declare your sources in nao_config.yaml: databases with their type, project id, credentials path and include or exclude lists, git repositories, and Notion pages with an API key
- Run $ nao sync to pull context automatically — columns, previews, profiling and descriptions from the warehouse, dbt models from your repositories, pages from Notion
- Write the business layer by hand: RULES.md for agent behaviour, business_defs.md, main_metrics.md and example_queries.md
- Run $ nao test to launch question-to-SQL unit tests and read the reliability metrics they return
- Run $ nao chat to deploy the chat interface, where users ask their questions in natural language
- Connect the surfaces your teams already live in: Slack, Teams, WhatsApp or Telegram bots, or the MCP endpoint for Claude, Cursor and Codex
- For production, self-host with the official Docker image or the Helm chart on Kubernetes, and point NAO_CONTEXT_SOURCE at your context repository
- Configure the admin panel: users, authentication, models, prompts, and LLM budgets per provider and per user
- Optionally download nao IDE for macOS or Windows and connect it to the same warehouse
Pros & Cons
Pros
- Fully open source, with a free self-hosted deployment that carries no time limit
- Your data does not transit through the vendor's servers: the connection to the warehouse is local, and credentials stay on your machine
- No training on customer data, backed by formal Zero Data Retention agreements with OpenAI and Anthropic
- SOC 2 Type II certification audited yearly, a public Trust Center, and European hosting in France for the server side
- Reliability is measured before production through question-to-SQL unit tests, which is rare in this category
- Broad coverage of the modern data stack: eight warehouses, twenty-four integrations, eleven of them via MCP
- You bring your own LLM keys, so token cost stays under your control, and budgets can be capped per user
Cons
- No postal address and no legal notice page: the legal identity comes down to the name nao Labs found in the terms
- A single email address handles support, legal matters, DMCA notices and disputes
- The GDPR is never mentioned, despite European hosting, and the terms are governed by Delaware law with mandatory arbitration
- Cloud and Enterprise pricing is not published, and nao Cloud is still behind a Get early access button
- The free offer assumes you can deploy and operate Docker or Kubernetes yourself
- No mobile app, and no published list of interface languages
- A very young company with a very small team: four people shown on the team page, and a team size of two on its Y Combinator profile
Pricing & Plans
A permanent free plan is available. nao Agent can be self-hosted at no cost, with your own LLM key, full access to the context builder and chat UI and Slack integration, while nao IDE offers a free STARTER plan capped at five agent requests per day and including a fifteen-day Pro trial. The lowest published paid price point is USD 30.00 per month for the nao IDE PRO plan, which lifts the request cap and adds team members and direct support on Slack. The Cloud and Enterprise tiers of nao Agent, and the ENTERPRISE tier of nao IDE, are quoted on request and no amount is published; Enterprise licences are issued per company on a yearly subscription. On self-hosted deployments, LLM token consumption is paid directly to the provider.
- self-hosted deployment
- bring your own LLM key
- full access to the context builder and chat UI
- Slack integration
- deployment on nao Cloud
- pre-configured email SMTP
- pre-configured Slack
- Teams and WhatsApp bots
- priority support (Get early access)
- self-hosted or Cloud
- SSO with Okta and Microsoft
- data rights impersonation from the warehouse
- row level security
- LLM budget per user
- company-branded UI
- SOC 2 Type II reports
- priority support and roadmap input
- 15 days of Pro trial
- unlimited data connections
- unlimited AI auto-complete
- up to 5 agent requests per day
- everything in STARTER plus direct support on Slack
- team member invitations and unlimited agent requests
- everything in PRO plus bring your own LLM key
- enterprise workspace
- centralised billing and a dedicated support team
Data, GDPR & hosting
A consolidated view of how nao handles your data.
GDPR overview
There is no mention of the GDPR anywhere on nao's website: the acronym appears neither in the visible text nor in the archived HTML of any page collected. No data protection officer is named, and no Article 27 EU representative is designated. What the site does document is factual rather than regulatory: servers located in the European Union (France), backend and Postgres database on Google Cloud Platform, and the Turbopuffer vector database also on EU servers. nao Labs is SOC 2 Type II certified with a yearly audit, publishes a Trust Center, and references a Data Processing Agreement from its terms. Authentication is SSO-only through Google, Apple or Microsoft, with no passwords stored; traffic uses TLS 1.2+ and stored data AES-256. The terms themselves are governed by Delaware law, with arbitration conducted in Delaware.
Who owns the data?
nao Labs draws a firm line between your data and its own material. Warehouse connection credentials are stored only on your machine and are never sent to nao's servers; the backend neither indexes nor reads the content of your tables, and BigQuery table content never leaves your computer, since the local instance queries Google directly. nao Labs states that it does not share, transfer, disclose or sell your data to third parties. Conversely, everything the service displays other than user-generated content — its proprietary material, the nao Labs name and logo — belongs to nao Labs or its licensors. The terms in force are dated 3 July 2025.
Reuse rights
Collection is limited to what you explicitly share: an indexed codebase and your use of the chat. Embeddings are generated through OpenAI's API and stored in the Turbopuffer vector database on EU servers. Source code itself is never stored, only file paths, and only warehouse metadata — table names, column names, types and descriptions — feeds those embeddings. Chat conversations are stored in nao's database, and depending on the model you choose the underlying provider may store them too, though you can disable the models you do not want to use. No customer data, whether metadata, queries, code or usage logs, is ever used to train nao's models or any external one, and formal Zero Data Retention agreements with OpenAI and Anthropic are meant to guarantee it. By default only non-sensitive metadata reaches the LLM; query results are sent only if you explicitly authorise it. On the other side of the line, nao's own proprietary material may not be reused without prior written consent.
Data retention & training
Hosting summary
nao's servers are located in the European Union, specifically in France, with the backend and the Postgres database running on Google Cloud Platform. The Turbopuffer vector database that holds the embeddings is also on EU servers. DNS resolution of the domain points to 34.155.160.35, an address located in Paris, France, on Google's AS396982. The important nuance is that most of your data never reaches those servers at all: the application runs locally and connects directly to your warehouse, so queries, results and table contents stay on your machine, and connection credentials are stored locally rather than sent to nao. Only metadata and embeddings travel. If you self-host, everything stays inside your own infrastructure. Traffic between your app, the warehouse and nao's servers uses TLS 1.2+, and stored data is encrypted with AES-256.
Things to keep in mind
Risks and trade-offs to weigh before adopting nao.
- The publisher's legal identity is thin: no postal address, no legal notice page, and the name nao Labs found only inside the terms of service
- A single email address carries support, legal questions, DMCA notices and disputes, which is fragile for anything that needs a paper trail
- The terms are governed by Delaware law and impose arbitration in Delaware, which is a real constraint for a European customer
- The GDPR is never mentioned, no data protection officer is named and no Article 27 representative is designated, despite servers in France
- Chat conversations are stored in nao's database, and depending on the model chosen the provider may retain them too; no retention period is published for them
- The Data Processing Agreement lives in a shared Google Docs file rather than on the site, so its permanence is not under the vendor's page control
- An agent that answers in plain English can quietly become the only reading of the data: the reliability numbers nao produces are useful precisely because they invite scepticism rather than replace it
Setup & Integrations
Technical difficulty
Moderate for the data team, negligible for everyone else. The free path is self-hosting: the official Docker image or the Helm chart on Kubernetes, configured through environment variables such as NAO_CONTEXT_SOURCE, with the context declared in a nao_config.yaml and warehouse credentials set up locally. Four CLI commands cover the cycle, and the quickstart claims a first agent in one minute. You must supply your own LLM key. nao Cloud removes the deployment work but is still early access, and nao IDE installs like any macOS or Windows application. Business users only open the chat or a Slack bot.
Deployment
Integrations
Behind nao
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What exactly is nao?
How do I deploy it?
Which warehouses and tools does it work with?
Is there a free plan?
Is my data used to train AI models?
Where is the data hosted?
Does nao offer an API?
Can I self-host it and keep everything internal?
Is there a minimum age?
Is there a mobile app?
Should you pick nao?
nao takes a position that sets it apart from most agentic analytics tools: it bets on the context and on measuring that context, rather than on the model. That choice is visible everywhere in the product — a context held as an inspectable file system, versioned in a git repository, and unit-tested with question-to-SQL cases that return a pass rate before anything reaches a business user. For a data team tired of being asked to trust a black box, this is a genuinely different proposition. The security posture is equally deliberate. The application runs locally and talks straight to the warehouse, credentials never leave the machine, only metadata reaches the LLM unless explicitly authorised, and the company holds a SOC 2 Type II certification with a Zero Data Retention policy and European hosting. Being fully open source, with an unlimited free self-hosted tier and your own LLM keys, lowers the barrier to entry a long way. The reservations are the mirror image of that youth. nao Labs was started in 2024, shows four people on its team page, publishes no postal address and no legal notice, and routes support, legal matters and disputes through a single email address. The GDPR is never mentioned even though the servers sit in France. Cloud and Enterprise prices are quoted on request, and the free path assumes you can run Docker or Kubernetes yourself. The result suits data teams that already have a warehouse, a modelling layer and the appetite to own their agent's behaviour. Organisations looking for a managed, turnkey answer will find nao Cloud still in early access, and should treat the missing legal and regulatory details as questions to raise before signing.
- Choosing a selection results in a full page refresh.
- Opens in a new window.