What's true
today.
How Neivan protects your data, where to find the docs, and every integration with its real status — what works now, what is rolling out, and what is still being built.
Your data,
kept apart and on record.
Agents touch your email, files and books, so the controls come first. Each company's data is kept apart at the database, connection tokens are encrypted, risky actions wait for a person, and every action is logged. Where something isn't done yet, this page says so.
No report yet; written policies modelled on SOC 2; independent audit not yet engaged
DPA + SCCs available; rights center built in
Consent management, data export, and deletion built in
Contact us about regulated workloads
Targeted 2027
Not required for current workloads
Independent test before general availability
At rest, HTTPS (TLS 1.2+) in transit, credentials sealed with AES-256-GCM
Not a policy document.
The controls, working.
A plain-English preview of what each connection can read and write. Four roles and about fifty permissions, overridable per person. Approval cards on anything risky, an audit log you can export — and every AI call metered against budgets you set for the company, a department or an agent.
Controls in detail
Tenant isolation
Postgres Row-Level Security at the database boundary. Every row carries an organization_id and queries are scoped automatically. Multi-tenant SaaS with strict row-level isolation per organization.
Identity & access
Four roles (owner, admin, manager, employee) and about fifty permissions, each overridable per person. Sign in with email and password. Authenticator-app MFA; an admin can require it, and then approvals, connection changes, invites and permission changes all need it. SCIM user provisioning is live (create, update, deactivate — no group sync yet). SAML single sign-on through your identity provider is available on request.
Encryption & key custody
Encryption at rest and in transit (HTTPS, TLS 1.2+). Connection tokens — and your own AI provider keys, where you bring them — are sealed with AES-256-GCM, decrypted only on our servers at the moment a call needs them, and never sent to a browser. Bring-your-own AI keys are built and switched on per customer; ask us.
Audit trail
Every meaningful action is logged with who did it, when, and the request id, in an audit log with an integrity check. Agent and connection actions carry their approvals. Admins can export the log as CSV, and each agent has an evidence pack export.
Approvals
Risky agent actions stop on an approval card. You choose who may approve — named people, a role or a team — and can require two or more approvers. The person who asked can't approve their own card unless nobody else can. Approval chains with steps and deadlines cover requests like expenses and time off.
Service-to-service auth
Every internal service-to-service call requires shared-secret auth. No implicit trust between agents, workers, or stores.
AI usage and spend
Customer data is not used to train models and is never visible to other customers. Every AI call is metered, and you can set monthly budgets for the company, a department or a single agent, with alerts before the limit and a hard stop at it.
Hosting
Hosted in the United States: the app on DigitalOcean (New York) and the database on Supabase (AWS us-east-1). We don't offer EU hosting or a dedicated deployment today.
How we earn the right to your most important data.
Boundary by default
Your data never leaves your tenant. Memory, models, and decisions are partitioned per customer with row-level security enforced at the database.
Explainable by construction
Actions are logged with the actor, input context, and approvals involved, so you can reconstruct what happened and why.
Approved before impact
High-impact actions pause in an approval queue until a human signs off — nothing irreversible happens silently.
Need the full security packet?
Our security overview, current SOC 2 progress summary, and answers to your vendor questionnaire are available to qualified prospects under NDA.
Every system needs a
map.
Short guides for the first week, written for the person running the business, not an engineer.
Getting started
GuideBuy a pack, connect your accounts, answer the setup questions, approve the first cards.
Open →How approvals work
GuideWhat stops for a person, who can approve, and what happens when you reject.
Open →What each pack needs
ReferenceThe accounts each Agent Pack connects to, and what it can and can't do with them.
Open →Integrations
ReferenceEvery app in the catalogue, with its real status today.
Open →Every app, with its
real status.
Our catalogue lists 214 apps. Today agents can act on 23 of them with ready-made actions, and call the API of 2 more with every call approved. The rest are signed in only, rolling out, being built, or planned — each one sits under a heading that says which.
Every provider below is wired into the model gateway — pick one per agent. Using your own provider API key is built and switched on per customer on request. Nothing is locked to a single vendor: switching an agent to a different model is a dropdown, and the cost of every run is recorded against whichever provider served it.
Claude — the default for agent reasoning and long-context work.
GPT models, including the cost-efficient mini tiers.
Gemini — strong multimodal and very large context windows.
European models, open weights, EU data residency.
Very low cost per token for high-volume batch work.
Grok, with real-time context from X.
On the roadmap: AWS Bedrock · Perplexity · Self-hosted — the gateway has no adapter for these yet, so they aren't listed above.
Connected in production — a real sign-in has completed with a real account. Split below by what an agent can do with it today.
Signed in, and an agent tool reaches it — these are the connections your agents can actually read from and write to right now.
No ready-made actions yet. An agent writes each request to the app’s API itself, and every call waits for your approval unless you set the agent to run unattended. Each of these has been used this way by a real agent in production.
The connection is real and stored, but we don’t yet claim an agent can act on it: either no agent tool targets it, or the tool is built and hasn’t yet been proven against a real account. Some of these surfaces are also reachable through their parent connection (Calendar and Docs via Google Workspace, for example).
You can connect these in the app today, but no agent has been proven on them in production yet. A few still wait on the vendor’s own app review.
The integration is written but not offered to customers yet — it still needs the vendor’s app registration or review.
Not built yet. Tell us which you need — it helps us decide what to build next.
Don't see a system you use?
Tell us which app you need. It helps us decide what to build next.
Start with one agent.
Add more when it earns its keep.
Agents are billed monthly, and what they send or change waits for your approval. The full Neivan platform opens as a free beta for selected companies in November 2026.
Not sure where to start? Take the free AI Readiness Check — free, about seven minutes.