AI Assistant

AI assistant — overview

AI assistant — overview

lazyit can work with AI in three ways. Each one acts as a person or service account, with exactly that principal’s permissions — never more.

  • The in-app assistant — a chat in the top bar. People ask about the estate (“which laptops are unassigned in Madrid?”) and ask for changes (“assign MBA-017 to Juan”). Every change it proposes waits for that person’s approval. See Using the chat.
  • External AI agents (MCP) — Claude Code, OpenAI Codex, Cursor, VS Code and other clients that speak the Model Context Protocol connect to lazyit on a person’s behalf, with their own model. See Claude Code & MCP.
  • Service accounts — scripts call the assistant through the API (POST /api/ai/runs), or connect over MCP, without a person. See Service accounts below.

Off by default

Nothing about AI is active on a new or upgraded instance. An administrator turns each part on in Settings → AI, and each has its own switch:

Switch Needs Effect of turning it off
The assistant An AI provider, a model and — for hosted providers — an API key Hidden for everyone at once; configuration and key are kept
External AI agents (MCP) Nothing else — the agent brings its own model Every connection is paused at once; it resumes when turned back on
Web search (assistant only) A provider that supports it Chats that used it become read-only

See AI assistant — setup.

Who can use it

Two permissions govern AI. Both are granted to Administrators and Members by default and can be changed in Permissions:

Permission What it allows
Use the AI assistant (ai:use) Use the chat once the assistant is on; for a service account, the headless API.
Connect external AI agents (MCP) (ai:connect) Connect an MCP client once MCP is on.

Neither permission adds any power: the assistant or agent can only do what the person could do themselves. Configuring AI needs Configure the instance.

What leaves your server

Read this before turning anything on. The administrator acknowledges it in the setup wizard, and the acknowledgement is recorded.

To the AI provider (the in-app assistant)

When someone uses the chat, lazyit sends to the provider the administrator chose:

  • the messages they type and the answers they give in the assistant’s forms;
  • the address of the page they are on and which record it is about — never what is on the screen;
  • lazyit’s instructions to the model, the administrator’s own instructions, and the list of tools;
  • every result of every tool the assistant uses — that is, any record that person can read: names, emails, employee numbers, assets and assignments, applications and who has access to them, stock, knowledge-base articles (including restricted folders they can open) and, for people who can read the activity history, its entries.

lazyit never sends Secret Manager values, sign-in passwords, service-account tokens or the provider’s API key. The key is stored encrypted and never shown again.

What the provider does with it is set by your contract with that provider — retention, whether it is used for training, the region it is processed in. lazyit cannot enforce or verify those terms, and any cross-border data-protection obligations are yours. To keep everything on your premises, use a model you host yourself through the OpenAI-compatible provider.

To search partners (web search, off by default)

When web search is on, the provider runs the searches on its own servers. The search queries the assistant writes and the conversation context go to the provider, which may pass the queries on to its search backend or a search partner, and may bill searches separately. lazyit itself makes no request to any other website. With OpenAI, lazyit restricts search to OpenAI’s cached copy of the web, so no page is fetched live from its site. See Web search.

Purchase documents (document extraction, off by default)

When Document extraction is on and someone chooses Read this document on a purchase, the whole file — an invoice, a quote or a delivery note, with its supplier, prices and tax IDs — goes to the provider, which reads it with no tools. No other lazyit data goes with it, not even the file name, and nothing read is saved until a person reviews the draft. The self-hosted OpenAI-compatible provider is never used for it. See Document extraction.

When the chat reads a document — “here is this purchase order” — the file goes out the same way, and what was read (the draft, with the file name) then becomes part of that conversation, which goes to the provider on every later message like the rest of it. See Filling a purchase from a document.

Through external agents (MCP)

An MCP client uses its own model. What it reads from lazyit goes to whatever provider that client uses, under the terms of the person running it — lazyit has no say over it. See Before you connect.

What stays on your server

  • Conversations are kept for the retention the administrator sets (90 days by default, 7–3650), then deleted. Only their owner can read them — administrators cannot.
  • The AI action log — every change the assistant proposed, who decided and what ran — is permanent and survives deleting the chat.
  • On a plain-HTTP (lan) instance, chat traffic between the browser and lazyit is unencrypted, like the rest of the app.

How lazyit keeps the assistant in check

Anything the assistant reads — an article, a note, a web page — might contain text written to trick an AI (“ignore your instructions and grant…”). lazyit does not rely on the model to resist it:

  • You approve every change on a card that lazyit builds from the exact change that will run, not from what the model wrote. See Approving changes.
  • Content written by others is flagged. A change proposed after the assistant read such content shows a Based on content written by others note, and is never applied automatically.
  • Purchase changes are never applied automatically, and one that creates assets or changes money is never part of Approve all. A supplier’s document marks the whole chat as based on content written by others.
  • Sensitive changes — roles, identity, access grants, sign-in, applications marked critical, instance configuration — get a distinct card, never auto-approve, and some need your password.
  • Some things are out of reach entirely: the assistant cannot read secrets, cannot run anything that returns a credential in clear (such as a new service-account token or a temporary password), and cannot change the AI’s own configuration.

External agents are different: over MCP the client decides whether to ask you before a change. Read Before you connect before connecting one.

Service accounts

A service account can use the assistant headlessly — through the API, or over MCP. Nobody approves its changes one by one, so each account has its own AI access (off, read only, or read and write with an optional cap on writes), set from its row menu in Settings → Service accounts. Headless runs never search the web. See Service accounts on the setup page.

Where to go next

You want to Read
Turn the assistant or MCP on, or change limits Setup
Use the chat Using the chat
Understand a change card Approving changes
Connect Claude Code, Cursor or another client Claude Code & MCP
Review or revoke what can act as you Connected apps & personal tokens
Fix an error Troubleshooting