Docs

Get started

Key concepts

The handful of words the rest of these pages use, and how they fit together.

#Workspace

Your company's space in IA4AI. Everything you send is filed under one workspace, and every read is limited to it: two companies on IA4AI never see each other's data. A workspace has an address of its own, https://ai.odux.io/<workspace>/, and the people you invite sign in to it with Google.

#Keys

Key Looks like What it can do Where it goes
Workspace key ia_… Send usage, read this workspace's own reports (and its captured content, while content mode is on), and manage its provider keys. Nothing in any other workspace, and no operator routes. Your agents, the gateway, your assistant's MCP connector
Extension key iax_… Only fetch the extension's settings and send usage events. It cannot read anything. Every browser running the extension
Provider key sk-… (the provider's) Pays your provider for your calls. Stored in IA4AI, encrypted, for the gateway to call with

IA4AI keeps only a hash of the workspace and extension keys, so it cannot show them again: each is shown once when you create it. Creating a new one replaces the old one, which stops working at once.

Your provider keys are encrypted under a key derived from your workspace key, which IA4AI does not store. So stored provider keys cannot be read at rest, even by us; each is opened in memory only while one of your calls passes through.

#The approval code

Your workspace key is pasted into your tools, so a copy of it could end up somewhere you did not intend. Three actions should not be possible with a copy alone: exporting captured content, replacing the key, and changing your provider keys. For those, your workspace sets an approval code in Setup. It is never pasted into a tool: a person types it at the moment one of those actions happens.

  • IA4AI keeps only a fingerprint of it and cannot show it again. If you forget it, write to us and we reset it; the reset is recorded.
  • Five wrong tries lock it for 15 minutes.
  • Someone signed in as the workspace's admin has already proven who they are and is not asked for it.
  • Until you set one, the key alone can do those three things. Set it when you create your key.

#Roles

A key or a signed-in session carries roles. Your workspace key and your team's sessions have viewer (read the workspace's reports) and investigator (send usage). Your workspace's admin signs in with Google, and only they can turn on content mode. The extension key has only browser. The API reference names the role each endpoint needs.

#Agent

Anything that calls a model: a support bot, a coding agent, an internal app. You name an agent with the x-ia-agent header or by putting the name in the gateway address (/gateway/agent/<name>/v1). Otherwise IA4AI names it from what the tool says about itself (Cline, Roo Code, Claude Code, n8n and the OpenAI SDKs are recognised), or from the label you gave the key. Anything else shows as unidentified, never as a guess.

#Model call, run and job

  • A model call is one request to a provider: model, tokens in and out (cached and uncached), latency, tools named, and its cost.
  • A run is the calls one agent made for one piece of work, in order, as numbered steps.
  • A job is a kind of work an agent does again and again. IA4AI groups runs into jobs from a fingerprint (a hash), without storing their text.

#Measured, estimated, untested

Every number and every finding says how sure it is:

Status Means
Measured Read from a record: a provider's reply, a call's token counts, a bill.
Estimated Worked out from measured records with a stated method, for example a cost at list price.
Untested A candidate, such as a cheaper model that has not been run on your traffic. It carries no saving.

A saving is stated only once it is verified: a cheaper model replayed on your own calls and shown to answer the same way. See verification.

#Investigation

A look at one period of your workspace, ending in findings (each measured, estimated or untested, each with what it costs), up to five conclusions in plain words, and one next step. An analyst may review it before you see it. You can accept, reject or question each finding. See Investigations and conclusions.

#Case

The record of one analysed agent run: what the agent said it did (the claim), what it actually did (the behaviour), and what independently proves it (the provenance), with the findings and cost. Cases come from runs sent to /ingest/trace.

#Evidence log

An append-only record of what IA4AI stored and decided, hash-chained and signed. Any change, reordering or deletion after the fact is detected, and anyone with a workspace key can check it (verify_evidence). It proves a record was not altered after it was stored; it does not prove the record was true when it arrived, which is why every finding carries its status.

#Capture and golden sets

Off by default. When your workspace admin turns on content mode, IA4AI keeps redacted prompts and answers for a limited time, so that a golden set (fixed real inputs with the answers your current model gave) can be built and a cheaper model tested against it. See Capture, golden sets and verification.