Docs · Getting started

From nothing to
a working coworker

Five steps. Twenty minutes, most of it spent waiting for other people’s OAuth screens.

Before you start

You need three things: a workspace (we create it on signature), an administrator on your identity provider, and one person willing to be the first agent owner. Pick someone whose week is full of drafting and chasing — the value shows up faster.

  • An admin on Slack or Microsoft Teams, to approve the app
  • An admin on your identity provider, if you want SSO on day one
  • A rough answer to “what counts as sensitive here?” — you can refine it later
You do not need a data migration

Agents read from your tools in place. Nothing is copied into a separate store unless you explicitly add it to the vault.

The five steps

  1. Connect the chat surface. Install the Wildfire app in Slack or Teams. Your admin approves it once; individuals do not need to.
  2. Set the tier policy. Three tiers, and a default. Tier 1 is public-grade, Tier 2 is client confidential, Tier 3 is regulated. If in doubt, start strict — loosening later is a one-line change and it is logged.
  3. Authorise connectors. Start with two: where the files live and where the deals live. Drive plus HubSpot covers most companies. Read-scoped by default.
  4. Bootstrap the vault. Point us at the folders that matter. The custodian structures them, flags duplicates, and tells you what it could not read.
  5. Name your first agent. One person, one agent, one name. Give it three real jobs from that person’s actual week and let it work them badly the first time.

The fifth step is the one people skip. Do not. An agent that has watched three real workflows is worth more than one configured perfectly in the abstract.

Core concepts

Agents

One per person, named by that person. Private memory: what you tell your agent stays with your agent. It drafts in your voice because it has read what you have written, not because you filled in a tone slider.

The orchestrator

One per team or company. Lives in channels rather than DMs. Pulls status, spots handoffs going stale, and writes the digest. It can see what the team has shared, never what an individual told their own agent.

The custodian

You will rarely talk to it. It keeps the vault clean, signs every action into the ledger, and produces the export your DPO asks for. It is the reason compliance is not a project.

Three memory spaces

  • Personal — visible to one person and their agent
  • Team — shared deliberately, by a human, once
  • Vault — company knowledge, versioned and attributable

Promotion goes one way and needs a human. Nothing drifts upward on its own.

Skills

A skill is a saved way of doing something — a report format, a review checklist, a follow-up pattern. Skills start personal, get promoted to the team when they prove out, and are sealed into the ledger when they touch regulated work.

Admin & SSO

SAML against Okta or Entra ID. Provisioning is SCIM if you want it, manual if you do not. Three roles: member, workspace admin, and policy admin. Only a policy admin can change tier routing, and every change writes an entry naming who made it.

  • Deprovisioning a user suspends their agent and freezes its memory rather than deleting it, so the audit trail survives the leaver
  • Connector tokens are per-user, so revoking a person revokes their reach
  • Nobody, including us, can read Personal memory. There is no support override

What next