This manual is for the people who operate AI Accelerator: they build agents, connect them to models and systems, put them in front of real users and then keep them running. It does not walk through the platform screen by screen. It explains how each part works, why it is built that way and what consequences each decision you make in the admin has.
The difference matters. A screen can be understood by looking at it: the Save button saves. What you cannot see is what happens next. That saving a published agent creates a new version and activates it on the spot. That an inherited connection disappears as soon as the client creates one of its own. That a WhatsApp session closes by itself after 24 hours of silence. That is what is expensive to discover in production, and that is what is here.
One example that runs through the whole manual
So that every mechanism has a concrete case, the manual follows the same organisation from start to finish.
Nortelink is an agency that builds agents for its clients. One of them is Clínica del Sol, which wants an agent that answers enquiries and manages appointments over WhatsApp and from its website. Nortelink has its own tenant, and Clínica del Sol is a child tenant underneath it. The screenshots in the manual show that scenario. The data is fictitious but has exactly the shape it has on the real platform.
When a chapter says "the clinic" or "the agency", it is talking about this case.
The chapters
The manual has four parts: building an agent, publishing and running it, refining it and going further. At the end there are hands-on guides for specific tasks.
Build
- How the platform is organised. Tenants, agents, versions and roles: the map you need before touching anything else.
- Language models. How a provider is connected, how the model that answers each step is decided and what happens when a provider goes down.
- Building an agent. The ways to create one, how a workflow runs, what a version is and how human approval is requested.
- Giving it knowledge. What happens to a document from the moment it is uploaded until the agent uses it to answer.
- Tools and integrations. How an external API becomes actions the agent can execute, and how credentials are protected.
Publish and operate
- Taking it to users. Widget, WhatsApp, Slack, Teams, voice and API, and how a conversation becomes a session.
- Operating and improving. What the platform records about each conversation and how to use it to find and fix problems.
- Clients, team and access. Client tenants, limits, users, roles and keys.
Refine
- Security and personal data. What is checked before and after the model and how personal data is protected.
- Memory and context. What the agent knows on each turn and where it comes from.
- What the end user sees. Appearance, branding, starter texts, cards and per-channel content.
- Reusing work: skills and templates. Procedures and agents ready to adapt.
Going further
- Quality and alignment. Testing the real agent with fixed cases and comparing before and after a change.
- Voice. Answering calls with Twilio, Telnyx or your own Asterisk PBX.
- Billing, balance and plans. Who pays for each run, top-ups, plans and balance cut-offs.
- Your own installation. Installing, upgrading and licensing your own copy of the platform.
Guides
- Calling an agent from your system. API contracts, keys and errors.
- Connect your AI assistant. Administer the platform from Codex, Claude Code or Cursor.
- Create tests with Quality and alignment builder. Generate test cases from an agent.
Integrations
- Connect Meta: WhatsApp and Instagram. What to create at Meta, what goes in each field and what to do when messages don't arrive.
- Connect Twilio: WhatsApp and voice. Credentials, sender, webhook URLs and outbound calls.
- Connect Slack. The Slack app, its scopes and events, and the URL you have to build.
- Connect Google Workspace and Microsoft 365. The OAuth app, per-service permission and re-authorization.
- Connect Telnyx (voice). The agent as the language model of a Telnyx voice assistant.
If this is your first time, read chapters 1 to 3 in order and then jump to whatever you need. If you already operate the platform, each chapter stands on its own.
How each chapter is written
Every chapter follows the same shape, so you know where to look:
- The mechanism, in order. What the platform does, step by step, with an example of what a piece of data looks like before and after.
- Where it is managed in the admin. The screen where each thing is decided, with a screenshot when the screen adds something the text does not.
- Where it stops. A short, concrete section on what the mechanism does not cover today. If you are going to rely on something for a client, you need to know where the edge is.
- To check. A list of things you can verify yourself, on your installation, to confirm it is configured the way you think.
What each chapter says has been checked against the code of the published version. The update date on each page is the date of that latest review.