The stack behind our agents.
The models, tools and infrastructure we actually use to build AI agents and products — what each one does for us and where it runs. No affiliate links, no sponsorship: when a tool stops earning its place, it goes.
Principles
- Models
- Chosen per task, never one provider for everything
- Hosting
- Containers we run, not serverless platforms
- Data
- Open-source databases we can dump and restore
- Messaging
- Official WhatsApp Business Platform for clients
Models and speech
| Tool | What it does for us | Where we use it |
|---|---|---|
| Anthropic Claude | Reasoning and tool use for agents; our team's coding assistant | Restaurant booking assistant, Mama Bot, the gym agent, the training platform's assistant, our own engineering |
| OpenAI models | Structured extraction and generation | Menu import and writing in Vimus, agents where it fits the task |
| Open-weight models via OpenRouter | Fast, low-cost models for high-volume internal assistants | The assistants of Takat OS |
| Whisper (open source) | Speech-to-text on a machine we run | The meeting notetaker, with no audio sent outside |
| ElevenLabs | Natural voice for a spoken agent | The voice-ordering prototype on WhatsApp calls |
| Google text-to-speech | Voice replies to voice notes | The gym agent's spoken answers |
Messaging and email
| Tool | What it does for us | Where we use it |
|---|---|---|
| WhatsApp Business Platform (Cloud API) | The official way for a business to talk to customers on WhatsApp | Client agents: bookings, ordering, gyms, insurance follow-ups — see WhatsApp AI agents |
| Mattermost | Open-source team chat we host, with our bots inside | Takat's own team chat and its assistants |
| Resend, SendGrid | Transactional email | Booking confirmations, notifications, forms |
Where it runs
| Tool | What it does for us | Where we use it |
|---|---|---|
| Docker Compose | One stack per project: app, database, workers | Every production project |
| Caddy | The single web gateway, with automatic TLS | In front of every site and app |
| GitHub Actions | Builds and tests every change, tags images with the commit | Every repository; nothing is built on a production server |
| Google Cloud (virtual machines) | The servers our stacks run on | Production, previews and development environments |
| Cloudflare | DNS | Our domains and those of the products we run |
| PostgreSQL, MySQL, Redis | Databases and cache, each in its project's own container | Booking platforms, menus, CRMs, bots |
| NocoDB | A database with an API and an admin view | The data behind Takat OS |
| Sentry | Error tracking | Client platforms |
What we build with
| Tool | What it does for us | Where we use it |
|---|---|---|
| React, TypeScript | Interfaces for staff, owners and customers | Booking dashboards, Vimus, this website |
| Next.js | Sites and apps that must be fast and indexable | Assermento, the creator platform, multi-brand sites |
| Vite | Fast single-page apps and prerendered sites | Vimus, this website |
| Astro | Static e-commerce and content sites | A custom online shop |
| Stripe, Paystack | Online payments, including African cards and wallets | A custom online shop, Vimus subscriptions, hospitality platforms |
Fewer moving parts, data we can restore
We prefer containers we run to serverless platforms: one place to look when something breaks, one bill, and data we can dump, restore and move. We pick a model per task — the best reasoning where the agent must act, a fast and cheap one where it must sort, an open-source one where the data must not leave. And every tool has to be replaceable: an agent's logic lives in our code, not inside a vendor's builder.
How these pieces fit together is described on how we build.
Frequently asked questions
Which AI model does Takat use?
Several, chosen per task: Anthropic Claude for agents that reason and act, OpenAI models for extraction and generation where they fit, open-weight models for high-volume internal assistants, and open-source Whisper for transcription on our own machine.
Is Takat tied to one AI provider?
No. The agent's logic, tools and data live in our code and your database, so the model behind a task can be changed.
Do you use no-code chatbot builders?
No. Our agents are written as code, tested, versioned in git and deployed as containers, so they can be reviewed, rolled back and moved.