- Home
- Tech stack
Tech stack
The technology behind every agent.
No black box. This is what MindLab runs on, layer by layer: the models you can choose from, the platform that executes the agents, the vision that runs on your own machine and the infrastructure that keeps it all online.
Models
Three providers with a real adapter in the platform. Each agent picks one, and you can change it from the dashboard without redeploying.
OpenAI
GPT models: the default for most agents, and the engine behind real-time voice.
Anthropic
Claude models for long context and cases that need careful reasoning.
Google Gemini
Gemini for high volume and low latency, plus the open-weight Gemma models.
Voice
Speech-to-speech on the phone, in the widget and in the branch avatar.
OpenAI Realtime
Real-time conversational voice over a persistent WebSocket session.
ElevenLabs
Neural text to speech and streaming transcription.
Gemini Live
The second live-voice engine, used by the avatar.
Web Speech API
The browser's own voice, for deployments that need zero cost per minute.
Computer vision
CP-Vision runs on the customer's own machine: frames are described locally and never leave it.
Python
Language of the vision service.
FastAPI
Its API and the WebSocket that streams descriptions frame by frame.
llama.cpp
Local inference of the vision-language model — no cloud call per frame.
MediaPipe
In-browser perception: people, gestures and face landmarks.
ONNX Runtime
Alternative perception backend, running on the GPU through WebGPU.
Branch avatar
The photorealistic MetaHuman that serves customers on a screen in the branch.
Unreal Engine
Renders the avatar and streams it as video to the browser.
WebRTC
Transport of Pixel Streaming: video, audio and input with no plugin.
Three.js
3D in the browser: the voice orb and the lightweight avatars.
Blender
Preparing and checking the avatar models.
Agent platform
The services that hold the agents, their tools, their memory and every conversation.
TypeScript
Every service is typed end to end, in strict mode.
Node.js
Runtime of the main API.
Bun
Runtime of the RAG service and the MCP servers.
NestJS
Framework of the API: modules, dependency injection and events.
Fastify
HTTP layer underneath, chosen for throughput.
Socket.IO
Live channel of the chat and the voice audio, scaled across instances with Redis.
BullMQ
Asynchronous work: campaigns, indexing, notifications and retries.
Model Context Protocol
How agents reach your systems — our own MCP servers for Shopify, Salesforce, Odoo, Genesys, HubSpot, Google Calendar and more.
Data and knowledge
Conversations, configuration and knowledge bases, isolated per customer.
PostgreSQL
Main database: agents, conversations, tenants and their configuration.
Redis
Cache, queues and the pub/sub that keeps every instance in sync.
Qdrant
Vector store of the knowledge bases: semantic search over your documents.
Neo4j
Knowledge graph: entities and relations for questions a plain search cannot answer.
Prisma
Schema and typed access to the database.
LangChain
Document splitting and the auxiliary chains of the RAG.
Interfaces
The chat widget, the dashboard and this site share one front-end stack.
React
Component model of every interface.
Vite
Build of every front end, including the widget's embeddable bundle.
Tailwind CSS
Design system: utilities and tokens instead of ad-hoc CSS.
Infrastructure
Containers on Kubernetes, described as code, deployed on every merge.
AWS
Where the platform runs: managed Kubernetes, databases, object storage and CDN.
Kubernetes
Orchestration of the backend services.
Terraform
The whole infrastructure as code, reviewable in a pull request.
Argo CD
GitOps: the cluster follows what the repository says it should run.
Docker
Every service ships as a container image.
GitHub Actions
CI/CD: tests, build and deploy, with no static credentials.
Cloudflare
DNS and the edge of this site, including its agent surface.
Observability and quality
What tells us something broke before a customer does.
OpenTelemetry
Distributed traces across services.
Grafana
Dashboards and trace storage.
Prometheus
Service metrics and alerting.
Sentry
Errors in the API and in the interfaces, with their stack trace.
Jest
Test suite of the platform API.
Vitest
Test suite of the front ends.
How we build
Choices we made on purpose
A stack is a set of trade-offs. These are ours, and why.
No model lock-in
Three providers with a real adapter and 48 selectable models. If a better model ships next month, you switch from the dashboard — the integration does not change.
We connect, we do not replace
The agent reaches the ERP, CRM or e-commerce you already run through MCP servers and APIs. Nothing has to be migrated to use MindLab.
Vision stays on your side
The vision model runs on the machine in the branch. Frames are described locally; what leaves is text, not video.
Boring where it matters
PostgreSQL, containers, infrastructure as code and typed services. The interesting part should be the agent, not the plumbing under it.
Next step
Want the details?
In a demo we walk your team through the architecture: integrations, data flows and what a deployment looks like on your side.