What TalkIDE is
TalkIDE is a platform where non-technical users build, run and publish their own web applications by talking to an AI Product Manager called Mara. The user describes what they want in natural language; Mara orchestrates a virtual team of Claude-powered subagents (backend developer, frontend developer, devops, tester, analyst) who write, test and deploy the code on the user’s behalf. There is no IDE to install, no infrastructure to configure, no code to read — the conversation is the product.
Each project lives in two parallel environments at once: a DEV preview under a generated
<uuid>.talkide.app URL that re-builds with every change Mara makes, and a PROD published
deployment under a stable, user-chosen <slug>.talkide.app URL that only changes when the
user explicitly clicks Publish. Both environments run side by side, each with its own
isolated Postgres database, in the same Kubernetes cluster.
TalkIDE itself is a multi-tenant SaaS application (Kotlin/Spring Boot backend, Vue 3 frontend, the Anthropic Agent SDK runtime, hosted on DigitalOcean Kubernetes in NYC3). It runs on two physically separate Managed Postgres clusters — a control-plane cluster for the platform DB and a data-plane cluster with schema-per-app for user apps — plus a self-hosted NFS for per-project working trees and DigitalOcean Spaces for blobs. The platform’s revenue model is AI token usage (prepaid credit) + hosting fees (postpaid); billing is live in test mode via Stripe.
System Context
The actors are an End User (vibecoder building their own app), a Platform Operator (TalkIDE staff running the platform — capacity dashboards, invite cascades, emergency brakes, issue triage, waitlist invites) and App End Users (the audience of a published user app). External systems TalkIDE depends on:
- Anthropic API — every Mara turn ultimately calls Claude; the platform gateway-proxy holds the API key.
- DigitalOcean — Kubernetes (compute), two Managed Postgres clusters (control-plane + data-plane), Spaces (blob storage), Container Registry, DNS, Load Balancer.
- Stripe — payment method on file, invoices, AI credit top-up, postpaid hosting charges (test mode today).
- Mailgun — transactional email: password reset, waitlist confirmation, invite emails (ADR-025).
- GitLab — source repositories and manual-gated CI for the platform itself.
- End user’s browser — both the TalkIDE app itself (
talkide.app) and the user’s generated apps (<uuid>.talkide.app,<slug>.talkide.app) are served to the same browser.
For the annotated picture see system-context.md; for the physical
DigitalOcean topology see infrastructure.md.
Where to read what — reading map
| You want to … | Read |
|---|---|
| Understand TalkIDE in 5 minutes | This page + glossary |
| See what runs where in the cluster | containers.md + specification/hosting-architecture.md |
| See the full DigitalOcean infra topology | infrastructure.md |
| Understand a specific UC group (e.g. project management) | components/uc-03-project-management.md → linked UC files |
| Know the contract of a specific endpoint | The relevant UC file under usecases/ |
| Know why a non-obvious decision was made | adr/ (platform) or specification/architecture.md (user-app) |
| Understand Mara, FUP, fatigue, quotas | specification/mara-fatigue-fup.md + ADR-020 |
| Understand how a user app is built and deployed | cross-cutting-flows.md § Deploy lifecycle + ADR-018/019/021 |
| Understand the entity model | model/README.md |
Status of this overview layer
Overview refreshed 2026-05-22 to reflect the two-cluster data-plane split (ADR-023), the worker extraction target (ADR-024), Stripe billing (UC-10), Mailgun (ADR-025) and the first-class Environment concept (ADR-026).
Diagrams are inline Mermaid rendered by the mermaid2 plugin — they version alongside
the prose, no build step. The legacy Structurizr DSL workflow (pre-rendered SVGs +
overview/workspace.dsl) was retired on 2026-05-23; see infrastructure.md § 6.
Thanks for the feedback.