DotSpace
An original Android-first AI agent workspace: specialist agents, persistent conversations, spaces, background tasks, approvals and memory.
- Kotlin
- Jetpack Compose
- Node.js
- Fastify
- TypeScript
- SQLite
Overview
DotSpace is an AI agent workspace designed from scratch: specialist AI agents (called Dots), persistent conversations, Spaces and Pages for organising work, background tasks with approvals, memory, voice, and MCP integrations.
Version 0.1.0 is complete as a working system: a Node 24 + Fastify + TypeScript backend with SQLite, and a native Kotlin + Jetpack Compose Android app. The backend runs on a laptop on the local network, and the app talks to it there.
Problem / Goal
General chatbots forget context, cannot run work in the background, and give the user no control over actions that matter. DotSpace explores the opposite shape: agents that remember, tasks that run with explicit approvals, and a workspace that keeps conversations and files organised by space.
Approach
- Monorepo with a typed Fastify backend (28 of 28 backend tests passing at v0.1.0) and a native Compose app (29 of 29 app tests passing).
- Approvals as a first-class concept: consequential actions pause for the user instead of executing silently.
- Honest integration boundaries: features that are not built yet (realtime voice, computer screen access, chat attachments) are reported as gaps, not hidden.
Selected features
- Specialist agents (Dots) with persistent conversationsBuilt
- Spaces and Pages for organising workBuilt
- Background tasks with an approvals flowBuilt
- Memory across conversationsBuilt
- Debug APK for local testing (13.6 MB)Built
- Always-on hosted deployment (Docker assets prepared, not yet live)In progress
- Realtime voicePlanned
Constraints & challenges
Deployment needs an always-on host
The backend is a long-running server with SQLite, so static and serverless hosts do not fit. A Dockerfile, docker-compose setup and deployment guide were prepared, but the hosted deployment waits on an always-on machine rather than being faked on the wrong platform.
Keeping the scope honest
An agent workspace invites endless features. The v0.1.0 line was drawn at a tested core (agents, spaces, tasks, approvals, memory), with the missing pieces listed plainly in the delivery notes.
Outcome & lessons
- 01Approvals are a product feature, not a permission dialog: they are where trust is built.
- 02A tested core beats a broad demo: 57 passing tests made v0.1.0 a system, not a sketch.
- 03Match the host to the architecture instead of bending the architecture to a free tier.