AI Workspace

Under active development. The repository is private for now, so there's no public link — what's described here is the project's current state.

The problem#

Every hosted AI chat platform charges the same hidden price: your context. The conversations, the documents you upload, the integration keys and the memory the model builds up about you all live on someone else's infrastructure.

AI Workspace inverts that. It runs entirely in Docker Compose on your machine or your VPS, and the only outbound traffic is the call to the model provider and whatever you tell the tools to reach. Secrets are stored encrypted at rest in the database.

It isn't a chat wrapper. It's the whole stack.

What's inside#

Chat and models#

Multi-model through OpenRouter — any OpenAI-compatible model — plus local models via Ollama. Streaming is resumable: closing the tab or hitting F5 doesn't cancel the generation.

Each custom model carries its own prompt, parameters, tools, capabilities, filters, voice, memory and sub-agents. In practice, it's building your own "GPT".

Two modes step outside linear chat: the round table, with several models talking to each other while you steer, and sub-agents, where an orchestrator delegates to worker models sequentially or in parallel.

Tools#

Tool-calling runs on SIFT — the other project of mine in this list. Web search, page reading, an AI-driven headless browser, charts, diagrams, quotes, video and audio transcription, and Python code written by the AI itself, executed in an isolated sandbox with CPU and memory limits.

Memory and knowledge#

Long-term memory through mem0, scoped globally, per model or per chat, with shareable stores. Alongside it, a knowledge base with RAG over pgvector, answering with source citations.

The second brain keeps interlinked notes in a graph, and a background review proposes new memories and skills — always going through your approval first.

Automation and channels#

Schedules and monitors for price, page, search and RSS, notifying you in the app, via Web Push or through the channels. You talk to your models over WhatsApp, Telegram and Discord, and each conversation becomes a chat in the sidebar.

Platform#

An OpenAI-compatible public API at /v1/chat/completions, with per-user keys, limits, quotas, cost control and per-key memory. Every request becomes a trace carrying latency, database time, reads and writes and LLM calls, in a waterfall panel.

Architecture#

A monorepo with four pieces:

Postgres 16 with pgvector holds data and vectors in the same place — a deliberate call, to avoid keeping a separate vector database alive.

Heavy features come up on demand through Compose profiles: self-hosted search with SearXNG, local voice with Kokoro, WhatsApp through Evolution, headless browsing with browserless. If you don't use them, you don't pay for having them running.

Security#

The whole point is trusting it with sensitive data, so this couldn't be deferred:

• Passwords with Argon2; sessions as a JWT in an httpOnly cookie, with a rotating refresh and a token_version to revoke every session at once; optional 2FA (TOTP).
• Per-user secrets encrypted with Fernet, keyed from APP_SECRET, plus a rotation tool that re-encrypts the entire database.
• Restricted CORS, security headers on every response, HSTS under HTTPS, and rate limiting on login, registration and the public API.
• Under APP_ENV=production, the server refuses to start with a weak APP_SECRET.

Current state#

Working end to end, with a Windows desktop app built by CI on every release. Migrations run automatically when the server starts. What's left is the decision I haven't made yet: which license to adopt before opening the repository.