Cesar Petrescu

Work

Adastrion Core

Self-hosted chat and coding agents with persistent projects and isolated workspaces

A workspace for conversations and code

I built Adastrion Core to bring conversations, code, browser tools and project context into one self-hosted web application. It runs Pi and DeepSeek Harness agents through a common interface, with persistent workspaces, shared project files and explicit controls for long-running work.

Open Adastrion Core ↗

  • React
  • TypeScript
  • Fastify
  • PostgreSQL
  • pgvector
  • Docker
  • LiteLLM
  • Pi
  • DeepSeek Harness

Why it exists

A coding conversation needs more than a chat history: it needs the files, browser state, tools and decisions that make the work reproducible. Core keeps those together while allowing each agent to work inside its own sandbox.

Components

Chat and agent modes

  • Chat provides streaming conversations, attachments, memory and connected MCP tools.
  • Agent mode adds an isolated Docker workspace with file tools, Bash, browser access and an embedded VS Code editor.
  • Pi and DeepSeek Harness retain their native execution and history while sharing the same application interface.

Projects and persistent context

  • Projects hold instructions, shared uploads, conversations and membership controls.
  • Workspaces and native session histories survive sandbox sleep, so work can resume without reconstructing the project.
  • Persistent memory and Ledger integration carry useful context between tasks.

Browser and editor

  • Chrome supports testing applications and local previews; Camoufox supports research with temporary browser profiles.
  • An embedded VS Code workbench and terminal operate on the agent’s workspace files.
  • The editor uses a separate origin and authenticated handoff to keep workspace code separate from the main application.

Parallel agents

  • A project coordinator delegates tasks to workers using shared files or isolated workspace snapshots.
  • Workers exchange durable messages; owners can inspect progress, answer questions and stop work.
  • Integration compares worker changes with the original and current files. Conflicting edits require review before they are combined.

Goals and automation

  • Persistent goals track progress and usage, with pause, resume and stop controls.
  • Schedules and scoped webhooks start repeatable work.
  • Ledger can dispatch research to Core; results return for review before acceptance and publication.

How it flows

  1. Open a project and select a chat or coding-agent conversation.
  2. Give the agent a task with project instructions, reference files and permitted tools.
  3. Inspect streamed work, browser previews and file changes; answer questions as needed.
  4. Review and integrate worker results, resolve conflicts and verify the combined output.

Production notes

Deployed Core
The chat and coding platform is deployed at chat.adastrion.com. This description reflects the October 2026 project inventory; DeepSeek Harness support remains alpha.
Earlier Adastrion work
The earlier Audio and Agent projects explored speech-to-speech, Flutter Android and Wear OS clients, and planner/executor/checker workflows. Those historical features should not be read as features shipped in the current Core.
Next work
Voice integration, native mobile and desktop applications, public sharing and a general dev-server proxy remain planned. The older voice-to-production concept is not the current product description.

Security boundaries

  • Agent sandboxes run without root privileges and have network isolation from the host and other sandboxes.
  • Project membership governs access to shared files, conversations and tools.
  • Stop controls, durable run state and explicit resumption keep long-running tasks under owner control.

Outcome

A deployed workspace for self-hosted chat, coding and research, with persistent context, isolated execution and reviewable collaboration between agents.