Running an entire AI workflow from a phone with a shared MCP memory layer

Asly97 · reddit · 2026-09-14

A developer shares how he replaced copy-pasting context between ChatGPT, Claude Code, and Codex with one phone-controlled system.

Architecture

Shared memory: Vilix AI exposes an MCP server; every tool connects via OAuth or API key and reads/writes the same memory (getcontext / saveturn). Retrieval is semantic RAG over a vector DB, storing conversations, projects, tasks, personal rules, and reusable skills.
Skills as memory: workflows an AI figures out (e.g., a video editing pipeline) get saved as skills, so every connected tool can reuse them instead of re-learning.
Integrations: Gmail, Notion, and social accounts connect once through Composio, which centralizes all MCP tool connections.

In practice

A 24/7 agent setup ("Hermes") runs on an Azure VM where Claude Code and Codex build; he plans tasks from his phone, and the builder picks up full context — a website change requested from his phone went live in minutes.
Content pipeline: record video on phone, Claude Code edits and renders it with Remotion following the saved skill, Canva handles design, and scheduled overnight jobs surface content ideas and analytics reports.

More from coding & agent