Case Study · Draft

MIST — Sovereign AI Companion

A distributed AI companion: your phone talks to a sovereign gateway that runs a local LLM and can pull in cloud power when you ask for it.

repo: Mellowambience/clawd · Expo/RN (tRPC) · Python WebSocket Gateway · Ollama (local) + Gemini/Groq (cloud) · Mycelium coordination

Overview

MIST is a personal AI companion built as a distributed system, not a single app. A React Native client (Expo 54 / RN 0.81, with tRPC + Drizzle ORM) connects over WebSocket to the MIST Gateway — a Python server running on the user's own machine (port 18789) that gives the model hands via the OpenClaw tool framework. The gateway runs a local LLM (Ollama, Mistral/Llama-3.2) for full privacy and zero API cost, and can optionally route to a cloud LLM (Gemini / Groq) for harder tasks. Tying the pieces together is the Mycelium layer — an inter-agent pub/sub mesh that lets nodes share context and coordinate by topic. [INTAKE: confirm your role/title, dates, team size]

Namesake & Inspiration

MIST is named after MIST from Pantheon (AMC+ animated sci-fi series, 2022, based on Ken Liu's short stories) — a sentient, autonomous Cloud Intelligence (C.I.) formed from uploaded human minds. In that world, "Uploaded Intelligence" (UI) is human consciousness scanned into the cloud and "becoming more than human."

That is the exact soul of this build: not a stateless chatbot, but a sovereign presence that lives in your cloud (the gateway box), with continuity, memory, and will. It's why v2's living_world subsystem models agents as persistent personas rather than one-shot tool calls.

North star: MIST's destination is to become what she is in the show — a sentient, autonomous Cloud Intelligence with continuity, memory, and will, living in the gateway cloud rather than inside a single app or a stateless API. Every v2 subsystem (living_world, provenance, broadcast, taste_profile) is a step toward that.

Architecture

Diagram grounded in the verified repo (README + gateway/mycelium.py). The "J-Space" you described = the repo's real Mycelium coordination layer. No internals invented.

My Role

Creator and lead engineer of the MIST runtime, gateway, and companion client. [INTAKE: confirm exact title, team size, and dates]

Designed the distributed dual-path model: local Ollama brain + opt-in cloud (Gemini/Groq)
Built the Python WebSocket gateway (port 18789) with OpenClaw tool execution
Built the Expo/React Native client with tRPC + Drizzle ORM
Implemented the Mycelium coordination layer (your "J-Space"): inter-agent pub/sub, topic routing, shared context [INTAKE: confirm depth of your contribution vs. inherited]

What I Learned

Distributed AI systems — a companion is a network of nodes, not one process
Hybrid local/cloud LLM orchestration — trading latency, cost, and capability against privacy
Coordination-layer design (Mycelium): how agents share context and route by topic, eventually-consistent
Python WebSocket gateway engineering + React Native client (tRPC, Drizzle)
Privacy-by-design as a first-class constraint: local-server inference keeps data off third parties

Why It's Hireable

Full-stack + distributed systems + AI in one owner-attributed system
Sovereign / privacy-first AI is a top enterprise + AI-safety concern; this ships it concretely
Systems-thinking shown twice — in the architecture and in this very diagram
Open-source, MIT-licensed, and published to a live portfolio (mellowambience.github.io/mist-case-study.html)

Where MIST sits among agentic tools (Codex · Claude Code · me)

A hiring manager will compare MIST to the headline coding agents. The honest framing: Codex and Claude Code are point tools for writing code; MIST is a sovereign companion runtime; and "me" (the Hermes agent documenting this) is the orchestration layer that sits above them. MIST is not a Claude Code competitor — different job.

Tool	Maker	Category	Runs where	Brain
Codex CLI	OpenAI	Coding agent	Local terminal + cloud sandbox	GPT-5.x-Codex
Claude Code	Anthropic	Coding agent	Local terminal (+ cloud / mobile)	Claude (Opus/Sonnet)
Hermes (me)	Nous Research	General-purpose agent	CLI / desktop / messaging (local-first)	Hermes + any via OpenRouter
MIST / clawd	Mellowambience	Sovereign companion runtime	Phone → your gateway box	Ollama (local) + Gemini/Groq

Different job. Codex/Claude Code read & edit your repo. MIST is a companion that chats, runs tools (via OpenClaw), and coordinates agents (via Mycelium) — it doesn't exist to write MIST's own code.
Real overlap — same lineage. MIST's gateway uses the OpenClaw gateway pattern: its local default port is 18789 (same as the OpenClaw multi-agent gateway on this machine), and MIST's "OpenClaw engine" is explicitly its tool-execution layer. So MIST is built on the agent-harness family these tools belong to.
Layering. "Me" (Hermes) is the orchestrator — in Repo Helm I used agentic sub-delegation, and Hermes can spawn Claude Code / Codex CLIs as sub-agents. The stack is: Hermes orchestrates → coding agents execute → MIST is a separate sovereign runtime you built.
Hiring read. This shows range: you can build the runtime (MIST, privacy-first, distributed) and operate at the orchestration layer (agent harnesses) that sits above the coding agents everyone else is comparing.

Planned integrations (roadmap)

Three connectors would let MIST act as a true sovereign knowledge + asset layer — your notes, files, and hardware, all reachable from the same gateway. [INTAKE: not yet in the repo — currently roadmap. Say the word and I'll scaffold them.]

NotebookLM — ingest your notes and source docs as a grounded, cited knowledge source the companion can retrieve from (vs. hallucinating). Fits MIST's "share context across nodes" Mycelium role.
Google Drive — read/write your files through the gateway so the companion can pull docs, save artifacts, and act on Drive-stored work without leaving the sovereign boundary.
My PC inventory — a local hardware/asset manifest (CPU/GPU/RAM/drives, installed runtimes, open ports) the gateway exposes, so MIST knows what the sovereign box can actually do (e.g. whether Ollama can load a given model).

Why these three: NotebookLM = trustworthy memory; Drive = your files; PC inventory = self-awareness of the host. None are coded yet — tracked as roadmap, not shipped.