Personal AI OS tools: choose by layer, not hype

Updated May 15, 2026.

Do not pick a Personal AI OS by GitHub stars. Pick it by layer.

A Personal OS is the control plane for memory, identity, permissions, tools, workflows, actions, receipts, and rollback. No single project owns that whole stack yet. ChatGPT, Claude, Gemini, Copilot, Cursor, Open WebUI, Khoj, Letta, mem0, Open Interpreter, n8n, Composio, browser-use, OpenHands, Home Assistant, and Hermes all attack different layers.

The useful question is not “which one is best?” It is: which layer are you missing right now, and where will the receipts live?

One-screen chooser

choose the missing layer receipts beat hype

Phone or desktop context

vendor rails

Gemini Intelligence on Android, Apple Intelligence, Microsoft Recall

replaceable start

Home Assistant voice, local scripts, Hermes-style local runtime

caveatGreat context, but policy and memory live on vendor rails

receiptSetting changed, device action log, screenshot

Chat workbench

vendor rails

ChatGPT, Claude, Gemini, Copilot

replaceable start

Open WebUI, Khoj, local model frontends

caveatA chat UI is a surface, not custody

receiptConversation export, prompt/config file

Durable memory

vendor rails

ChatGPT memory, Claude project memory

replaceable start

mem0, Letta, Markdown, SQLite, Git

caveatSeparate memory from retrieval; keep corrections and expiry

receiptMemory diff, source file, correction log

Knowledge retrieval

vendor rails

Vendor file search and project uploads

replaceable start

Quivr, LlamaIndex, Haystack, memory system guide

caveatRetrieval can cite documents; it does not decide what should be remembered

receiptSource document, index config, query trace

Tool connectors

vendor rails

OpenAI Apps SDK, OpenAI MCP connectors, Claude MCP

replaceable start

MCP servers, MCP server stacking, Composio/Rube

caveatConnector sprawl is account-sprawl with better branding

receiptScopes, token inventory, revocation path

Browser action

vendor rails

Operator, ChatGPT agent, Claude computer use

replaceable start

Playwright MCP, browser-use, Skyvern, Browserbase

caveatBrowser automation is an action layer, not a security boundary

receiptScreenshots, DOM/log trace, confirmation screen

Coding workspace

vendor rails

Claude Code, Codex, Cursor, Windsurf, Copilot agent mode

replaceable start

AI coding assistants, OpenHands, Superpowers, skills

caveatCoding has good receipts; life-admin actions often do not

receiptGit diff, test log, commit, PR

Workflow daemon

vendor rails

Copilot Actions, scheduled assistants, app automations

replaceable start

n8n, Flowise, AutoGPT, CrewAI, LangGraph

caveatBackground autonomy needs logs before cleverness

receiptRun ID, inputs, outputs, error log

Local/private interface

vendor rails

Vendor desktop apps

replaceable start

Open WebUI + Ollama/local model, Open Interpreter, 01, Home Assistant voice

caveatLocal does not automatically mean safe; tools still need permissions

receiptConfig file, command log, backup

User-owned network

vendor rails

Vendor account graph

replaceable start

Hermes Agent, network.self.md-style encrypted state

caveatEmerging infrastructure; do not treat it like polished SaaS

receiptState export, peer/member log, approval record

If you cannot name the receipt for a layer, you are not choosing a Personal OS. You are choosing a place to lose track of side effects.

Personal OS registry

This is the seed map, not a “top AI tools” graveyard. Every row has to answer four questions: what layer does it replace, who owns the state, where does it run, and where do the receipts live?

The public registry starts with the verified core seed. The verify-first pile stays off the page until the source, custody model, and receipt story are checked again.

Coverage labels are editorial plumbing, not scores: exact means a dedicated self.md page exists, related means the row is covered by a broader page, backlog means a canonical page is planned later, and registry-only means the row stays here for now.

87core seed entries

36user-owned

15hybrid

36vendor-owned

11exact pages

6related coverage

1backlog

69registry only

Search Layer Custody Runs Coverage

87 entries visible. Coverage labels mean exact page, related coverage, backlog, or registry-only. Core seed only; verify-first rows stay out of the public UI until checked.

Vendor Suite 5 entries

Apple Intelligence

vendor-owned registry only

Vendor Suite mixed high

iPhone/Mac context

layer: interface/surface
receipts: device logs, cloud logs

caveatstate stays on Apple rails; export is limited

official

Listed in the registry. No canonical self.md page yet.

ChatGPT / Operator / ChatGPT agent

vendor-owned registry only

Vendor Suite SaaS high

chat + browser tasks

layer: action surfaces
receipts: cloud logs, transcripts, screenshots

caveatdurable state stays on OpenAI rails

official

Listed in the registry. No canonical self.md page yet.

Claude / Claude Desktop / Skills

vendor-owned

Vendor Suite SaaS high

chat + coding + computer use

layer: tools/connectors
receipts: cloud logs, transcripts

caveatcustody stays Anthropic-owned

related note official

Related Claude skills/tools coverage; no single Claude suite page yet.

Google Gemini on Android

vendor-owned registry only

Vendor Suite mixed high

Android device context

layer: interface/surface
receipts: cloud logs, device history

caveatworks best when you already live on Google rails

official

Listed in the registry. No canonical self.md page yet.

Microsoft Copilot Actions

vendor-owned registry only

Vendor Suite mixed high

background tasks in Windows

layer: orchestration/runtime
receipts: task traces, cloud logs

caveatrollout and policy are still changing

official

Listed in the registry. No canonical self.md page yet.

Local Workbench 11 entries

AnythingLLM

user-owned registry only

Local Workbench mixed medium

doc-grounded personal workspace with RAG

layer: retrieval/index
receipts: local DB, vector store, files

caveatreal value is retrieval/plumbing, not just chat

official

Listed in the registry. No canonical self.md page yet.

GPT4All

user-owned registry only

Local Workbench local medium

simple local-model starter kit

layer: interface/surface
receipts: local app data, models, chat history

caveatmore bundle than a full OS layer

official

Listed in the registry. No canonical self.md page yet.

Jan

user-owned registry only

Local Workbench local medium

local-first desktop assistant with flexible model backends

layer: interface/surface
receipts: local app data, chat history, cache

caveatecosystem and extension surface move fast

official

Listed in the registry. No canonical self.md page yet.

LibreChat

user-owned registry only

Local Workbench self-hosted medium

self-hosted ChatGPT-style multi-provider chat

layer: interface/surface
receipts: local DB, attachments, logs

caveatauth/connectors surface grows fast; governance is required

official

Listed in the registry. No canonical self.md page yet.

LM Studio

user-owned backlog

Local Workbench local medium

local model playground and desktop-server gateway

layer: interface/surface
receipts: local app data, model cache, transcripts

caveatproprietary app; assistant story is model-first, not state-first

related note official

Related local-runtime coverage exists; canonical LM Studio page is backlog.

LobeChat

user-owned registry only

Local Workbench mixed medium

polished consumer-grade chat UI

layer: interface/surface
receipts: local storage, sync logs, exports

caveatverify state/export portability separately

official

Listed in the registry. No canonical self.md page yet.

Logseq

user-owned registry only

Local Workbench local medium

daily notes and local knowledge graph

layer: memory/state
receipts: local markdown/org files, git, sync logs

caveatsync/plugin ecosystem can be uneven

official

Listed in the registry. No canonical self.md page yet.

Obsidian

user-owned registry only

Local Workbench local medium

durable local vaults and manual knowledge curation

layer: memory/state
receipts: local markdown files, git, sync logs

caveatAI is add-on, not core memory engine

official

Listed in the registry. No canonical self.md page yet.

Open WebUI

user-owned exact

Local Workbench self-hosted medium

self-hosted chat hub for multiple backends

layer: interface/surface
receipts: local DB, exports, logs

caveatUI is portable, but durable state lives where the backend/storage lives

self.md page official

Dedicated self.md tool page.

Raycast AI

hybrid registry only

Local Workbench mixed high

command palette with personal device context

layer: interface/surface
receipts: local config, command history

caveatdeep state fragments across extensions

official

Listed in the registry. No canonical self.md page yet.

01

user-owned registry only

Local Workbench local medium

local personal-assistant shell

layer: interface/surface
receipts: local transcripts, config, files

caveatsibling project; fast-moving, verify the current release cadence

official

Listed in the registry. No canonical self.md page yet.

Memory 10 entries

Graphiti

user-owned registry only

Memory self-hosted medium

temporal knowledge graph memory for agents

layer: memory/state
receipts: graph store, database, logs

caveatgraph/schema discipline adds complexity

official

Listed in the registry. No canonical self.md page yet.

Letta / MemGPT

user-owned registry only

Memory mixed medium

long-lived agent state and memory-centric agents

layer: memory/state
receipts: database, memory store, logs

caveatarchitecture-heavy; still evolving

official

Listed in the registry. No canonical self.md page yet.

Mem.ai

vendor-owned registry only

Memory SaaS medium

AI-native notes and context recall

layer: memory/state
receipts: cloud workspace data, exports

caveatmigration/export path is the main risk

official

Listed in the registry. No canonical self.md page yet.

mem0

user-owned exact

Memory mixed medium

portable memory layer for agents

layer: memory/state
receipts: database, vector store, logs

caveatmemory primitive, not a full OS

self.md page official

Dedicated self.md tool page.

MemU

hybrid exact

Memory mixed medium

agent memory API / personal memory buffer

layer: memory/state
receipts: database, vector store, logs

caveatpublic docs are thin; verify hosting/export model

self.md page official

Dedicated self.md tool page.

Pieces OS

user-owned registry only

Memory mixed high

developer memory and local context recall

layer: memory/state
receipts: local index, local files, sync logs

caveatnarrower than a full Personal AI OS

official

Listed in the registry. No canonical self.md page yet.

Reflect

vendor-owned registry only

Memory SaaS medium

private note capture and recall

layer: memory/state
receipts: notes, cloud sync, exports

caveataction surface is limited

official

Listed in the registry. No canonical self.md page yet.

Supermemory

vendor-owned registry only

Memory SaaS medium

async memory service for apps

layer: memory/state
receipts: database, vector store, logs

caveatearly product; vendor dependency is high

official

Listed in the registry. No canonical self.md page yet.

Tana

vendor-owned registry only

Memory SaaS medium

structured knowledge graph and tasks

layer: memory/state
receipts: cloud workspace data, exports

caveatvendor-held state and export path deserve scrutiny

official

Listed in the registry. No canonical self.md page yet.

Zep

hybrid registry only

Memory mixed medium

conversation memory service and retrieval over chat history

layer: memory/state
receipts: database, vector store, logs

caveatservice shape and hosting model need recheck

official

Listed in the registry. No canonical self.md page yet.

Retrieval 8 entries

Cognee

user-owned registry only

Retrieval self-hosted medium

graph-based knowledge engine and retrieval

layer: retrieval/index
receipts: database, graph store, vector store, logs

caveatgraph abstraction can be overkill for simple notes

official

Listed in the registry. No canonical self.md page yet.

Haystack

user-owned registry only

Retrieval mixed medium

production search/RAG pipelines

layer: retrieval/index
receipts: code/config, vector store, traces/logs

caveatpipeline complexity rises fast

official

Listed in the registry. No canonical self.md page yet.

Khoj

user-owned registry only

Retrieval self-hosted medium

personal search across notes/docs

layer: retrieval/index
receipts: local notes index, vector store, DB

caveatvalue depends heavily on source and indexing hygiene

official

Listed in the registry. No canonical self.md page yet.

LlamaIndex

user-owned registry only

Retrieval mixed medium

building RAG and knowledge workflows

layer: retrieval/index
receipts: code/config, vector store, traces/logs

caveatframework, not end-user memory product

official

Listed in the registry. No canonical self.md page yet.

Notion AI / Q&A

vendor-owned registry only

Retrieval SaaS high

Q&A over workspace notes

layer: memory/state
receipts: workspace data, cloud logs

caveatplatform-owned state and export limits

official

Listed in the registry. No canonical self.md page yet.

Perplexity

vendor-owned registry only

Retrieval SaaS high

answer engine with citations

layer: retrieval/index
receipts: cloud logs, citations

caveatnot a durable memory layer

official

Listed in the registry. No canonical self.md page yet.

Quivr

user-owned registry only

Retrieval self-hosted medium

self-hosted personal knowledge base over docs

layer: retrieval/index
receipts: local files, vector store, logs

caveatcloud/self-host split and UX maturity should be checked

official

Listed in the registry. No canonical self.md page yet.

Saner.AI

vendor-owned exact

Retrieval SaaS medium

AI search/organizer over scattered knowledge

layer: retrieval/index
receipts: cloud workspace data, exports, logs

caveatproduct shape and export semantics are fast-moving

self.md page official

Dedicated self.md tool page.

Tool Connectors 11 entries

App Intents

vendor-owned registry only

Tool Connectors mixed high

system app actions

layer: tools/connectors
receipts: app logs, device state

caveatonly works with what the app exposes

official

Listed in the registry. No canonical self.md page yet.

Arcade.dev

vendor-owned registry only

Tool Connectors SaaS high

secure tool calling/auth layer for agents

layer: identity & auth
receipts: auth logs, tool call logs

caveatvendor auth custody

official

Listed in the registry. No canonical self.md page yet.

Composio

vendor-owned exact

Tool Connectors SaaS high

managed app action connectors

layer: tools/connectors
receipts: tool call logs, scopes

caveatauth custody sits with platform

self.md page official

Dedicated self.md tool page.

Glama MCP

hybrid

Tool Connectors SaaS high

MCP server directory

layer: tools/connectors
receipts: registry metadata

caveatdirectory is not custody

related note official

Covered by the MCP servers connector-layer page; no Glama-specific page yet.

LiteLLM

user-owned registry only

Tool Connectors self-hosted medium

provider abstraction, routing, and cost control

layer: tools/connectors
receipts: proxy logs, cost logs, db, traces

caveatcan become a hidden policy chokepoint

official

Listed in the registry. No canonical self.md page yet.

mcp.so

hybrid

Tool Connectors SaaS high

MCP server discovery

layer: tools/connectors
receipts: registry metadata

caveatrequires per-server verification

related note official

Covered by the MCP servers connector-layer page; no mcp.so-specific page yet.

OpenAI Apps SDK

vendor-owned registry only

Tool Connectors SaaS high

building apps inside ChatGPT

layer: tools/connectors
receipts: app/tool logs, ChatGPT context

caveatstrongly ChatGPT rail

official

Listed in the registry. No canonical self.md page yet.

OpenAI MCP connectors

vendor-owned

Tool Connectors SaaS high

connecting tools to OpenAI agents

layer: tools/connectors
receipts: connector logs, scopes

caveatvendor policy/runtime rail

related note official

Covered by the MCP servers connector-layer page; no OpenAI-specific connector page yet.

Pipedream

vendor-owned registry only

Tool Connectors SaaS high

API/event connector layer

layer: tools/connectors
receipts: workflow run logs, event payloads

caveatSaaS custody unless self export is enough

official

Listed in the registry. No canonical self.md page yet.

Rube

vendor-owned

Tool Connectors SaaS high

Composio MCP/action rail

layer: tools/connectors
receipts: MCP call logs, scopes

caveatvendor connector rail

related note official

Related Composio coverage; no Rube-specific page yet.

Smithery

hybrid

Tool Connectors SaaS/mixed high

MCP server discovery and install

layer: tools/connectors
receipts: server listings, config, install logs

caveatregistry quality varies

related note official

Covered by the MCP servers connector-layer page; no Smithery-specific page yet.

Browser Action 5 entries

Bardeen

vendor-owned registry only

Browser Action SaaS/browser high

browser productivity automations

layer: action surfaces
receipts: automation history, browser state

caveatextension permissions matter

official

Listed in the registry. No canonical self.md page yet.

Browserbase

vendor-owned registry only

Browser Action SaaS high

managed browser infra for agents

layer: action surfaces
receipts: session recordings, logs, traces

caveatvendor browser rail

official

Listed in the registry. No canonical self.md page yet.

browserless

hybrid registry only

Browser Action mixed high

headless browser infrastructure

layer: action surfaces
receipts: browser session logs, traces

caveatnot an agent by itself

official

Listed in the registry. No canonical self.md page yet.

Hyperbrowser

vendor-owned registry only

Browser Action SaaS high

cloud browsers and web agents

layer: action surfaces
receipts: session logs, screenshots

caveatvendor rail for browser state

official

Listed in the registry. No canonical self.md page yet.

Steel.dev

hybrid registry only

Browser Action mixed high

browser infra for agents

layer: action surfaces
receipts: browser sessions, screenshots, logs

caveatverify open-source/cloud boundary

official

Listed in the registry. No canonical self.md page yet.

Coding Agent 8 entries

Claude Code

hybrid exact

Coding Agent mixed high

repo-level agent work with strong receipts

layer: action surfaces
receipts: git diff, terminal logs, transcripts

caveatvendor model rail; memory/policy still needs project files

self.md page official

Dedicated self.md tool page.

Cursor

hybrid exact

Coding Agent mixed high

AI-native code editor

layer: interface/surface
receipts: repo diffs, chat history, editor state

caveateditor lock-in and cloud context deserve scrutiny

self.md page official

Dedicated self.md tool page.

Devin

vendor-owned registry only

Coding Agent SaaS high

hosted software engineering agent

layer: orchestration/runtime
receipts: cloud workspace logs, PRs

caveathigh vendor custody

official

Listed in the registry. No canonical self.md page yet.

Open Interpreter

user-owned registry only

Coding Agent local medium

local shell/browser assistant with real side effects

layer: action surfaces
receipts: command logs, screenshots, files

caveatextremely powerful; dangerous without sandbox and approval gates

official

Listed in the registry. No canonical self.md page yet.

Replit Agent

vendor-owned registry only

Coding Agent SaaS high

app generation inside Replit workspace

layer: orchestration/runtime
receipts: workspace diffs, deploy logs

caveatbest inside Replit rail

official

Listed in the registry. No canonical self.md page yet.

Sourcegraph Amp

hybrid registry only

Coding Agent mixed high

agentic coding workflow

layer: action surfaces
receipts: agent task logs, git diffs

caveatfast-moving; verify features per current docs

official

Listed in the registry. No canonical self.md page yet.

Sourcegraph Cody

hybrid registry only

Coding Agent mixed high

large-codebase context and code search

layer: retrieval/index
receipts: repo search logs, chat, enterprise logs

caveatless a personal OS, more code intelligence layer

official

Listed in the registry. No canonical self.md page yet.

Windsurf

hybrid exact

Coding Agent mixed high

agentic IDE workflows

layer: interface/surface
receipts: repo diffs, editor logs

caveatfast-moving vendor rail

self.md page official

Dedicated self.md tool page.

Workflow Daemon 11 entries

Dify

user-owned registry only

Workflow Daemon self-hosted medium

self-hosted app builder and workflow-backed chat apps

layer: orchestration/runtime
receipts: local DB, workflow logs, traces

caveatplatform holds state and complicates lifecycle control

official

Listed in the registry. No canonical self.md page yet.

Dust

vendor-owned registry only

Workflow Daemon SaaS high

company assistants over tools/data

layer: orchestration/runtime
receipts: assistant logs, data source logs

caveatworkspace/vendor custody

official

Listed in the registry. No canonical self.md page yet.

Flowise

user-owned registry only

Workflow Daemon self-hosted medium

quick self-hosted LLM workflow composition

layer: orchestration/runtime
receipts: flow JSON, local DB, logs

caveatgovernance and tracing are thin by default

official

Listed in the registry. No canonical self.md page yet.

Gumloop

vendor-owned registry only

Workflow Daemon SaaS high

AI workflow builder

layer: orchestration/runtime
receipts: flow runs, logs

caveatvendor flow state

official

Listed in the registry. No canonical self.md page yet.

Langflow

user-owned registry only

Workflow Daemon self-hosted medium

visual prototyping of LLM flows and agent graphs

layer: orchestration/runtime
receipts: flow configs, local DB, logs

caveatbetter as a builder than as a final Personal OS runtime

official

Listed in the registry. No canonical self.md page yet.

Lindy

vendor-owned registry only

Workflow Daemon SaaS high

business assistant workflows

layer: orchestration/runtime
receipts: agent runs, app logs

caveatopaque vendor autonomy if ungated

official

Listed in the registry. No canonical self.md page yet.

MindStudio

vendor-owned registry only

Workflow Daemon SaaS high

build/deploy AI workflow apps

layer: orchestration/runtime
receipts: app runs, logs

caveatplatform lock-in risk

official

Listed in the registry. No canonical self.md page yet.

n8n

hybrid exact

Workflow Daemon self-hosted/SaaS high

visible workflow daemons with logs

layer: orchestration/runtime
receipts: run history, payloads, error logs

caveatcan become spaghetti without gates

self.md page official

Dedicated self.md tool page.

Relay.app

vendor-owned registry only

Workflow Daemon SaaS high

human-in-the-loop workflows

layer: orchestration/runtime
receipts: workflow runs, approvals

caveatstill SaaS-owned state

official

Listed in the registry. No canonical self.md page yet.

Stack AI

vendor-owned registry only

Workflow Daemon SaaS high

enterprise AI workflow builder

layer: orchestration/runtime
receipts: flow logs, data source logs

caveatless personal/user-owned

official

Listed in the registry. No canonical self.md page yet.

VectorShift

vendor-owned registry only

Workflow Daemon SaaS high

AI workflows and assistants

layer: orchestration/runtime
receipts: pipeline logs, data connectors

caveatvendor data custody

official

Listed in the registry. No canonical self.md page yet.

Agent Framework 2 entries

CrewAI

hybrid registry only

Agent Framework mixed high

multi-agent workflows

layer: orchestration/runtime
receipts: run logs, task traces

caveatcrew metaphors can hide poor state design

official

Listed in the registry. No canonical self.md page yet.

Pydantic AI

user-owned registry only

Agent Framework local/self-hosted high

typed Python agent apps

layer: orchestration/runtime
receipts: structured logs, eval traces

caveatdeveloper framework only

official

Listed in the registry. No canonical self.md page yet.

Smart Home Voice 4 entries

Home Assistant

user-owned registry only

Smart Home Voice self-hosted high

local control plane for home and devices

layer: orchestration/runtime
receipts: local logs, automations, device state

caveatvoice/AI pipeline depends on which components you connect

official

Listed in the registry. No canonical self.md page yet.

Home Assistant Assist

user-owned registry only

Smart Home Voice mixed high

smart-home voice control with local mode

layer: interface/surface
receipts: local logs, transcripts, automation traces

caveatquality depends heavily on the ASR/TTS/LLM pipeline

official

Listed in the registry. No canonical self.md page yet.

OpenVoiceOS

user-owned registry only

Smart Home Voice self-hosted high

open voice assistant stack

layer: interface/surface
receipts: local logs, transcripts, config

caveatsmaller and more DIY ecosystem than commercial rails

official

Listed in the registry. No canonical self.md page yet.

Wyoming Satellite

user-owned registry only

Smart Home Voice self-hosted high

edge microphone/speaker endpoint for a local voice stack

layer: interface/surface
receipts: audio buffers, logs, device state

caveatneeds a backend such as Home Assistant/Wyoming

official

Listed in the registry. No canonical self.md page yet.

Personal Data / Lifelog 5 entries

Granola

vendor-owned exact

Personal Data / Lifelog SaaS high

meeting capture and summaries

layer: receipts/observability
receipts: notes, meeting transcript logs

caveatnot a general OS

self.md page official

Dedicated self.md tool page.

Limitless

vendor-owned registry only

Personal Data / Lifelog mixed high

meeting and life capture

layer: receipts/observability
receipts: transcripts, cloud logs

caveatprivacy-sensitive and vendor-owned

official

Listed in the registry. No canonical self.md page yet.

Microsoft Recall

vendor-owned registry only

Personal Data / Lifelog mixed high

search over previous Windows screens

layer: receipts/observability
receipts: local screenshots, local index

caveatprivacy-sensitive, Windows-only

official

Listed in the registry. No canonical self.md page yet.

Plaud

vendor-owned registry only

Personal Data / Lifelog mixed high

hardware-first capture for calls/field notes

layer: receipts/observability
receipts: audio files, transcripts, cloud notes

caveatcapture device first; general OS use limited

official

Listed in the registry. No canonical self.md page yet.

Rewind

vendor-owned registry only

Personal Data / Lifelog mixed high

personal time-travel memory

layer: receipts/observability
receipts: screen/audio history, local cache, cloud logs

caveatprivacy-sensitive and vendor-owned

official

Listed in the registry. No canonical self.md page yet.

Network / Runtime 6 entries

Hermes Agent

user-owned registry only

Network / Runtime local high

local agent runtime with skills/tools/gateway

layer: orchestration/runtime
receipts: files, tool logs, scheduled outputs, git

caveatemerging; not polished SaaS

official

Listed in the registry. No canonical self.md page yet.

llama.cpp

user-owned registry only

Network / Runtime local medium

minimal portable inference runtime

layer: orchestration/runtime
receipts: CLI logs, model files, config

caveatthis is an engine, not a product

official

Listed in the registry. No canonical self.md page yet.

LocalAI

user-owned registry only

Network / Runtime self-hosted medium

OpenAI-compatible self-hosted inference

layer: orchestration/runtime
receipts: local DB/config, logs, model cache

caveatops surface is wider than the quick start makes it look

official

Listed in the registry. No canonical self.md page yet.

Ollama

user-owned exact

Network / Runtime local medium

simplest local model-serving rail

layer: orchestration/runtime
receipts: model cache, daemon logs, config

caveatno first-class chat UI or durable memory; needs a front-end

self.md page official

Dedicated self.md tool page.

vLLM

user-owned registry only

Network / Runtime self-hosted medium

high-throughput self-hosted serving

layer: orchestration/runtime
receipts: server logs, metrics, traces

caveatruntime only; UI, auth, and memory are separate

official

Listed in the registry. No canonical self.md page yet.

Wyoming protocol

user-owned registry only

Network / Runtime mixed high

transport rail between voice satellite and backend

layer: tools/connectors
receipts: local logs, stream traces

caveatplumbing, not a full assistant

official

Listed in the registry. No canonical self.md page yet.

Custom Stack 1 entries

Fabric

user-owned registry only

Custom Stack local medium

repeatable prompt chains over retrieved context

layer: custom-stack / recipe
receipts: markdown, git, prompt files

caveatnot a state layer; it only transforms what you already have

official

Listed in the registry. No canonical self.md page yet.

Starter stacks by job

1. Local/private assistant

Use this when the first requirement is “I want an assistant on my machine before I wire it to accounts.”

Surface: Open WebUI or a local chat frontend.
Model: Ollama, llama.cpp, vLLM, or a paid API if local quality is not enough.
Knowledge: local files plus a small retrieval layer; keep the source files outside the chat app.
Actions: start with read-only tools. Add shell/browser control only after logging is boringly clear.
Receipt: config file, model name, prompt file, source folder, command log.

This stack is not automatically a Personal OS. It becomes one only when memory, permissions, and action receipts are explicit.

2. Memory-first stack

Use this when your main pain is forgotten context, repeated briefings, and agents that keep relearning you.

Memory primitive: mem0 , MemU , or Letta for agent memory.
Source of truth: Markdown, SQLite, Git, or another store you can inspect.
Retrieval: LlamaIndex, Haystack, Quivr, or a small custom index.
Correction path: every memory should be editable, attributable, and removable.
Receipt: memory diff, source citation, expiry/correction note.

Do not collapse this into “RAG.” Retrieval answers “what file is relevant?” Memory answers “what state should this agent carry forward?” Different jobs. Different failure modes.

For the concepts, read agent memory systems and the memory system guide .

3. Browser-heavy stack

Use this when the work happens in web apps: dashboards, bookings, forms, admin panels, research flows.

Vendor path: OpenAI Operator or ChatGPT agent , Anthropic computer use , or Claude Cowork-style desktop work.
Replaceable path: browser-use, Skyvern, Browserbase, Playwright MCP .
Policy: manual approval for purchases, deletes, sends, account changes, and irreversible form submissions.
Receipt: screenshot before/after, URL, run log, browser trace, confirmation ID.

Browser agents are powerful because they operate where APIs are missing. They are dangerous for the same reason. Treat the browser as a messy action surface, not a trusted kernel. See browser agents and browser automation .

4. Coding-agent stack

Use this when the “personal OS” is really your software workbench.

Vendor path: Claude Code , OpenAI Codex, Cursor, Windsurf, Replit Agent, GitHub Copilot agent mode.
Open path: OpenHands , Superpowers, repo-level skills, project-local rules.
Memory: AGENTS.md, CLAUDE.md, rules files, project memory, design specs, issue context, and plugin/skills systems where they are explicit enough to review.
Policy: tests before merge, no secret exfiltration, no surprise restarts, no broad rewrites without a diff.
Receipt: Git diff, test output, build log, issue/PR link.

Coding agents are the clearest preview of Personal OS because they already have a loop: read context → plan → edit → run tools → leave a diff. That loop should be copied to non-code work, not the other way around.

5. Workflow daemon stack

Use this when you need work to happen without a chat window open.

Workflow engine: n8n, Flowise, AutoGPT, CrewAI, LangGraph.
Connectors: MCP servers , Composio/Rube, direct APIs.
Triggers: cron, webhooks, file changes, inbox events, calendar events.
Policy: event filters, budgets, retry limits, human checkpoints.
Receipt: run ID, input payload, output artifact, failure log, notification target.

This is where “agentic” often turns into unattended spaghetti. Start with small, visible loops. If you cannot replay a run, you cannot debug it.

6. User-owned network stack

Use this when the real requirement is not just local control, but portable identity and state across agents.

Runtime: Hermes Agent-style local runtime: tools, skills, gateway surfaces, cron, model/provider config, approvals.
State: files, Git, local databases, encrypted state, or a network.self.md-style peer state.
Protocol: MCP where it helps expose tools; encrypted messaging/state where agents need to collaborate.
Policy: approval-gated bridges to external channels and accounts.
Receipt: state snapshot, member/peer log, message ID, action approval, exported artifact.

This is the most “Personal OS” direction, but also the least plug-and-play. Treat it as infrastructure. Keep the claims boring: emerging, inspectable, replaceable, not magic.

Recipe queue for community stacks

Custom stacks are allowed into the registry, but only as reproducible recipes: component graph, custody model, risk flags, and receipts. Otherwise it is just fanfic with tool names.

12 recipe archetypes Use these as submission templates, not magic blueprints.

Local private assistant user-owned

assistant on my machine before accounts

components: Open WebUI, Ollama or llama.cpp, Markdown/Obsidian, Git, read-only MCP servers
layers: chat, retrieval, memory, receipts
receipts: config files, source folder, git commits, command logs
setup: Install local model runtime, point the UI at local docs, add read-only tools, log every tool call.

caveatLocal is not safe by default; permissions still matter.

Memory-first personal context hybrid/user-owned

stop re-briefing agents

components: mem0 or Letta, SQLite/Postgres, Markdown source files, LlamaIndex/Haystack retrieval
layers: memory, retrieval, receipts
receipts: memory diffs, source citations, expiry/correction notes
setup: Keep memory editable and separate from retrieval; add the correction workflow first.

caveatDo not call RAG memory.

Browser-heavy admin stack hybrid

forms, dashboards, bookings, web apps

components: Playwright MCP or browser-use, Browserbase/Hyperbrowser optional, approval gate, screenshot archive
layers: browser, tools, workflow, receipts
receipts: screenshots, DOM traces, run IDs, confirmation pages
setup: Start with test accounts; require human approval for sends, purchases, deletes, and account changes.

caveatBrowser automation is not a security boundary.

Coding-agent workspace hybrid

software workbench with real receipts

components: Claude Code/Codex/Cline/Aider, AGENTS.md, GitHub/Linear, tests
layers: coding, memory, tools, receipts
receipts: git diff, tests, commit, PR, issue link
setup: Keep project rules in the repo; no broad rewrites without a diff; tests before merge.

caveatVendor model quality can mask bad project custody.

Workflow daemon stack hybrid/user-owned

work happens without a chat window open

components: n8n or Dify/Flowise, MCP/API connectors, Discord/Telegram notifications, run ledger
layers: workflow, tools, receipts
receipts: run ID, input payload, output artifact, error log
setup: Name triggers, cap retries, log payloads, require approvals for risky writes.

caveatUnattended spaghetti if no replay.

Local voice + smart home stack user-owned

voice as home/device interface

components: Home Assistant Assist, Wyoming Satellite, local ASR/TTS/LLM, MQTT/device logs
layers: voice, device-context, action surfaces, receipts
receipts: Home Assistant logs, transcripts, automation traces, device state
setup: Keep local voice first; gate destructive device actions.

caveatVoice feels magical until it flips the wrong switch.

Lifelog capture stack hybrid/vendor-owned

meetings, calls, screen/audio memory

components: Granola/Limitless/Plaud/Rewind, export pipeline, Obsidian/Markdown archive
layers: receipts, memory, retrieval
receipts: audio files, transcripts, meeting notes, export dumps
setup: Treat capture as raw evidence; summarize into editable memory after review.

caveatPrivacy blast radius is huge.

Self-hosted business ops stack hybrid/user-owned

small team internal AI ops

components: Dify or LangGraph, Postgres, n8n, Pipedream/Composio, auth inventory
layers: workflow, retrieval, tools, identity/auth, receipts
receipts: DB rows, workflow logs, connector scopes, approvals
setup: Separate the data store, workflow engine, and auth ledger.

caveatEnterprise-ish complexity arrives fast.

Hermes/network runtime stack user-owned/experimental

user-owned agent runtime and P2P state

components: Hermes Agent, skills, cron, gateway, Git, network.self.md encrypted state
layers: runtime, tools, memory, network-state, receipts
receipts: tool logs, scheduled outputs, state exports, peer/member logs
setup: Keep claims boring: local runtime first, network state second.

caveatEmerging infra; not polished SaaS.

High-privacy offline-ish stack user-owned

sensitive personal docs

components: Jan/LM Studio/Ollama, local files, local vector store, no cloud connectors
layers: chat, retrieval, memory, receipts
receipts: local config, index files, source docs, backups
setup: Disable external tools; export configs; snapshot indexes.

caveatQuality and convenience cost is real.

Low-budget student stack hybrid

cheap learning / hobby Personal OS

components: free Claude/ChatGPT, Open WebUI/Ollama, Obsidian, GitHub, simple MCP tools
layers: chat, coding, retrieval, receipts
receipts: notes, git diffs, config files
setup: Use free vendor rails for reasoning, local files for state.

caveatDo not pretend free SaaS is user-owned.

Mobile/device-context stack hybrid

phone/desktop context without surrendering all state

components: Apple/Google/Microsoft device AI, Raycast/Pieces, local export/backup routine
layers: device-context, interface, memory, receipts
receipts: settings changes, screenshots, local index, exports
setup: Use vendor context as a surface; keep durable state exportable elsewhere.

caveatOS vendors will always pull state toward their account graph.

Submit your stack

If your Personal OS setup has a real component graph and receipts, submit it through the structured GitHub form. Required evidence: URL, layers, custody model, where it runs, receipt locations, caveats, and last verified date.

Submit a Personal OS stack →

Build/evaluation checklist

Before you add a tool to your Personal AI OS stack, answer this table.

Question	Good answer	Bad smell
Who owns memory?	I can inspect, correct, export, and delete it	“The assistant remembers somehow”
Who owns auth?	Tokens/scopes are listed and revocable	One agent has every account forever
Are actions logged?	Every side effect has an external receipt	The chat says it worked
Is there a sandbox?	Risky actions run in dry-run, test, or limited scopes first	Browser/terminal control is always-on
Is there an export path?	State survives a vendor switch	The product is the only source of truth
Are triggers visible?	Cron/webhooks/events are named and logged	Background runs nobody can explain
Is there human override?	Kill switch, approval gates, manual takeover	“Autonomous” means unstoppable
Is rollback real?	Git/backups/snapshots/checkpoints exist	The only rollback is apologizing
Is prompt injection handled?	Untrusted content is data, not policy	Webpages/emails can rewrite instructions
Can agents be swapped?	Tools and state are decoupled from one model	The whole stack dies with one subscription

The boring columns are the product. The model is just one replaceable part.

Market notes that change the choice

The big platforms are not sitting still. Google is pulling Gemini into Android. Apple is exposing app actions through Apple Intelligence and App Intents. Microsoft is turning Windows/Copilot into a memory-and-action surface. OpenAI is connecting ChatGPT to apps, tools, browsing, and computer-use-style workflows. Anthropic is expanding from Claude Code into computer use, Cowork, skills, and MCP. GitHub, Cursor, Windsurf, Replit, and OpenHands are turning code repos into agent workspaces.

That means your safest bet is not “pick the winner.” The safer bet is portable state plus replaceable action surfaces.

The OpenClaw story is a good warning label. Public reports said creator Peter Steinberger was joining OpenAI, and other reporting framed it as smart-agent talent capture. Useful signal: the access-control layer matters. Unsafe leap: turning that hire into a project-purchase claim, or assuming every open personal agent is doomed. Do not build strategy on rumors.

What not to build

Do not build another generic dashboard with a chat box in the middle.
Do not choose tools by stars, badges, or “active” labels. Those age badly and hide architecture.
Do not call retrieval memory unless it has correction, provenance, and expiry.
Do not stack MCP servers without an auth inventory and revocation plan.
Do not let a browser agent control accounts without screenshots, approvals, and logs.
Do not present a vendor chat app as user-owned custody.
Do not chase full autonomy before receipts. That is how you get confident bullshit with side effects.

Practical recommendation

Start with one missing layer:

If you need a local interface, start with Open WebUI and local files.
If you need durable context, start with mem0 or Letta plus an editable source store.
If you need tool access, start with MCP servers and a tiny allowlist.
If you need browser work, start with Playwright MCP or browser-use in a test account.
If you need coding leverage, start with AI coding assistants and Git receipts.
If you need recurring workflows, start with n8n/Flowise/LangGraph and run logs.
If you need a user-owned agent runtime, look at Hermes Agent and keep state exportable from day one.

Then add exactly one next layer. Personal OS work compounds; tool sprawl just rots.

Sources worth checking

Source	Use it for
Google Gemini Intelligence for Android	OS/device-context direction.
Apple Intelligence and App Intents	Personal intelligence and app action exposure.
Microsoft Recall and Copilot Actions	Windows memory/action surface.
OpenAI Operator , ChatGPT agent , Apps SDK , MCP connectors	Browser/action/connectors layer.
Claude Code , Claude memory , computer use , skills and MCP	Coding-agent and tool-use surface.
Open WebUI , Khoj , Letta , mem0	User-owned interface and memory layers.
Open Interpreter , 01 , browser-use , Skyvern , Browserbase	Local computer and browser action layers.
n8n AI docs , Flowise , CrewAI , AutoGPT	Workflow and agent-graph layer.
Hermes Agent docs	Local agent runtime shape: tools, skills, gateway surfaces, cron, approvals.

Export Config for Claude Code / Codex / Agents

One-screen chooser

Personal OS registry

Starter stacks by job

1. Local/private assistant

2. Memory-first stack

3. Browser-heavy stack

4. Coding-agent stack

5. Workflow daemon stack

6. User-owned network stack

Recipe queue for community stacks

Submit your stack

Build/evaluation checklist

Market notes that change the choice

What not to build

Practical recommendation

Sources worth checking

Get the next field note