local-first / ai-ready

Your work leaves traces.

YZOS turns them into knowledge your AI can use, entirely on your machine.

$ git clone https://github.com/StringKe/yzos && cd yzos && mise run dev-install
mcp { "mcpServers": { "yzos": { "url": "http://127.0.0.1:17432/mcp" } } }
pipeline

How it works

Observe

Screen recording with OCR, clipboard history, microphone transcription and AI tool sessions are captured passively in the background. You never operate YZOS directly; it watches while you work.

Crystallize

A local LLM condenses raw traces into people, projects, tasks and first-person memory points. Knowledge lives in plain Markdown files you can read and edit.

Supply

Claude Code, Codex and any MCP client connect to a local HTTP endpoint and query your activity, entities, knowledge and clipboard, with real data instead of guesses.

supply
MCP tools served by YZOS: what each returns and which observed data source backs it
toolreturnssource
activity
yzos_activity_summaryday digestscreen ocr
yzos_activity_sessionssession listwindow focus
yzos_ai_sessionsai tool sessionsagent logs
entity
yzos_entity_listentities by kindpostgres
yzos_entity_getentity detailpostgres
yzos_entity_searchranked matchespgvector
knowledge
yzos_knowledge_searchokf sectionsokf markdown
yzos_knowledge_readmarkdown sourceokf markdown
clipboard
yzos_clipboard_recentredacted historypii model
yzos_clipboard_searchredacted matchespii model

# Your AI assistant asks YZOS, not you

$ claude "what was I working on this morning?"

yzos_activity_summary -> 3 sessions: cluster autoscaling fix,

model registry refactor, PR review for auth module

YZOS.app  (menu bar, single process)
|- observe      screen / clip / mic / ai
|- crystallize  llama.cpp -> entities + okf
|- postgres     ~/.yzos/pg  (pgvector)
\- http         127.0.0.1:17432  /mcp
promise

Why YZOS

Everything stays local

Recording, OCR, inference and storage all run on your Mac. Embedded PostgreSQL, bundled llama.cpp, no cloud account, no telemetry.

Five model slots

Reasoning, vision, speech, embedding and PII detection each run their own model. Swap any of them from the built-in Hugging Face marketplace.

Privacy by default

A dedicated PII model redacts passwords, names, phone numbers and ID numbers before anything is stored or served.

Free and open source

MIT licensed. No paid tier, no feature gates. Build it, run it, fork it.

install

# 1

$ git clone https://github.com/StringKe/yzos && cd yzos

# 2

$ mise trust && mise install && mise run dev-install

# 3

$ claude mcp add --transport http yzos http://127.0.0.1:17432/mcp

ready state=observing pg=up llm=ready mcp=:17432