GraphyOS

Compile a codebase into an AI-native operating system. Point GraphyOS at a repo; it smashes the code, its docs, and its dependencies into walkable JSON substrates, then derives a fan-out — a routing layer a cold AI model boots through to operate the code like a maintainer who's been there for years.

The model is replaceable. The substrate is the moat. Not "an AI that knows your codebase." An AI that knows how to walk your codebase cold — because the repo was compiled into a map it can traverse, instead of stuffed into a prompt it has to guess from.

GraphyOS is a distribution of an AI coding agent (reference-implemented on Claude Code): an opinionated operating layer — a constitution, a substrate toolchain, and enforcement gates — the way a Linux distro is an opinionated assembly on a kernel. Minimal kernel + optional modules + a way to grow your own.

The pattern

smash   →   walk   →   build the TOC from the walk

Compile the repo (and its dependencies) into JSON graphs on their own literals. Walk them. Derive the routing layer from the clusters the walk finds — never hand-author it, so it can't name a symbol the code doesn't have. Then a cold agent boots the fan-out and answers maintainer-style questions without prompt-stuffing.

Layout

path	what
CLAUDE.md	the kernel — the constitution: reflexes · the autonomy loop · the gate model. Start here.
AGENTS.md	the cold-boot drill (the six moves an agent runs per task).
FLEET.md	optional module — running GraphyOS as a multi-agent, multi-machine fleet.
tenants/fastapi/	the flagship proof: FastAPI, ripped cold — repo + Starlette + Pydantic smashed into walkable joined substrates with a walk-derived fan-out. See its DEMO.md.

Try it (the flagship tenant)

cd tenants/fastapi
./rebuild.sh          # clone FastAPI → smash repo + starlette + pydantic → walkable substrates
                      # then boot a cold agent on tenants/fastapi/CLAUDE.md and ask a maintainer question

The FastAPI tenant was built cold in one session: fastapi 499 nodes + starlette 618 + pydantic 2,011, with the cross-substrate joins (APIRouter→starlette.Router, FastAPI→Starlette, →pydantic) proven from the walk. A cold model routed "how would I add validation to APIRouter?" to the right seam — because the join told it, not because it happened to remember. See tenants/fastapi/DEMO.md.

Why it works

The power isn't private data — it's owned literals, schema admission, substrate ingestion, a walkable AST hierarchy, dependency joins, agent-boot discipline, and gates that enforce all of it at the tool boundary. Any repo the toolchain can walk becomes agent-native the same way.

Status

Early and opinionated. The kernel + the FastAPI tenant are the proof; the standalone graphy CLI (so a stranger can clone → smash with one command) is the next milestone — today the substrate toolchain rides its source checkout. Issues and tenants welcome.