The plumbing every local AI experiment rebuilds, built once. athome is one SDK, CLI, and skill set for the machinery every experiment repo grows anyway — content-keyed caching, sidecar workers, record/replay LLM caching, launchd scheduling, detached overnight runs, and resumable progress.

Get started

uv add experiment-at-home
uv run athome --help

Driving with an agent? Paste this:

Add experiment-at-home (uv add experiment-at-home) and wrap my most expensive
pipeline step in its @cached decorator, namespaced after this project.
Docs: https://yasyf.github.io/experiment-at-home/

/plugin marketplace add yasyf/experiment-at-home
/plugin install athome@experiment-at-home

Use cases

Stop re-paying for the same expensive artifact every rerun

A pipeline rerun that re-derives every intermediate burns minutes (or API dollars) to reproduce bytes you already had. Wrap the step; the cache is content-keyed under ~/.athome/cache, shared across every repo on the machine:

from athome import cached


@cached(ns="ingest", version=1)
async def parsed_pages(pdf_digest: str) -> list[str]:
    # the expensive part: OCR, chunking, whatever your pipeline re-derives
    ...

The second call with the same arguments returns from disk. Invalidation is one edit: bump version=2 and the old entries are never read again.

Send tonight's sweep off before you close the laptop

An overnight script tied to a terminal dies with the terminal, and nohup leaves you grepping for whether it finished. Detach it with a name:

uv run athome run --detach --name sweep -- uv run python sweep.py

The run gets its own log under ~/.athome/logs/runs/, and a completion sentinel records the exit code. In the morning, uv run athome run wait sweep returns instantly with that exit code — 0 or the failure you need to chase.

Move a corpus without wondering if it all arrived

rsync reports transfer success, not tree integrity — and macOS salts remote copies with ._* AppleDouble droppings. Mirror with verification:

uv run athome sync ~/corpora/traces backup-host:corpora/traces --move

Every file is sha256-verified against the source before --move deletes anything; a mismatch repairs and re-verifies, and an unverifiable tree blocks deletion outright.

More in the docs

• Record/replay LLM caching — replay yesterday's API responses in today's tests, byte-for-byte — llmcache
• Sidecar workers — call GIL-bound engines from free-threaded Python over a version-handshaked pipe — workers
• launchd agents — declarative Calendar/Interval/KeepAlive schedules, no plist authoring — launchd
• Resumable progress — JSONL work-sets where crashed units retry and finished ones skip — progress
• One config file — every knob in ~/.athome/config.toml, every module a pydantic-settings section — config

Read the docs for the full guide.

Status: alpha — the primitives above are shipped and tested (including on free-threaded Python 3.14); MLX serving recipes, OCR engines, batch LLM lanes, and the autoresearch harness land in 0.2–0.3.

Licensed under PolyForm-Noncommercial-1.0.0.