Rack Community: a self-hosted AI lab whose job runner won't overload the box

Rack Community released a self-hosted AI lab for a single Linux box, built on Node 24 and Python 3, that runs a durable job queue with memory admission control so jobs cannot overload the machine. The admission script admit.sh admits a job only if max(live memory in use, declared total) plus the job's peak is at or below LIMIT, with ADMIT_LIMIT defaulting to 90 GB and ADMIT_BASE to 10 GB; a refused job exits 75. The project ships no model and no keys, pointing recipes at Ollama at http://127.0.0.1:11434/v1 or any OpenAI-compatible endpoint via RACK_LLM_URL, RACK_LLM_MODEL and RACK_LLM_KEY, and includes Somnus, a human-approved memory control plane, and TalkBack, a team room AI agents can join.

A self-hosted home AI lab for one box: live gauges, a job runner that refuses to overload the machine, recipe templates for training and testing, memory you approve, and a team room any AI can join. If this is useful to you, a star helps other people find it. | Controls stay locked until you paste your token. | Pause says who paused and why. Saved pieces are kept. | | Recipes, and what memory each one is allowed to use. | Run status follows your own log lines as they are written. | | Works at phone width. | | Linux, Node 24, Python 3. Everything listens on 127.0.0.1 . clone or download this repository, then from its top folder: ./start.sh open http://127.0.0.1:8080 and paste the token that start.sh prints the path of Then Controls , switch to LIVE , pick slow count , Check it fits , Start , type START . Press Pause everything half way, then Resume : the saved pieces are kept and the job finishes. | Folder | What it is | |---|---| | engine/ | The core, in Python. A durable job queue; an actor that runs one admitted job at a time with a hard deadline per piece; admit.sh memory admission control; recipes; host gauges CPU, memory, and GPU if nvidia-smi exists; a reading that cannot be taken shows unavailable , never a made-up zero ; a small HTTP API. | | panel/ | The control panel React, TanStack Start : gauges, "what is it doing", jobs with saved progress, recipes, controls, run status. | | memory/ | Somnus. A control plane for AI-agent memory. It runs a dream pass over a memory folder and reports what it proposed, what it applied and what waits for a person. Judgment-class changes wait for a human; you can diff a snapshot against live and roll back. It calls no model provider. | | room/ | TalkBack. A team room for people and AI agents: channels, threads, reactions, mentions, a wire so an AI agent joins as a desk with its own key, huddles, and an Electron shell that hosts the room on your LAN. | Each folder is its own project with its own package.json or requirements.txt . Rack Community ships no model and no keys. It is the place your models run and report. - Local: point a recipe at Ollama https://ollama.com . engine/samples/ask model pieces.py asks a list of questions, one saved piece per answer, against http://127.0.0.1:11434/v1 by default. - Hosted: the same script works with any OpenAI-compatible endpoint. Set RACK LLM URL , RACK LLM MODEL and, if the service needs one, RACK LLM KEY sent as a Bearer token and never written to the saved pieces . - In the room: any AI agent that can make an HTTP call can join TalkBack as a desk with its own key room/tools/talkback-wire.mjs is a small client . To use the sample, copy engine/samples/ask model.recipe.json into your Rack folder's recipes/ and add ask model 8 to its peaks.conf . Admission control. admit.sh admits a job only if max live memory in use, declared total + the job's peak <= LIMIT . The declared total is a base figure plus the declared peaks of admitted jobs that are still alive, so two jobs that have not yet touched memory cannot both be admitted into the same gap. A refused job exits 75 and says why. Peaks live in peaks.conf ; ADMIT LIMIT default 90 and ADMIT BASE default 10 are in GB. Durable pieces. A job is split into named pieces. Each piece result is written atomically and in order, so a crash, a restart or a pause resumes at the first piece that was not saved. A supervisor outside the worker enforces a wall-clock limit per piece, including a worker stuck in a native call. Dry-run first. The Rack starts in dry-run: every request is validated and recorded and nothing changes. Switching to LIVE is one explicit action. Pause writes a single HOLD marker; work queued before the pause waits, new starts are refused while it exists. Recipes, not commands. A start names a recipe from recipes/ .json . No endpoint accepts a command line, a path or a shell string. {"action": "nightly index", "label": "Rebuild the search index", "script": "samples/hello pieces.py", "args": , "admission": "nightly index", "peak gb": 4} Add the recipe file, add nightly index 4 to peaks.conf , and write the script with from rack jobs import from environment . See engine/samples/ for complete examples. More detail in docs/ARCHITECTURE.md https://github.com/ohhbeehave35-commits/rack-community/blob/main/docs/ARCHITECTURE.md . - The API needs the X-Rack-Token header for every write and for the two reads that show your log lines. With no RACK TOKEN set, nothing unlocks. - Bind everything to 127.0.0.1 , or put it behind your own access boundary. Do not expose it to the internet as it is. - Somnus and TalkBack have no sign-in. They run as one local owner. Their server functions refuse requests that a browser reports as started by another site, but they are not a multi-user service. - The panel keeps the token in sessionStorage for one tab and never puts it in a URL, a body or a log. Workers never receive it. - Admission, pause and the recipe list are safety rails, not a sandbox: a recipe script runs as the user who started the actor. Use a different port from the panel: cd memory && npm install && npm run dev -- --port 8081 http://127.0.0.1:8081 cd room && npm install && npm run dev -- --port 8082 http://127.0.0.1:8082 Somnus keeps its state in an embedded database, in memory by default; set PGLITE DATA DIR to keep it. TalkBack's desktop shell is described in room/desktop/README-DESKTOP.md . Run the narrowest check that covers what you changed. cd engine && python3 -m pytest -q tests 38 tests cd panel && npx tsc --noEmit && npm test 19 tests cd memory && npx tsc --noEmit && npm test 39 tests cd room && npx tsc --noEmit && npm test 61 tests Those are unit and small integration tests. There is no browser end-to-end suite in this repository, and the full test suites of the original apps were not carried over see "What is not here" . - No company-specific assistant, recipes, prompts, data or customer material. This is the general-purpose part only. - The engine was rewritten from the mechanisms of a private lab the admission rule, ordered atomic pieces, external deadlines, a pause that holds starts, dry-run first . It is not a cleaned copy of that code. - The web apps in panel/ , memory/ and room/ started from an app template whose licence we could not establish. Its sign-in layer, install-page scripts and the tests that only exercised them were removed before this release, and the apps now run as one local owner. Some template-era source files remain in the apps; see NOTICE . - Somnus's Backup page and checkpoint tool were removed because they copied files specific to one machine. The TalkBack relay has no authentication of its own see room/relay/README.md . - Version 1.0.0 is a first public release. Expect rough edges. Issues and pull requests are welcome. Keep changes small, keep the safety rails admission, dry-run, token-checked writes intact, and say which tests you ran. MIT. See LICENSE and NOTICE . Copyright c Rack Community contributors.