The operating system for autonomous software engineering. LOOPRITE gives AI coding agents durable, file-based memory and a deterministic execution protocol — vendor-neutral, resumable, and gated by explicit human confirmation. Now running in your pocket.
AI coding agents forget everything between sessions, and every vendor locks that memory into its own proprietary state. LOOPRITE fixes both: it puts the agent's memory, plans, events, and safety rules into plain files inside your repository — readable and writable by any agent. Claude can plan, Codex can execute, Gemini can verify and respond to review comments, and whatever ships next can resume all of it from files. No backend. No hosted service. No lock-in.
The .l00prite/ folder holds the blueprint, run ledger, durable decisions,
constraints, known failures, and todos — the project's whole working memory as
Markdown and JSON, versioned with your code.
Six canonical prompts — resume, heartbeat, event, review, handoff, execute — define exactly how any agent picks up the next unit of work, verifies it, and persists what happened before stopping.
PR review comments and failed CI runs become first-class JSON events with a real
lifecycle — pending → processing → completed — handled one at a time,
always treated as untrusted input.
heartbeat.json carries iteration budgets, continue/pause/stop signals,
and review gates — so a loop knows when to keep going and when a human has to look
first.
A lock/lease convention (lock.json) keeps agents working close together
in time from silently corrupting shared memory — check, acquire, respect, reclaim
stale, release.
ledger.md records what actually happened, run by run, with the
verification evidence — the commands that ran and what they proved. Progress you can
audit, not vibes.
Autonomy without a gate is a liability. LOOPRITE splits agent work into two modes — and the autonomous one can only be entered through an explicit, in-session human confirmation. Every run. No carry-over grants, and no persisted flag can ever stand in for you.
Clarifies requirements, generates a blueprint, scaffolds the repo and its
.l00prite/ memory — then stops. Planning never executes the
project it scaffolds, and always ships Execution Mode disarmed.
Reads the blueprint and iterates autonomously — plan a unit, execute it, verify it with recorded evidence, persist memory, repeat — until Definition of Done or a run boundary stops it.
iteration_limit_reached to destructive_operation_required, every exit is resumable
The beta APK ships the unmodified LOOPRITE gateway as a native arm64 binary and
runs it entirely on-device — setup wizard, dashboard, playground, and autonomous runs
in a WebView on 127.0.0.1. Your provider keys are sealed by an
Android-Keystore-wrapped master key and never leave the device. There is no
LOOPRITE cloud: your phone is the server.
cli-os/scripts/build-apk.sh.EXECUTE. Pushes,
merges, and anything destructive still wait for your explicit approval.LOOPRITE is under active development and this Android build is a beta: it is debug-signed, distributed outside the Play Store, and rough edges are expected. Point it at repos you can afford to experiment on, keep Execution Mode's approval prompts on, and expect breaking changes between releases. Found something weird? Open an issue — beta reports shape the release.
LOOPRITE is built off the l00prite source code — the vendor-neutral protocol, the gateway, the run engine, and this very APK pipeline are all MIT-licensed and public. Read the prompts your agents will follow. Audit the run boundaries. Build the APK from source and diff the checksum. That's the point.
Protocol spec, canonical loop prompts, vendor adapters, the Go gateway, the run engine, validators, and the hermetic Android build — one repo, MIT license.