Lightclaw

LightClaw

**Control local Codex and Claude coding agents from your phone via Telegram—and keep the proof of what they delivered.**

LightClaw is a self-hosted mission-control layer, not a generic remote terminal. Send a goal, review its paths, risk, workers, and acceptance checks before execution, then receive the diff, test evidence, artifacts, private run receipt, and scoped recovery instructions.


[Live site](https://othmaneblial.github.io/lightclaw/) · [Five-minute demo](#see-it-work-without-a-token) · [Install](docs/INSTALL.md) · [Security boundary](docs/THREAT_MODEL.md) · [Releases](https://github.com/OthmaneBlial/lightclaw/releases) · [Changelog](CHANGELOG.md)

Alpha software with meaningful host access. Telegram access fails closed, delegated workers do not inherit provider secrets, and trusted host execution always requires per-run confirmation.

Why LightClaw

Phone access, Telegram, diffs, and multi-agent execution already exist elsewhere. LightClaw's narrower job is the **governed handoff**: approve a delivery contract, let bounded local workers execute it, and judge the result from durable evidence instead of a reassuring chat summary.

If you need…	Best fit
A live mobile terminal or full session steering	A dedicated remote-control client
A broad multi-channel personal assistant	A general agent gateway
Pre-run scope review, bounded Codex/Claude workers, acceptance checks, a patch, receipt, and undo path	LightClaw

One complete story

1. From Telegram: “Add a health check, test it, and return only verified work.”
2. LightClaw shows builder → verifier, requested paths, risk, and capability before execution.
3. You approve, edit, or cancel the plan.
4. Workers run in a dedicated task directory with a minimal environment.
5. LightClaw returns actual test evidence, a reviewable Git patch/branch, file h