Skip to content

Latest commit

 

History

109 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

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.

Silent 24-second LightClaw walkthrough: request, plan, workers, tests, and run receipt

Live site · Five-minute demo · Install · Security boundary · Releases · Changelog

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 hashes, a private JSON/Markdown receipt, and a scoped undo path.

That loop—not a long channel list—is the product.

See it work without a token

The deterministic demo contacts no model and needs no Telegram account:

git clone https://github.com/OthmaneBlial/lightclaw.git
cd lightclaw
python3 -m venv .venv && . .venv/bin/activate
python -m pip install -e .
lightclaw demo

It replays a recorded phone request and approval, creates a disposable Git-backed Python service, runs a real unit test, and finishes with changes.patch, artifact.json, receipt.json, and receipt.md. Nothing is pushed. Try every product story:

lightclaw demo --scenario memory
lightclaw demo --scenario repo-task
lightclaw demo --scenario multi-agent
Story What it proves locally Replay notes
Repository task A real unit test, Git patch, artifact, and accepted receipt Exact prompt and cleanup
Persistent memory A synthetic fact survives a real SQLite restart and is recalled lexically Exact prompt and limits
Multi-agent plan Dependency order, machine-readable handoffs, visible failure, and one bounded repair Exact plan and limits

For public, sanitized evidence that can be forked and replayed, browse the privacy-checked showcase. Its three starting recipes are maintainer fixtures, not community submissions or claims about live provider quality.

Measured proof, not a blanket performance claim

The current full benchmark at commit ca1aea7 used macOS 26.6, Python 3.13.1, five runs, and live PyPI resolution with a potentially warm cache.

Measurement Result
Base direct dependencies 3
Clean-wheel installed distributions 10
Clean-wheel install 4.245 s
CLI import median 223.499 ms
Lexical memory fixture 8 / 8 top-1
Orchestration fixture 4 / 4 contracts

These numbers describe that commit and machine. They do not predict live Telegram latency, provider cost, model quality, or availability. Reproduce them from bench/.

LightClaw is / is not

LightClaw is LightClaw is not
A Telegram-first review and control surface for local coding agents A hosted multi-tenant agent service
Local SQLite memory, task workspaces, receipts, and skills Fully local inference when a hosted model is configured
Human approval around plans and trusted execution A guarantee that model-generated changes are correct
A small Python product with optional edges A feature-for-feature OpenClaw clone
Auditable fixture stories that run for $0 Proof of real-provider latency, cost, or availability

Safe installation

Use an isolated tool environment and install only the SDK for the provider you actually use. OpenAI, xAI, DeepSeek, and Z-AI share the openai transport extra:

pipx install 'lightclaw-ai[openai] @ git+https://github.com/OthmaneBlial/lightclaw.git'
# or
uv tool install 'lightclaw-ai[openai] @ git+https://github.com/OthmaneBlial/lightclaw.git'

Replace openai with claude or gemini, or use providers to install all three SDK families. The token-free demo needs no provider extra.

Then:

lightclaw demo
lightclaw onboard --configure
lightclaw doctor
lightclaw run

The supported paths are app-specific:

  • config: ~/.config/lightclaw/config.env (mode 0600);
  • memory, skills, logs, and receipts: ~/.lightclaw/;
  • owned task workspaces: ~/.lightclaw/workspace/.

Existing configuration is backed up before reset. The compatibility installer uses its own virtual environment and never installs into global Python. See install, upgrade, undo, and uninstall.

Signature workflow

Telegram text or voice goal
  -> scoped DAG with risk, paths, and expected outputs
  -> approve / edit / deny
  -> isolated Codex or Claude workers
  -> compact live progress
  -> tests + Git patch + file evidence + receipt
  -> accept / reject / selective apply / optional PR preview

Current capabilities include:

  • fail-closed Telegram allowlisting and privileged-command rate limits;
  • OpenAI, xAI, Anthropic, Gemini, DeepSeek, and Z-AI routing;
  • Codex and Claude delegation profiles: observe, workspace-write, trusted-command;
  • DAG planning, owned paths, JSON handoffs, acceptance checks, and bounded repair;
  • namespaced SQLite FTS5 lexical recall with retention, export, and selective delete;
  • permission-manifest skills with pinned provenance, hash review, and prompt-only activation;
  • voice transcription, scheduled jobs, heartbeat, Telegram, and terminal chat;
  • token-free fixture adapters used by the full CI matrix.

Security model

An empty TELEGRAM_ALLOWED_USERS blocks startup. Intentionally public bots require LIGHTCLAW_PUBLIC_BOT_ACK=yes. Delegated processes get a minimal environment that excludes Telegram/provider keys. lightclaw undo refuses paths that lack a LightClaw ownership record.

These controls do not protect the host after you explicitly enable trusted-command, install malicious instructions, weaken an external CLI sandbox, or place secrets inside a readable task workspace.

Read SECURITY.md and the threat model. Report vulnerabilities through private vulnerability reporting, never a public issue.

Documentation

Honest alpha limits

  • Memory retrieval is bounded SQLite FTS5 lexical search, not semantic understanding. Embeddings are an optional adapter and are never required.
  • Fixture demos prove LightClaw contracts, not external model quality.
  • External coding-agent CLIs remain separate security boundaries with their own versions and settings.
  • The package is installable from Git today; the first stable PyPI release follows the release gate in the roadmap.

Roadmap and release state

  • Now: make the base install smaller, keep the three fixture stories reproducible, and collect privacy-bounded external install evidence.
  • Next: publish v0.1.0, PyPI attestations, and a versioned container only after 10–20 external self-hosters exercise the clean install and the release gate passes.
  • Later: improve acknowledgement/reconnect proof and real-world operator ergonomics before adding new channels or chasing broad assistant parity.

No stable release exists yet. The release page, draft notes, and evidence-derived gate make that status explicit.

Help qualify v0.1.0

Follow the five-minute quickstart, then submit the bounded alpha report. Never include prompts, receipts, repository content, local paths, Telegram identities, tokens, request IDs, or screenshots. Failures and missing timings remain visible evidence.

Community and updates

Release notes and the update thread publish fixes, workflows, benchmarks, limitations, and explicit pauses. See the maintenance and badge policy.

Development

python3 -m venv .venv
. .venv/bin/activate
python -m pip install -e '.[dev]'
python scripts/quality.py

This is the canonical local quality command. The GitHub workflow runs the key-free suite on Python 3.10–3.13 across Ubuntu and macOS, installs the wheel in a clean environment, audits dependencies, and replays all three deterministic stories. Read CONTRIBUTING.md, the support routes, and the Code of Conduct before participating.

Licensed under MIT.

About

Control local Codex and Claude coding agents from your phone via Telegram — review plans, approve bounded work, inspect diffs/tests, and keep receipts.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

20 stars

Watchers

1 watching

Forks

Used by

Contributors

Languages