chrysb / alphaclaw  ·  @chrysb/alphaclaw

Deploy once.
Stay running forever.

AlphaClaw is the operations plane of the ClawFamily. It wraps OpenClaw in a self-healing harness — watchdog, crash-loop detection, Git rollback, prompt hardening, and one-click Railway/Render deployment. No ops expertise required.

Stars 639
Forks 79
Age 16 days
Language JavaScript
Runtime Node ≥ 22.12
License MIT
Deploy Railway · Render · Docker
Last pushed 2026-03-13 — actively maintained at < 3 weeks old. Fastest-growing ops wrapper in the ClawFamily ecosystem.

AlphaClaw wraps OpenClaw — ops plane around agent plane

AlphaClaw runs OpenClaw as a managed child process on 127.0.0.1:18789, proxying all traffic through its own Express server. OpenClaw never faces the internet directly. The watchdog monitors health, detects crash-loops, and auto-repairs.

Public-facing — Express Server
AlphaClaw
Handles all inbound requests. Runs the setup UI (Preact), manages webhooks, injects prompt hardening (AGENTS.md + TOOLS.md), runs hourly Git commits, manages Google Workspace OAuth, and exposes the 7-tab admin dashboard.
Express HTTP Setup UI (Preact) Watchdog Git Sync (cron) Prompt Hardening Google OAuth Webhook engine Telegram + Discord pairs
proxy → child process :18789
Agent Plane — Child Process
OpenClaw Gateway
Spawned by AlphaClaw at startup. Listens on loopback only — never exposed directly. AlphaClaw proxies requests in, intercepts responses for webhook transforms and prompt injection. Working directory is ALPHACLAW_ROOT_DIR (/data).
MCP agent execution Tool invocations 127.0.0.1 only
watchdog events + webhook history
Persistence
SQLite
Two tables: watchdog event log (crash history, repair actions, timestamps) and webhook request log (endpoint, payload, transform output, response). Zero external DB dependency.
Watchdog events Webhook history No external DB
graph TD Client["External Client\n(Telegram / Discord / HTTP)"] Alpha["AlphaClaw\n(Express :3000)"] WD["Watchdog\n(crash detection)"] GS["Git Sync\n(hourly cron)"] PH["Prompt Hardening\n(AGENTS.md + TOOLS.md)"] WH["Webhook Engine\n(transforms + logging)"] OC["OpenClaw Gateway\n(:18789 loopback)"] DB["SQLite\n(events + webhooks)"] GR["Git Repo\n(workspace backup)"] Client -->|"HTTPS"| Alpha Alpha --> PH PH -->|"inject per-message"| OC Alpha --> WH WH -->|"transform + log"| DB WD -->|"health check"| OC WD -->|"log events"| DB OC -->|"child process"| Alpha GS -->|"hourly commit"| GR style Alpha fill:#10b98120,stroke:#10b981,color:#dff2eb style OC fill:#22d3ee15,stroke:#22d3ee,color:#dff2eb style WD fill:#f87171,stroke:#f87171,color:#060d0f,fill-opacity:0.15 style DB fill:#f59e0b15,stroke:#f59e0b,color:#dff2eb style GR fill:#a78bfa15,stroke:#a78bfa,color:#dff2eb style GS fill:#10b98115,stroke:#10b981,color:#dff2eb style PH fill:#10b98115,stroke:#10b981,color:#dff2eb style WH fill:#10b98115,stroke:#10b981,color:#dff2eb style Client fill:#0d1a16,stroke:#5a8a75,color:#dff2eb

8 feature areas — pure operations, zero AI

AlphaClaw adds no intelligence to OpenClaw. Its job is to keep OpenClaw alive, reachable, and disciplined. Every feature is operational: health, rollback, auth, deploy.

🖥️
Setup UI — 7-Tab Dashboard
Preact + htm + Wouter (no build step). Tabs: General, Browse, Usage, Watchdog, Providers, Envars, Webhooks. Runs in the browser — no React build tooling required.
Preact · No Build
🩺
Watchdog + Crash-Loop Detection
Health checks on a timer. Detects crash-loops: 3 crashes within 300 seconds triggers auto-repair mode. Sends Telegram/Discord alerts. Logs all events to SQLite.
3 crashes / 300s threshold
↩️
Git-Backed Rollback
Hourly automated workspace commits via cron. If auto-repair fails, AlphaClaw rolls back to a known-good Git snapshot. Zero manual intervention required.
Hourly commits · Auto rollback
🧠
Prompt Hardening
Injects AGENTS.md and TOOLS.md on every message to prevent agent drift. The agent always has its behavioral constraints in-context, regardless of conversation length.
Per-message injection
🔐
Auth + Google Workspace OAuth
Password auth with exponential backoff against brute force. Full Google Workspace OAuth for 8 services. OpenAI Codex PKCE flow. All tokens managed by AlphaClaw, never exposed to OpenClaw directly.
Exp. backoff · PKCE · 8 Google services
🔗
Webhook Engine
Named webhook endpoints with transform modules. Every request is logged to SQLite with full payload, transform output, and response. Inspect from the dashboard.
Named endpoints · Transform modules
📡
Channel Orchestration
Telegram + Discord pairing via wizard. Topic groups support — multiple topics per Telegram supergroup, each routed to a different agent context. Extends OpenClaw's own channel support.
Topic groups wizard
🚀
Zero-Config Cloud Deploy
One-click Railway button, Render config, and Docker image. Deploy in under 5 minutes with no ops knowledge. AlphaClaw is the only ClawFamily member optimized end-to-end for instant cloud deployment.
Railway · Render · Docker

Radar — Operations vs Intelligence trade-off

AlphaClaw scores maximum on deployment ease, observability, and reliability — and intentionally scores 1 on agent intelligence and memory. It adds no AI. It keeps AI alive.

AlphaClaw
Operations plane
OpenClaw
Agent plane
ScalyClaw
Complete product
NanoClaw
Minimal footprint
MetaClaw
Learning plane
Dimension Score
Deployment Ease10
Observability9
Reliability9
Security6
Multi-channel5
Scalability3
Agent Intelligence1
Memory1

Four planes — each member owns a distinct concern

AlphaClaw sits between the user and OpenClaw: it is the ops glue that makes OpenClaw production-ready. Paperclip controls goals and budgets above AlphaClaw. MetaClaw learns below. ScalyClaw is a parallel fully-integrated product.

graph TD PC["Paperclip\n(Control Plane)\nGoals · Budgets · Org"] AC["AlphaClaw\n(Operations Plane)\nDeploy · Monitor · Heal"] OC["OpenClaw\n(Agent Plane)\nTask Execution · MCP"] MC["MetaClaw\n(Learning Plane)\nLoRA Fine-tuning"] SC["ScalyClaw\n(Complete Product)\nAll-in-one · Parallel track"] PC -->|"orchestrate"| AC AC -->|"wrap + proxy"| OC OC -->|"telemetry → fine-tune"| MC SC -.->|"parallel track\n(independent product)"| OC style PC fill:#d4a73a20,stroke:#d4a73a,color:#dff2eb style AC fill:#10b98125,stroke:#10b981,stroke-width:2px,color:#dff2eb style OC fill:#22d3ee15,stroke:#22d3ee,color:#dff2eb style MC fill:#f59e0b15,stroke:#f59e0b,color:#dff2eb style SC fill:#a78bfa15,stroke:#a78bfa,color:#dff2eb

AlphaClaw's position

AlphaClaw is the only ClawFamily member whose entire surface area is operational. It does not route messages intelligently, it does not learn, it does not own goals. It keeps things running. This makes it uniquely complementary — it can wrap any agent plane, not just OpenClaw.

Dependency on OpenClaw

AlphaClaw's tightest coupling is to OpenClaw's startup sequence and port contract. If OpenClaw changes its process interface, AlphaClaw breaks. This is by far its largest architectural risk — high probability, high impact. No abstraction layer exists between the two.

Member Plane Deploy Reliability Intelligence Memory Multi-channel
AlphaClaw Operations 10 9 1 1 5
OpenClaw Agent 4 5 8 7 8
ScalyClaw Complete product 5 7 7 9 10
NanoClaw Minimal 7 4 6 5 3
MetaClaw Learning 3 4 10 3 1

3 patterns worth borrowing

AlphaClaw solves three concrete problems that Lyra also faces. None require JavaScript — all three can be implemented in Python with existing Lyra primitives.

1
Prompt Hardening — per-message discipline injection
AlphaClaw injects AGENTS.md and TOOLS.md into every message sent to OpenClaw. This prevents behavioral drift as conversations grow long. Lyra faces the same problem: long sessions cause Claude to forget its operating constraints (tool limits, tone, persona boundaries). A system-message injection layer — triggered per-message, not per-session — would fix this.
Lyra impl: middleware in message_pipeline.py — append behavior_constraints.md to every LLM call
2
Crash-Loop Detection — upgrade supervisord awareness
AlphaClaw detects crash-loops at the application level: 3 crashes in 300 seconds triggers auto-repair (rollback, env reset) before alerting. Lyra's supervisord only restarts blindly. A thin Python layer wrapping supervisord's event bus could detect loops and either pause restart attempts or trigger a diagnostic routine before resuming.
Lyra impl: supervisord event listener script — count restart events in sliding window, alert + pause on loop
3
Telegram Topic Groups — multi-topic routing
AlphaClaw's wizard enables a single Telegram supergroup to have multiple topics, each routed to a different agent context. Lyra's Telegram adapter currently treats all messages from a chat identically. Supporting topic-based routing would allow #dev-questions, #personal, and #reminders to each go to a specialized context without separate bots.
Lyra impl: extract message_thread_id from Telegram Update, map to context_profile in session store

What Lyra does not borrow

AlphaClaw's deployment story (Railway, Docker) is irrelevant — Lyra runs on local hardware with supervisord. Its JavaScript stack, Google Workspace OAuth, and single-process Express model are all out-of-scope. The Git rollback pattern is interesting but Lyra's workspaces are not stateful in the same way. 3 borrows maximum — don't port the whole ops model.

6 risks — two are high, watch them

OpenClaw coupling
High
AlphaClaw is architecturally fused to OpenClaw's process interface. Any breaking change in OpenClaw's startup, port, or signal contract propagates immediately to AlphaClaw. No abstraction layer. Hard fork risk is real.
Very early project
High
Created 2026-02-25 — less than 3 weeks old at time of analysis. API surface, file structure, and feature set are all subject to rapid, breaking change. Production dependency today is a significant bet.
JavaScript, no TypeScript
Medium
The codebase uses plain JavaScript. No type safety, no compile-time checks. As the project grows, this will make refactoring riskier. Community contributions will be harder to validate. A TypeScript migration is unlikely given the "no build step" Preact philosophy.
Security trade-offs
Medium
AlphaClaw intentionally trades security for ease. Password auth (no MFA), basic exponential backoff only, OAuth tokens stored locally. Score 6/10 on security. Not suitable for multi-tenant or enterprise environments without hardening.
Docker/Linux only
Medium
Child process management and file path assumptions are Linux-first. Docker deployment abstracts this for cloud, but local development on Windows/macOS may have gaps. The Railway/Render focus means non-Docker deployment paths are undertested.
Single-process model
Low
One Express process handles everything: proxy, watchdog, webhooks, dashboard, Git sync. Scalability score: 3/10. For personal or small-team use (AlphaClaw's target), this is fine. Becomes a bottleneck under significant webhook load or many concurrent users.

Bottom line

AlphaClaw delivers exceptional value for its age: 639 stars in 16 days signals genuine market resonance for "zero-ops OpenClaw deployment". The prompt hardening and crash-loop detection patterns are production-grade ideas worth extracting. Do not run it in production against a critical workload yet — wait 6 weeks. The 3 borrow patterns for Lyra are low-risk to implement independently.