Quick Start

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/alib8b8/aflare/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/alib8b8/aflare/main/install.ps1 | iex

# Manual binary download
#   GitHub:  https://github.com/alib8b8/aflare/releases
#   CN accelerated: https://ghproxy.com/https://github.com/alib8b8/aflare/releases

- uses: alib8b8/aflare/action@v0.12.0
  with:
    workflow: .aflare/pr-review.yaml

Try it in 60 seconds:

aflare doctor                      # environment self-check (zero-config)
aflare run examples/content-processor.yaml   # read post.md → HTML → write post.html
aflare init                        # configure an LLM (local Ollama or cloud provider)

aflare create "monitor BTC price, alert via Telegram when > 70000"
aflare run btc-monitor.yaml        # generate a workflow from keywords (add --ai for LLM generation)

aflare chat                        # interactive ReAct Agent chat

Optional: install bubblewrap for full sandbox isolation (code_interpreter node) — sudo apt install bubblewrap / brew install bubblewrap.

Market data in generated monitoring workflows comes from public quote APIs — for personal research only, not investment advice.

What is aflare?

A local-first automation Agent and a deterministic workflow engine in a single binary. You explicitly grant access to your data (directories, note libraries, local databases), and the AI works deterministically inside the permission ceiling you define.

aflare chat / agent          aflare create
  ReAct Agent                  → YAML workflow
  (conversational)               ↓
       ↓                    DAG scheduled execution
  node tools                (WAL recovery · Saga · retry · audit)

Currently at v0.12.0, targeting local users first — local data lives on your machine, aflare is the deterministic and secure control layer between AI and that data.

Key Features

• Local-first, data stays local — single binary, zero runtime deps, ~10–30MB RAM; workflows, history, memory and secrets all stay on local disk; fully offline-capable; no usage telemetry.
• Connect your own LLM — Ollama / vLLM / LM Studio / any OpenAI-compatible endpoint; loopback needs no API key; without an LLM, keyword matching keeps everything working offline. Multi-provider routing cuts spend and avoids lock-in: one OpenRouter endpoint for every vendor's models, or the native llm_router node routing by cost / latency with automatic fallback. See LLM Routing.
• Local data & API connectors — named, explicitly-authorized connectors for directories, databases and HTTP APIs (files / notes / sqlite / mysql / postgres / http); credentials live only in the secrets store, permission ceilings can be tightened but never loosened. See Connector API.
• Deterministic runtime — DAG parallel scheduling (TLA+ formally verified), WAL crash recovery + --resume, Saga transaction compensation, idempotency, retry / rate limit / circuit breaker. Every operation is traceable, replayable, verifiable.
• Dual Agent + Workflow mode — conversational ReAct Agent (aflare chat) and daemon Agent (aflare agent) share one core; 6 pluggable capabilities (reflection / human-in-the-loop / utility / memory / planning / workflow).
• Agent interconnection & commanding — aflare directs and supervises other agents: CLI channel (codex / claude / gemini or any generic CLI) and A2A protocol channel, with real delegation via the supervisor node and failure isolation per agent.
• Security built in — HMAC tamper-evident audit chain, AES-GCM encrypted secrets, automatic secret redaction, SSRF / path-traversal / command-injection defenses, outbound anomaly monitoring + auto circuit-break, four security levels (L0–L3).
• Extensible ecosystem — MCP Server / Client, custom nodes in Go, community plugins, GitHub Action for CI, 30+ built-in LLM providers, OpenClaw plugin for the OpenClaw ecosystem.
• Ready-to-run examples — real-world workflow packs under examples/real-world/: industrial monitoring (OpenFOAM divergence watchdog, similarity-RAG incident triage), DevOps CI pipelines, research, batch processing, multi-agent role pipelines (analyst→researcher→trader→risk trading crew, digital-company marketing & sales departments), and AI after-sales customer service (emotion detection → product disambiguation → photo error-code OCR → RAG troubleshooting → self-fix or escalation → ticket archiving, fully offline-capable via layer-by-layer fallback).

Security

Four security levels (--security-level): L0 relaxed → L3 maximum (L2 refuses unsandboxed code_interpreter; L3 disables it). CI runs gofmt / go vet / gosec / govulncheck on every PR.

Security Guide →

Documentation

• Getting Started · Tutorial · YAML Syntax
• Dataflow · Scheduling · MCP · Plugins · Connectors · LLM Routing
• Web UI · Visualizer · Custom Nodes
• API Reference · Nodes Reference
• Deployment · Docker · Multi-Tenancy
• Troubleshooting · Changelog

Contributing

We welcome contributions! Contributing →

License — Dual Licensing

aflare ships under a dual license:

• Community Edition — GNU AGPL v3.0, free (LICENSE). Note AGPL §13: offering aflare-based functionality as a service triggers source-disclosure obligations for the entire combined work.
• Commercial License — for embedding aflare in closed-source products or SaaS without AGPL obligations (LICENSE-COMMERCIAL.md, local_first_agent@126.com).