LoomFlow

English | 中文

A lightweight AI workflow builder for individuals and small teams.

An AI-native workflow runtime with built-in evaluation, evolution, and quality gates.

Describe your idea in natural language → generate a runnable workflow → customize it on a visual canvas → publish it as an API.

Everyone should be able to create their own AI automation flows.

⚡ Quick Start (Docker, self-hosted):

git clone https://github.com/banmu123/LoomFlow.git && cd LoomFlow && bash scripts/init-env.sh && docker compose up -d

→ open http://localhost:5000 · default admin admin / 123456 (change immediately) · full guide: 🚀 Quick Start

📸 Screenshots

🎯 Who is it for?

• ✅ Indie developers — ship an AI feature without building an orchestration platform
• ✅ Content creators — turn recurring production steps into one-click automated flows
• ✅ Small studios — deliver demos and client work faster with shareable, runnable flows
• ✅ AI automation enthusiasts — prototype your idea in minutes, then deploy it for real

🤔 Why LoomFlow instead of Dify / n8n?

In one sentence: Dify builds LLM app platforms for teams, n8n does general automation for everyone — LoomFlow turns "one sentence" into a runnable, publishable workflow for individuals and small teams. It doesn't chase comprehensiveness — it pursues a 30-second learning curve and a one-minute deployment.

Scope: LoomFlow is not an AI coach, a personal-growth app, or a coding-practice platform. It stays deliberately narrow — one pipeline: natural language → workflow → canvas → API. Anything that doesn't serve that loop is kept out of the core product.

✨ Highlights

💬 Natural Language → Workflow

No drag-and-drop required to start — just describe your process in plain language, and the AI generates an executable workflow directly onto the canvas:

You: "Build me a workflow: input a product name → AI generates selling-point copy → generates a promo video script"

AI: ✅ Generated a 4-node workflow (Start → LLM Copy → LLM Script → End), loaded to canvas

Then fine-tune visually and publish as an API.

🎨 Visual Canvas

Tinyflow canvas editor: drag nodes, connect flows, configure parameters. 12 node types (LLM / HTTP / Code / Template / Search / Excel / Loop / Human Confirm, etc.).

🚀 Workflow as API

One-click publish a workflow as an HTTP endpoint — one global API Key calls all of your published workflows, with unlimited calls and call logs:

curl -X POST https://your-host/api/publish/{workflowId}/execute \
  -H "Authorization: Bearer ffk_xxxxx" \
  -H "Content-Type: application/json" \
  -d '{"inputs": {"query": "..."}}'

The API Key is auto-generated on first publish and shown only once; when it expires, regenerate it on the API Keys page.

🔗 Shareable Workflow Pages

Generate a public link — recipients can view nodes, fill inputs, and run the workflow without signing in. Perfect for demos and delivery.

🏢 Team & Permissions

• Full data isolation (including admin)
• No call limits — chat & API calls are unlimited
• Audit logs for all critical operations

📊 Admin Dashboard

User management, usage statistics (trends), audit logs, API call logs.

🌐 Internationalization

Chinese/English one-click switch (framework supports any language).

🔒 Self-Hosted & Private

App, data, and storage can all be deployed on your own server. Database can be switched to self-hosted PostgreSQL (see Deployment Manual).

🧩 Extensible Node System

NodeRegistry + NodeDefinition — a single source of truth for nodes. Custom nodes can be registered with one entry, with automated validation (executor binding, start/end singleton).

📝 Brew Notes (Workflow Notes)

Record why a workflow is designed this way — decisions, problems, solutions, optimizations and usage. AI can summarize design intent, suggest notes from run history, and the canvas AI assistant answers "why did I choose X" from your notes.

🕵️ Execution Trace & Debug Assistant

Node-level execution trace after each run: per-node status, duration, model, tokens and errors in a timeline — see exactly where a workflow got stuck. The canvas AI assistant reads run history and answers "why did this run fail" with root cause and fix suggestions.

🤖 Canvas AI Assistant

Chat with the canvas: ask about the current workflow, describe a change ("add a summarizer after search"), and the AI outputs a complete workflow JSON you apply to the canvas in one click.

🧠 Bring Your Own Model

Add any model (DeepSeek / Ark / any OpenAI-compatible endpoint) through the admin UI — no code changes:

• Per-model API key & base URL (overrides environment defaults)
• Declare capabilities (text / vision / ...) — vision models enable image input automatically
• Canvas & chat model lists sync instantly after adding

Admin → Model Settings → Add Model
  id: qwen-vl-max · provider: openai-compatible
  base URL: https://... · api key: sk-...
  capabilities: [text, vision]  →  ✅ instantly available everywhere

🔄 Workflow Evolution Engine

Workflows continuously improve — without manual intervention. The Evolution Engine automatically detects performance regressions and proposes optimizations:

• Trigger rules: scheduled (cron), metric-based (latency +30%, failure rate > threshold), event-based (N consecutive failures)
• Regression detection: automatic baseline comparison (version / production / rolling) across 5 metrics (success rate, failure rate, P95 latency, cost, test score) with relative + absolute thresholds
• AI analysis: metrics → bottleneck detection → static analysis → AI generates optimization patch
• Human approval: proposals require explicit user confirmation before modifying production workflows
• Evolution history: full traceability — trigger → analysis → proposal → decision → version change → outcome (before/after metrics)
• Full audit trail: every trigger, analysis, and decision is recorded in evolution events

Evolution Dashboard → Workflow Health (score, trend, bottlenecks)
  → AI Proposals (view diff, approve, reject)
  → Trigger Rules (create cron / metric / event triggers)
  → Regression Detection (baseline vs candidate, 5 metrics, severity policy)
  → Evolution History (timeline, before/after outcome, version diff)

🏗️ Architecture

flowchart TB
    subgraph Client["Clients"]
        Browser["Browser — Next.js frontend<br/>AI Chat / Canvas / Admin / Share"]
        External["External systems<br/>curl / API consumers"]
    end

    subgraph App["LoomFlow App (Next.js)"]
        UI["App Router pages<br/>Chat / Workflows / Admin / Share"]
        API["API Routes<br/>auth / chat-ai / workflow-history / publish / api-key<br/>search-providers / nodes / schedules"]
        Engine["Workflow engine<br/>FlowEngine + NodeRegistry + Executors"]
        Registry["Registries<br/>Model Registry · Search Provider Registry"]
    end

    subgraph Data["Data layer"]
        PostgREST["PostgREST"]
        PG[("PostgreSQL<br/>conversations / workflow_history / workflow_versions<br/>user_api_keys / ai_models / search_providers<br/>node_definitions / audit_logs ···")]
    end

    subgraph ExternalSvc["External services"]
        LLM["LLM Providers<br/>DeepSeek / any OpenAI-compatible endpoint"]
        OSS["Object storage<br/>Aliyun OSS / S3-compatible"]
        Search["Search Providers<br/>Tavily / Exa / Google"]
    end

    Browser --> UI
    External -->|"Authorization: Bearer API Key"| API
    UI --> API
    API --> Engine
    API --> Registry
    API --> PostgREST --> PG
    Engine --> LLM
    Engine --> Search
    API --> OSS

How it flows: describe a process in natural language → AI generates a workflow (validated & auto-repaired) → edit on the visual canvas → save as versioned history → publish a chosen version as a secured HTTP API (one global API key per user) → external systems call it with Authorization: Bearer <key>.

Docker self-hosted deployment:

flowchart LR
    User["User"] -->|":5000"| App["loomflow app"]
    App -->|"http://nginx:80/rest/v1"| Nginx["Nginx reverse proxy"]
    Nginx --> PostgREST["PostgREST"]
    PostgREST --> PG[("PostgreSQL 16<br/>volume: loomflow-pgdata")]
    Mig["migration container<br/>idempotent SQL on every `up`"] -.-> PG

🚀 Quick Start

🐳 Docker (recommended, one-click self-hosted)

Includes **PostgreSQL + PostgREST