AI Fabric

AI Fabric is a UX-first orchestration studio for building and running connected sub-agent workflows with:

• multi-workspace management
• visual graph editor for agent topology
• configurable worker/reviewer nodes
• conditional human-in-the-loop gate (only when worker has assigned reviewer)
• persistent run history (prompts, outputs, decisions)
• full flow execution from entry prompt (entire configured chain)
• model discovery via OpenAI models.list()
• local RAG memory with SQLite chunks + embeddings JSON + rerank
• runtime provider per node (openai_api or cursor_bridge)
• session controls per node (clear context, switch dialog)
• bridge API contract document at docs/cursor-bridge-contract.md

Current MVP

Implemented now:

• Workspace creation/list page
• Workflow editor based on React Flow
• node configuration (name, role, model, instructions, reviewer, success criteria)
• node runtime controls (provider, temperature, clear-context-on-run)
• workspace switcher directly in studio header
• graph persistence to SQLite via Prisma
• run creation and per-node execution
• run full chain in topological order with pause/resume on review gates
• modal quality gate with Approve / Reject with note for reviewer-linked workers
• event persistence in WorkflowRun, NodeRun, HumanDecision
• API endpoint for available models with fallback list
• local knowledge base API and UI:
• add documents
• auto-chunking
• embeddings generation (OpenAI or fallback local embedding)
• hybrid rerank search (semantic + lexical)
• RAG context injection into node execution
• agent session API:
• list sessions by workspace
• clear context
• switch dialog
• cursor bridge runtime adapter with OpenAI fallback

Tech Stack

• Next.js 16 (App Router, API routes)
• TypeScript
• React Flow
• Prisma 7 + SQLite (better-sqlite3 adapter)
• OpenAI Node SDK (responses API)

Run Locally

1. Install dependencies:

npm install

2. Set environment variables:

# .env
DATABASE_URL="file:./dev.db"
OPENAI_API_KEY="" # optional
CURSOR_BRIDGE_URL="http://127.0.0.1:4141" # optional, bridge adapter endpoint
# optional codex fallback controls:
# CODEX_BIN="C:\\Users\\User\\.cursor\\extensions\\openai.chatgpt-...\\bin\\windows-x64\\codex.exe"
# CODEX_EXEC_TIMEOUT_SECONDS="300"
# CURSOR_BRIDGE_TIMEOUT_MS="315000"

3. Push schema and generate Prisma client:

npx prisma db push
npm run db:generate

4. Start development server:

npm run dev

5. (Optional) Start local Cursor bridge runtime:

npm run bridge:dev

6. Open:

http://localhost:3000

One-click Windows launch

You can also start everything by double-clicking:

start-ai-fabric.bat

This script checks:

• Node.js and npm availability
• dependency installation (node_modules)
• .env presence (creates from .env.example if missing)
• Prisma client generation + DB sync
• free ports 3000 and 4141

Useful Scripts

npm run dev
npm run bridge:dev
npm run lint
npm run build
npm run db:generate
npm run db:migrate
npm run db:studio

Data Model (Core)

• Workspace
• Workflow (stores graph JSON)
• WorkflowRun
• NodeRun
• HumanDecision
• AgentSession (prepared for per-node memory/session metadata)
• KnowledgeDocument
• KnowledgeChunk

Next Up (Recommended)

1. Full graph execution engine:

• topological traversal
• retries by quality gate
• conditional branching

2. Cursor bridge hardening:

• stream token-level events from bridge to UI
• strict bridge-only mode toggle (disable fallback by policy)
• richer bridge diagnostics per node run

3. Optional RAG upgrade path:

• move from local SQLite retrieval to pgvector or dedicated vector DB
• add metadata filters and advanced rerank strategies

4. Multi-user readiness:

• auth and roles
• audit trail and immutable event log
• workspace permissions