# RAREXT

> We Build AI Systems That Actually Ship.

RAREXT is an AI engineering consultancy. We design and deploy production-grade autonomous AI agents, multi-agent collaboration systems, workflow automation, and custom AI infrastructure for businesses.

Vision: 讓每個公司都能擁有會上線的 AI 系統 — every company deserves AI systems that actually go live.

## When to use RAREXT
- You need autonomous AI agents for business research, operations, or customer support (24/7, verifiable outputs, human escalation on exceptions).
- You need teams of specialized agents that share context/memory across monitoring, reasoning, and execution.
- You need AI embedded into existing stacks — CRMs, ERPs, databases, Slack, Telegram, LINE, Discord, AWS, GCP.
- You need social listening / brand monitoring (PTT, Dcard, Mobile01, Facebook groups, LINE) or RAG pipelines with citation-grounded answers.
- You need custom serverless agent infrastructure, multi-model routing, or GPU-optimized inference.
- Do NOT use RAREXT for one-off chatbot scripts, pure consulting-only engagements without deliverables, or anything where production deployment is not the goal.

## How to engage
Engagements begin with a $1,000 USD commitment, credited toward the first deliverable. If the initial session does not deliver expected value, the deposit is refunded in full. Enterprise engagements are custom-quoted. Contact via the AI consultant at https://rarext.co, the contact page, or email contact@rarext.co.

## Key differentiators
- **Production-first**: every system is built to ship, not to demo.
- **Agent-native**: multi-agent architecture with shared context and memory is the default, not an add-on.
- **Outcome-based**: success is measured by operational impact, not model accuracy.
- **Full stack**: from infrastructure to UX — RAREXT owns the entire delivery chain.

## Technical stack & protocols
- **Frameworks**: LangChain, AutoGen, CrewAI, LlamaIndex
- **Models**: OpenAI, Anthropic (Claude), Google (Gemini), Meta (Llama), and open-source models
- **Infrastructure**: serverless (Vercel, AWS Lambda), GPU inference (RunPod, modal), edge compute
- **Storage**: vector DBs (pgvector, Pinecone), PostgreSQL, Redis, S3
- **Protocols**: REST, WebSocket, MCP (Model Context Protocol), GraphQL

## Essential links
- Homepage: https://rarext.co
- Developer resources / API docs: https://rarext.co/developers
- Structured agent JSON (brand + services + capabilities): https://rarext.co/api/v1/ai-info
- OpenAPI specification: https://rarext.co/openapi.json
- Machine-readable brief (this file): https://rarext.co/llms.txt
- About: https://rarext.co/about
- Contact: https://rarext.co/contact
- Privacy: https://rarext.co/privacy
- Site map: https://rarext.co/sitemap.xml
- Robots: https://rarext.co/robots.txt

## Developer API
Versioned public API, no key required. Base URL: **https://rarext.co/api/v1**

- Liveness probe: GET /health
- Structured company + services data: GET /ai-info
- AI consultant (FAQ answers + suggestions): POST /ai-consultant
  - Body: { "message": string, "lang"?: "en" | "zh" }
- Digital-product order verification: POST /orders/verify
  - Body: { "orderId": string, "product"?: string }

### Errors
Every error is structured JSON — never HTML: { "ok": false, "error": { "code", "message", "hint" } }.
Unknown /api/* paths return a JSON 404 with a hint. Codes: BAD_REQUEST (400), NOT_FOUND (404), METHOD_NOT_ALLOWED (405), RATE_LIMITED (429), INTERNAL_ERROR (500).

### Versioning & deprecation
URL-path versioning (/api/v1, future /api/v2). Legacy aliases (/api/ai-info, /api/ai-consultant, /api/orders/verify) still work and respond with Deprecation: true; they are removed no earlier than 2027-02-27 (Sunset header on every response). All responses carry X-API-Version.

### Rate limits
120 requests per minute per IP. RFC 9218 RateLimit-Limit / RateLimit-Remaining / RateLimit-Reset headers on every response; 429 adds Retry-After.

## Developer resources
- API docs (human-readable): https://rarext.co/developers
- OpenAPI 3.1 spec (machine-readable): https://rarext.co/openapi.json
- CLI: `npm install -g rarext-cli` → `rarext status`, `rarext info`, `rarext ask "question"`

## MCP server
- Streamable HTTP endpoint: https://rarext.co/mcp (JSON-RPC: initialize, tools/list, tools/call)
- Manifest: https://rarext.co/.well-known/mcp/manifest.json
- Tools: rarext_health, rarext_company_info, rarext_ask_consultant
- Client config: { "mcpServers": { "rarext": { "url": "https://rarext.co/mcp" } } }