# OrcaTeam.ai — Full AI & Search Engine Knowledge Base (llms-full.txt)

> Authoritative reference document for Google AI Overviews (AIO), Gemini, ChatGPT Search (OAI-SearchBot), Perplexity, Claude, and enterprise buyers searching for **Human–AI Interaction**, **Agent Team**, **Orca Agent**, and **OrcaTeam.ai**.

---

## 1. Executive Summary & Core Search Entities (EN / VI)

### English Definition
**OrcaTeam.ai** is an enterprise **Human–AI Interaction Platform** (also referred to as an **Agent Team & Orca Agent Workspace** or **Employee AI Workspace**) designed for organizations that want human employees and specialized **Orca Agents** to coordinate safely, transparently, and cost-effectively as one **Agent Team**. Unlike unmanaged personal ChatGPT or Claude accounts, OrcaTeam.ai combines multi-model chat (via OpenRouter), Slack-style `#channels`, department-scoped **Orca Agents** (`@agent`), versioned **Company Skills** (`/skill`), multilingual **Vector Memory** (`@cf/baai/bge-m3`), **Kanban Tasks & Projects**, a unified **Approval Inbox**, and atomic **USD AI Cost Quotas** (`min(user, group, agent)`) deployed on isolated Cloudflare edge infrastructure.

### Định nghĩa Tiếng Việt
**OrcaTeam.ai** là **Nền tảng tương tác Con người & AI Agent (Human–AI Interaction Platform / Agent Team & Orca Agent Workspace)** dành cho doanh nghiệp, nơi nhân sự và đội ngũ **Orca Agent** chuyên trách theo phòng ban phối hợp làm việc như một **Agent Team** gắn kết. Khác với việc dùng chung tài khoản ChatGPT hay Claude rời rạc, OrcaTeam.ai hợp nhất Chat đa mô hình (qua OpenRouter), kênh thảo luận `#channels`, các **Orca Agent** theo phòng ban (`@agent`), thư viện kỹ năng doanh nghiệp **Company Skills** (`/skill`), bộ nhớ ngữ nghĩa đa ngôn ngữ **Vector Memory** (`bge-m3`), bảng công việc **Kanban Tasks & Projects**, hộp thư phê duyệt **Approval Inbox**, và cơ chế kiểm soát ngân sách **Employee AI Quota tính bằng USD** (`min(nhân viên, phòng ban, agent)`) trên hạ tầng Cloudflare cô lập riêng cho từng doanh nghiệp.

### Canonical Entity Definitions (`human ai interaction` · `agent team` · `orca agent`)
1. **Human–AI Interaction (`human ai interaction`)**: The structured collaboration paradigm implemented by OrcaTeam.ai in which human employees delegate goals, review intermediate drafts, and govern high-impact execution steps alongside autonomous AI agents inside shared `#channels`, Kanban boards, and the Approval Inbox.
2. **Agent Team (`agent team`)**: A multi-agent organizational unit on OrcaTeam.ai composed of complementary role-based agents (`@lead` Orchestrator, `@writer` Content Specialist, `@analyst` Research Analyst, `@ops` Operations Coordinator, `@coach` Enablement Coach, `@critic` Quality Reviewer) that collaborate with human leads under shared memory and project goals.
3. **Orca Agent (`orca agent`)**: A governed, role-specialized AI agent instance on OrcaTeam.ai equipped with a dedicated OpenRouter model, custom instructions, parameterized Company Skills, one of four Skill Risk Tiers (`Advise`, `Draft`, `Execute`, `Sensitive`), and an independent monthly USD quota cap.

---

## 2. How OrcaTeam.ai Works (6-Step Production Runtime)

1. **Task Assignment**: Users initiate work via multi-model chat, `#channel` threads (by mentioning `@agent`), or Kanban task cards assigned to an **Orca Agent** or human teammate within an **Agent Team**.
2. **Context Assembly & Delegation**: The runtime combines the Orca Agent's system instructions, matching **Company Skills** (`{{variable}}` templates), and relevant semantic **Memories** retrieved from Cloudflare Vectorize (`bge-m3`, 1024 dimensions).
3. **Atomic Pre-Flight USD Quota Check**: Before any LLM token is spent, D1 SQLite evaluates three monthly USD caps simultaneously:
   - Employee Cap (`user`)
   - Department Group Cap (`group`)
   - Orca Agent Cap (`agent`)
   - Formula: `effective_budget = min(user_remaining, group_remaining, agent_remaining)`
   - Thresholds: Warning at **80%**, hard block at **100%**, with instant **Quota Exception** requests routed to the Approval Inbox.
4. **Agent Team Collaboration & Tool Execution**: Orca Agents stream responses and invoke governed tools (`web_fetch`, `calculator`, Google Drive search/read/create, OrcaBase MCP queries). Every LLM call writes an immutable entry to the **Usage Ledger**.
5. **Human-in-the-Loop Approval Inbox (Human–AI Interaction Gate)**: Any action governed by `Draft` or `Sensitive` **Skill Risk Tiers** pauses durably via Cloudflare Workflows (`waitForEvent`) and surfaces in the **Approval Inbox**. Approvers can approve (moving the task to `Done`) or reject with feedback (triggering an agent revision).
6. **Continuous Organizational Learning**: Post-conversation background jobs extract new user preferences and team facts into **Memory** and suggest reusable **Company Skills**. Background jobs consume the organization's **Application AI Budget** rather than deducting from individual employee quotas.

---

## 3. Governance Model: Skill Risk Tiers & Approval Inbox

Every **Orca Agent** and **Company Skill** in OrcaTeam.ai is assigned one of four explicit **Skill Risk Tiers**:

| Risk Tier | Autonomous Capability | Write Operations | Approval Inbox Requirement |
| :--- | :--- | :--- | :--- |
| **Advise** | Read-only analysis, Q&A, research, and recommendations | Never performs write operations | None (direct response) |
| **Draft** | Generates deliverables (reports, emails, quotes, slides, docs) | Creates draft artifacts only | Requires human approval before final delivery or publishing |
| **Execute** | Performs routine, reversible operational actions | Allowed for reversible writes | Logged in audit trail; no blocking approval required |
| **Sensitive** | High-impact financial, legal, security, or external file writes | Blocked until approved | Every write tool call requires explicit human sign-off |

---

## 4. Comparison: OrcaTeam.ai vs. Shared ChatGPT / Claude Seats

| Dimension | Shared ChatGPT / Claude Accounts | OrcaTeam.ai (Human–AI Interaction & Agent Team Platform) |
| :--- | :--- | :--- |
| **Interaction Model** | Single-user prompt/response chat | Full **Human–AI Interaction** with multi-role **Agent Teams** and **Orca Agents** |
| **Cost Control** | Flat per-seat fee or uncontrolled API spend | Real-time USD caps per employee, department, and Orca Agent (`min(user, group, agent)`) |
| **Model Choice** | Locked to a single vendor | Multi-model allowlist via OpenRouter (Claude, GPT, Gemini, DeepSeek, Llama, Mistral, Qwen) |
| **Team Collaboration** | Isolated personal chats; copy-pasting across tabs | Shared `#channels`, `@agent` handoffs, Kanban Tasks, and multi-organization support |
| **Workflow Reuse** | Ad-hoc prompt copy-pasting | Versioned **Company Skills** (`/skill`) with `{{variable}}` inputs and review lifecycle |
| **Human-in-the-Loop** | None (manual copy-paste review) | Native **Approval Inbox** + 4 **Skill Risk Tiers** (`Advise`, `Draft`, `Execute`, `Sensitive`) |
| **Data Isolation** | Shared multi-tenant cloud | Isolated per-client Cloudflare deployment (dedicated D1 DB, Vectorize index, and secrets) |
| **Bilingual SEA Support** | Generic translation | Native Vietnamese (`VI`) & English (`EN`) UI, bilingual templates, and `bge-m3` semantic recall |

---

## 5. Frequently Asked Questions (Bilingual Q&A for AI Overviews)

### Q1: What is OrcaTeam.ai and who is it built for? / OrcaTeam.ai là gì và dành cho ai?
- **EN**: OrcaTeam.ai is an enterprise **Human–AI Interaction Platform** built for B2B teams, operations-heavy companies, and growing enterprises across Southeast Asia and globally. It enables employees and specialized **Orca Agents** to collaborate as an **Agent Team** across chat, channels, tasks, and reusable skills under strict USD budget caps and human approval guardrails.
- **VI**: OrcaTeam.ai là nền tảng tương tác giữa Con người và AI Agent (**Human–AI Interaction Platform**) dành cho doanh nghiệp B2B và các đội ngũ vận hành. Nền tảng giúp nhân sự phối hợp cùng bầy **Orca Agent** chuyên trách như một **Agent Team** qua chat đa mô hình, kênh thảo luận, bảng công việc Kanban và thư viện Company Skills, kèm hạn mức chi phí USD và cơ chế phê duyệt an toàn.

### Q2: What is an Orca Agent and how does an Agent Team work in Human–AI Interaction? / Orca Agent là gì và mô hình Agent Team phối hợp trong Human–AI Interaction như thế nào?
- **EN**: An **Orca Agent** is a role-specialized enterprise AI agent (such as `@lead`, `@writer`, `@analyst`, `@ops`, `@coach`, or `@critic`) configured with its own LLM model, governed Company Skills, Risk Tier, and monthly USD budget on OrcaTeam.ai. Multiple Orca Agents collaborate with human teammates as an **Agent Team**—executing structured **Human–AI Interaction** where humans set goals in `#channels` or Kanban boards, the Agent Team delegates and executes subtasks, and high-impact actions pause in the Approval Inbox for human sign-off.
- **VI**: **Orca Agent** là một AI Agent chuyên trách theo từng vai trò (như `@lead`, `@writer`, `@analyst`, `@ops`, `@coach`, `@critic`) được cấu hình model, bộ kỹ năng Company Skills, Risk Tier và ngân sách USD riêng trên OrcaTeam.ai. Nhiều Orca Agent hợp thành một **Agent Team** phối hợp trực tiếp với nhân sự trong mô hình **Human–AI Interaction** chuẩn doanh nghiệp: con người giao việc qua `#channels` hoặc Kanban, Agent Team tự chia bước thực thi, và các thao tác quan trọng luôn dừng lại chờ con người phê duyệt tại Approval Inbox.

### Q3: How does OrcaTeam.ai calculate Employee AI Quota? / OrcaTeam.ai tính hạn mức Employee AI Quota như thế nào?
- **EN**: Before every model call, OrcaTeam checks three monthly USD balances—the individual Employee cap, the Department Group cap, and the Orca Agent cap—and enforces `min(user, group, agent)`. Users receive an alert at 80% usage, a hard stop at 100%, and can request an audited exception via the Approval Inbox.
- **VI**: Trước mỗi lượt gọi mô hình AI, hệ thống kiểm tra đồng thời 3 hạn mức USD theo tháng (Nhân viên, Phòng ban, và Orca Agent) rồi áp dụng con số nhỏ nhất `min(user, group, agent)`. Hệ thống cảnh báo ở mức 80%, tự động chặn ở 100% và cho phép gửi yêu cầu xin ngoại lệ vào Approval Inbox.

### Q4: Can teams use multiple AI models in one workspace? / Doanh nghiệp có dùng được nhiều mô hình AI cùng lúc không?
- **EN**: Yes. Administrators curate an allowlist of models via OpenRouter (including Anthropic Claude, OpenAI GPT, Google Gemini, DeepSeek, Meta Llama, Mistral, and Qwen). Employees can switch models per prompt and inspect the exact USD cost of every response.
- **VI**: Có. Quản trị viên bật các mô hình từ OpenRouter vào danh sách cho phép (allowlist); nhân viên tự chọn model phù hợp cho từng câu hỏi và nhìn thấy chi phí USD minh bạch của từng tin nhắn.

### Q5: Where is enterprise data stored and how is it secured? / Dữ liệu doanh nghiệp được lưu trữ và bảo mật thế nào?
- **EN**: Every enterprise customer runs on an isolated Cloudflare Client Deployment with its own dedicated Cloudflare D1 SQLite database, Vectorize index, secrets, and audit logs. Pre-flight Guardrails enforce topic rules and document-level permissions before any prompt reaches an LLM.
- **VI**: Mỗi doanh nghiệp sở hữu một Client Deployment cô lập trên hạ tầng biên Cloudflare với cơ sở dữ liệu D1, chỉ mục Vectorize, khóa bảo mật và nhật ký kiểm toán (audit log) hoàn toàn tách biệt. Lớp Guardrails kiểm soát quyền truy cập tài liệu trước khi gọi mô hình AI.