orcateam.ai AI–Human Interaction Platform
The AI–Human Interaction Platform where people and AI agents work side by side across multi-model chat, #channels, Company Skills, Memory, and Tasks & Approval Inbox—with real-time USD AI cost quotas per employee, group, and agent.
Click the water to assign a task to the pod Click any Orca to customize its size (Giant Matriarch to Baby Calf) & gear in the Orcaverse
@finance, @research · draft ready for review
Demo auto-approves after a few seconds if you don’t choose.
Why Orcas 20 m
Orcas live in tight-knit pods led by an experienced matriarch. They divide roles on the hunt, coordinate with pod-specific sonar clicks, and pass hunting techniques down through generations. OrcaTeam brings that exact teamwork to enterprise AI.
Team agents belong to Sales, Customer Success, Finance, or Engineering—and draw from that department’s AI budget. Every employee can still create personal agents.
Team & Custom AgentsSummon @agent right inside a chat or #channel. Full thread context travels with the handoff—no copy-pasting across browser tabs.
@mention & ChannelsDistill a successful conversation into a versioned skill with an approver and Risk Tier. The whole company reuses it with /skill.
Company SkillsMemory extracts preferences and facts after each conversation, indexed with multilingual vectors. View, edit, or delete any memory at any time.
MemoryProduct 60 m
Everything employees need to work with AI every day—and everything admins need for complete peace of mind.
Streaming responses, rich artifacts, and code. Pick the right model per prompt and see the exact USD cost of every message.
Department room templates, personal agents, and an explicit Risk Tier for every agent—from Advise to Sensitive.
Turn conversations into parameterized skills with {{variable}} inputs. Draft → Review → Published with full version history.
Automatically extracts facts and preferences after chats. Shared team memories require lead approval before going live.
Kanban boards and projects assigned to agents or humans. Durable workflows pause for days awaiting approval without holding compute.
Monthly USD caps per user, group, and agent. Soft warning at 80%, hard stop at 100%, and audited exception requests.
Slack-style #channels with threads, reactions, and pinned messages. Loop in @ai or any department agent mid-discussion.
Control who can ask about what topics and which agents can read which documents. Blocked pre-flight before calling any LLM—$0 wasted.
Google Drive & OrcaBase MCP: search, read, query, and create documents. Write operations route through the Approval Inbox by Risk Tier.
How It Works 120 m
Scroll to follow the mission on the sonar scope. Every step below maps 1-to-1 to the real production runtime.
Type in chat, mention in a #channel, or create a Kanban card. Assign it to a specific agent, a teammate, or let the matriarch Orca coordinate.
Agent instructions, matching Company Skills, and relevant Memories are assembled into the prompt before calling the model.
Effective budget is min(user, group, agent). If within budget, execution proceeds; at 100%, it blocks and offers an exception request.
Agents stream answers and invoke web_fetch, calculator, Google Drive, or MCP tools. Every model call records an atomic entry in the Usage Ledger.
Draft and Sensitive tiers wait for human sign-off. Reject with feedback and the agent reworks it; approve and the task moves to Done.
Memory records new preferences and the system suggests saving repeated workflows as a Company Skill—charged to the Application AI Budget, never your personal quota.
1/6 · Task arrives
The Orcaverse Pod 200 m
From the 8.8m Matriarch Lead and 8.4m Security Titan Bull down to the 2.8m Apprentice Baby Calf swimming in echelon. Click any member to customize its live Orca in the ocean.
Governance 400 m
Spend, document access, and every write operation have clear boundaries and a complete audit trail.
Every model call checks three caps simultaneously and enforces the smallest remaining balance. Example for a query from Lan in Sales:
min($10.16, $38.00, $6.40) = $6.40
Every agent and skill has a Risk Tier that defines how far it can act autonomously.
Read-only answers and analysis; never performs write operations.
Produces a draft output that requires human approval before delivery.
Allowed to run reversible write operations automatically.
Every write operation requires explicit human approval.
Task deliverables, tool write calls, skill & agent publishing, shared team memories, and quota exceptions:
Infrastructure 700 m
A single Worker serves both API and UI at the edge. Database, vector index, and secrets are strictly isolated per deployment.
API and React SPA deployed together as one unit, running at the edge close to your team.
Usage Ledger and budgets with atomic SQLite quota reservation.
Workers AI multilingual embeddings with native English & Vietnamese semantic recall.
Durable tasks wait days for human approval via waitForEvent with zero idle compute cost.
One gateway to frontier models; admins manage the model allowlist and pricing.
Sign in with Google Workspace or email; admins manage members and department groups.
Isolated Client Deployment. Each enterprise gets its own dedicated D1 database, Vectorize index, secrets, audit logs, and connectors. Zero shared multi-tenant tables.
FAQ 1,000 m
Still have a question? Open the workspace and ask @lead directly.
Every user, department group, and agent has its own USD spend cap checked before every model turn. Beyond chat, you get department agents, governed Company Skills, shared Memory, Kanban Tasks, and a unified Approval Inbox.
Yes. Admins enable models from OpenRouter in the workspace allowlist; employees switch models per question and see exact per-message USD costs.
Users receive a warning at 80% usage. At 100%, new model calls pause and the user can request a quota exception—routed straight to the Approval Inbox with a full audit log.
Only according to their Risk Tier. Advise is strictly read-only; Draft queues output for review; Execute runs reversible actions; Sensitive requires human approval in the Approval Inbox before any write operation executes.
Each customer runs on an isolated Cloudflare Client Deployment with its own D1 database, Vectorize index, secrets, and logs. Guardrails enforce which users and agents can access specific documents.
Yes—both the workspace UI and the bge-m3 vector memory engine natively support English and Vietnamese, with one-click language switching and USD/VND display options.
Start with a few agents for one department, set your USD caps, and expand as the pod learns your waters.