🤖
AI Minis
17 specialists
17 role-based AI employees with distinct personalities
Custom knowledge bases per agent
Voice, text, phone, and workflow modes
Example: Marketing Mini drafts a LinkedIn post about your product launch, using your brand voice from training data
Founder, Developer, Marketing, Designer, Ops, Research, Copywriter, Outreach, Client, Analyst, Trading, Crypto, Studio, Vision, Aria, Travel, Events
System prompt templates with personality and domain expertise
Per-agent avatar colors and voice configurations
Agent-specific default tool assignments
Knowledge base loaded from markdown files per agent
Seamless switching between Minis in conversation
Avatar (AI Self) sits above all Minis with cross-company visibility
View all
🖥️
Mini Workstations
Local Mac compute fleet
Enroll a Mac mini (or any Mac) — Minis get iMessage + WhatsApp Personal
Browser automation, computer use, shell, AppleScript
Default-deny policy with per-contact ACL + kill switch
Example: "Send mom a happy birthday message on iMessage at 9 AM tomorrow" — the Mac mini sends it from your real number, no Apple Business Chat required
iMessage — send / read / watch via macOS chat.db + AppleScript
WhatsApp Personal — pair via Baileys (your real number, your contacts)
Browser automation — local Chromium with your saved logins
Computer use — full screen + keyboard + mouse access
Shell — allowlisted commands only
Ed25519-signed WSS — every frame verified
Capability tokens — HMAC-signed, scoped, replay-protected
Hash-chained audit log — tamper-evident JSONL of every action
Two-stage injection filter — LLM output parsed + planned + canary-leak-checked before any capability fires
Money-mention override — auto-forces approval on dollar-amount mentions even for trusted contacts
Inbound triggers — iMessage / WhatsApp arrive, fire matching workflow, Mini drafts reply
One-line install — node mini-workstation/agent.js enroll <token> + agent.js run
launchd persistence — survives reboots
One-click kill switch — /workstations.html
View all
✈️
Travel Mini
24/7 flight watcher + 1-click booking
Set a watch in natural language; Travel Mini calls Duffel on adaptive cadence
Three alert tiers: mistake fare, threshold hit, 30-day low
Rich email with flight times, layovers, airline link — one-click booking
Example: "Watch SFO → BLR June 8–14 under $900, prefer non-stop" — Travel Mini emails you when a $716 SFO→BLR via TYO drops
Intent parser via gpt-4.1-mini, JSON mode
City → IATA airport-cluster resolution (SFO/SJC/OAK)
Adaptive cadence: scan more often when prices are falling
Tick-scoped offer cache (30–70% hit rate on shared routes)
Per-watch budget cap (TRAVEL_MAX_CALLS_PER_WATCH_PER_SCAN)
Daily cost gate (isHalted) — halts on Duffel call-cost ceiling
HMAC-signed approval URLs in email
Refresh offer + 5% price-gate before booking
If order fails after payment, automatic refund rollback
Per-user passport profiles + multi-traveler support
Snapshot retention (default 90 days), atomic prune
Post-alert backoff to avoid re-alerting same route
Test mode by default; DUFFEL_ENV=live to take payment
View all
🏠
Property Manager Mini
Replaces Rentvine — full property mgmt
Manages properties, tenants, leases, tickets, vendors, ledger
Parallel quote requests + owner approval cards for repairs
Per-property P&L + year-end Schedule E export
Example: Tenant Sarah texts the property number "AC broken." Mini triages severity, contacts top 3 HVAC vendors in parallel, surfaces 3 quotes to owner with approve buttons, schedules with the chosen vendor, and tracks to closed.
Properties — address, parcel, mortgage, escrow, insurance, photos, knowledge base
Tenants — phone-indexed for inbound channel routing; unique-phone-per-company invariant
Leases — pending → signed → active → ending → ended state machine
Maintenance lifecycle — OPEN → TRIAGED → QUOTING → AWAITING_OWNER → SCHEDULED → IN_PROGRESS → COMPLETED → CLOSED
Severity ladder — 1 (preventive) → 5 (life safety); auto-escalation thresholds
Vendor CRM — rating, response score, familiarity-with-property, jobsCompleted, blocklist, W9 + 1099 tracking
Ranking algorithm — 0.4×rating + 0.3×responseScore + 0.2×familiarity + 0.1×(1-load)
Append-only ledger — every rent + expense + late fee event with Schedule E line mapping
Schedule E lines — auto-categorized: 5 advertising, 7 cleaning, 9 insurance, 12 mortgage, 14 repairs, 15 supplies, 16 taxes, 17 utilities, 18 depreciation, 19 other
Per-property P&L — income vs expense by subtype, net, by Schedule E line
Year-end export — CPA-ready Schedule E grouped by property + line
Owner-approval gates — repairs > $300, severity ≥ 4, new vendors, lease renewals, deposit returns, evictions, capital improvements
Per-user / per-company isolation — principal-agent invariant; tenants of one user can never read another's records
14 LLM tools — property_create, tenant_create, lease_create, maintenance_open/triage/schedule/complete/close, vendor_create/request_quote/record_quote/rate, rent_record, expense_record, pl_report, ledger_query
View all
🎫
Events Mini
Concert + sports ticket watcher
Watch artists / sports teams across SeatGeek, Ticketmaster, Bandsintown
Alert on threshold, 30-day low, or new-show announce
Email card with performer photo, venue, listings, vendor link
Example: "Diljit Dosanjh in San Francisco June–Aug under $200" — Events Mini emails when SeatGeek lists tickets at $189
Multi-provider scanning with cross-provider deduplication
Three alert tiers configurable per watch
Performer headshot in email
"Also on Ticketmaster" cross-vendor chips
One-click "Open Vendor" CTA (Mini's World does NOT process event payments)
Heads-up disclaimer that resale prices move minute-by-minute
Per-watch dedup so a tour with 12 dates doesn't spam 12 emails
View all
🏦
Banking
Plaid — Bank of America + 12,000 banks
One-click connect via Plaid Link
Analyst Mini pulls transactions, balances, spending categories
Anomaly detection: z-score outliers + category spikes
Example: "What did I spend on Uber last month?" or "Any unusual charges this week?" — Analyst Mini answers with live Plaid data
plaid_get_transactions — date-ranged with merchant + category
plaid_get_balances — current + available per account
plaid_spending_by_category — aggregated totals
plaid_financial_summary — income/expenses/top merchants/net cashflow
plaid_detect_anomalies — z-score + category spikes
Access tokens encrypted per-user (AES-256-GCM)
Sandbox by default; PLAID_ENV=production when ready
View all
💬
Dual-Mode Chat
Text + Voice
Text chat with SSE streaming via gpt-4.1
Voice chat via WebSocket and OpenAI Realtime API
File upload support: images, PDFs, documents
Example: Ask Dev Mini to review a codebase screenshot — upload the image and get architecture feedback
Progress events: thinking, tool_start, tool_done, done
Full bidirectional PCM16 24kHz audio in voice mode
Toggle voice on/off per session from workspace header
Chat history persisted as append-only JSONL per agent per company
Tool-call loop: up to 10 iterations per message
Result compaction for long tool outputs
View all
⚡
Workflows
Multi-step AI orchestration
State machine: PENDING → RUNNING → APPROVED → COMPLETED
Human approval gates before sensitive actions
Max 2 revisions per step, 10-minute timeout
Example: Research Mini analyzes competitor → Copy Mini writes blog post → Founder reviews → Marketing Mini publishes to LinkedIn
AWAITING_REVISION state for iterative refinement
AWAITING_HUMAN state pauses for manual approval
Terminal state gate prevents re-execution
Loop cap of 100 iterations for safety
Background timeout monitor
Workflows can be triggered by schedules or events
Import/export in portable mw_schema 2.0 format
View all
🔧
Tool System
62 tools
3-tier routing: Local executors → API tools → MCP servers
Per-company tool assignments per Mini
Encrypted credential storage (AES-256-GCM)
Example: Outreach Mini uses Gmail to send personalized cold emails, then logs the interaction in HubSpot
Tool catalog with definitions, schemas, and descriptions
Local executors for Gmail, LinkedIn, WhatsApp, image gen, trading
Per-user API credentials with OAuth 2.0 flows
MCP server fallback for extensibility
Tool timeout: 15s default (configurable)
Approval gating for sensitive tool actions
View all
📞
Phone Calls
AI-powered calling
Outbound AI calls via Twilio + OpenAI Realtime
Sub-second latency with barge-in support
Live whisper: admin can inject real-time guidance
Example: Ops Mini calls a vendor to confirm delivery, takes notes, and creates a follow-up task
Call transcripts saved automatically
AI-generated call summaries
Call telemetry and quality metrics
Disconnect detection and reconnection
Configurable voice and personality per call
Integration with workflow triggers
View all
🎙
Podcast Studio
AI video production
AI-generated podcast scripts
HeyGen avatar video generation
DALL-E cover art + FFmpeg assembly
Example: Studio Mini produces a 5-minute weekly industry roundup video with AI avatar and b-roll
YouTube auto-publish with metadata
Customizable avatar selection
Background music and transitions
Script editing before generation
Thumbnail generation with DALL-E
View all
🌐
Website Builder
Natural language to web
Natural language to full HTML pages
Live preview in iframe
One-click deploy to Vercel or self-hosted
Example: Dev Mini builds a landing page from "create a SaaS pricing page with 3 tiers" and deploys to Vercel
Responsive design generation
Edit and refine with follow-up prompts
Inline CSS and JavaScript
Template library for common page types
Custom domain configuration
View all
📊
Trading
Stocks + Crypto
Alpaca stock and crypto trading
Paper and live trading modes
Capital rules and risk management
Example: Trading Mini recommends buying AAPL based on RSI analysis, waits for your approval, then executes
Approval gating on all trade executions
Position sizing rules and limits
Technical indicators: RSI, Bollinger Bands
Trade recommendation detection and cards
Position limit override nudge
Per-company Alpaca credential management
View all
🎯
Lead Generation & Campaigns
Full sales pipeline
Apollo.io, Google Maps, Hunter.io integration
Lead scoring and email verification
Multi-step email campaigns (4 emails)
Example: Outreach Mini finds 50 SaaS CTOs in NYC via Apollo, scores them, and launches a 4-email campaign
Reply tracking across campaign steps
Automatic lead enrichment
Personalized email templates per lead
Campaign analytics and open rates
CRM sync with HubSpot and Salesforce
View all
📅
Schedules & Triggers
Automation engine
Cron schedules: hourly, daily, weekly, custom
Event triggers: email, call, form, payment, webhook
Once vs continuous run modes
Example: Every Monday 9 AM, Analyst Mini generates a weekly performance report and emails it to the team
Configurable polling interval (default 5s)
Trigger states: WAITING_FOR_TRIGGER, STOPPED
Auto-delete trigger after one-time fire
Trigger-workflow linking and cleanup
Animated waiting node in flow visualization
Start/stop listening endpoints
View all
🖼
Image Gallery
AI image management
DALL-E 3 generation with prompt tracking
[[IMG-1035]] inline reference syntax
Tags, favorites, search, archive
Example: Designer Mini generates 5 logo variations, you favorite the best one, and reference it in chat with [[IMG-1042]]
Full-size image preview modal
Prompt history for each generated image
Bulk operations: archive, delete, tag
Image metadata and creation timestamps
Gallery view with grid layout
View all
🔒
Security
31 controls
AES-256-GCM encryption at rest
OAuth 2.0 + PKCE authentication
Per-company data isolation
Example: Each company's contacts, memories, and credentials are fully isolated — switching companies shows only that company's data
Cost circuit breaker to prevent runaway spend
Helmet security headers
CORS origin whitelisting
Express rate limiting on sensitive endpoints
Client-side encrypted credential vault
Per-user data directory isolation
View all
📦
Import / Export
Portable workflows
Portable mw_schema 2.0 JSON format
Export workflows with triggers and schedules
Import with selective checkbox UI
Example: Export your "Missed Call → SMS" workflow and share it with another company
Version-stamped schema for forward compatibility
Selective import: choose which workflows to bring in
Preserves trigger and schedule associations
Cross-company workflow sharing
View all
👁
Vision Browser
Autonomous web browsing
Autonomous browsing via Puppeteer / Browserless
Form filling, login, data extraction
Screenshot capture and multi-step tasks
Example: Vision Mini logs into your competitor's pricing page, extracts all plan details into a spreadsheet
Headless Chrome with stealth mode
Cookie and session management
JavaScript execution on page
PDF generation from web pages
Scheduled recurring scraping tasks
View all
✍
Mini Sign
E-signatures
Electronic document signing
PDF generation and signature tracking
Follow-up reminders
Example: Client Mini sends an NDA for e-signature, tracks when it's signed, and notifies you
Multi-party signing support
Signature status dashboard
Automated reminder emails
Signed document archival
View all
🤝
Microsoft Teams Bot
Enterprise integration
Direct agent chat in Teams
Adaptive cards for approvals
Proactive notifications
Example: Message @MiniBot in Teams to ask Analyst Mini for today's sales numbers
Azure Bot Service integration
Approval workflows within Teams cards
Channel and direct message support
Rich message formatting
View all
🏢
Multi-Company
Workspace management
Multiple company workspaces
Per-company data isolation
Company context and training
Example: Manage Agency A and Agency B from the same account — each has separate contacts, workflows, and tool configs
Emergency stop-all button
Per-company tool assignments
Separate chat histories per company per agent
Company-specific knowledge base training
Quick-switch between companies
View all
✦
Avatar (AI Self)
14 endpoints · 9 server files
Persistent AI self above all companies
Cross-company monitoring every 5 min
Morning briefings (chat/email/phone)
Example: Ask your Avatar "Find me a business idea under $200/month" — it runs a multi-step Claude research pipeline with unit economics
Dual-model routing: Claude for deep reasoning, GPT-4o for speed
Data-grounded responses — fetches real-time data before every factual answer
Opportunity research with unit economics modelling
Draft company builder (Minis, workflows, tools, knowledge)
Autonomous actions within configured boundaries
Self-correction detection — tracks and fixes factual errors
Cost registry for honest monthly estimates
Configurable autonomy boundaries, quiet hours, briefing schedule
Voice input via Web Speech API
Personalised: "{FirstName}'s Avatar"
View all
📊
Eval System
LLM-as-Judge · 8 dimensions
Scores every AI interaction on 8 dimensions
Failure categorization (7 types)
Admin dashboard with trace explorer
Example: Click any trace row to see full input/output, per-dimension scores with reasoning, and improvement suggestions
8 scoring dimensions: task completion, accuracy, tone, safety, tool use, efficiency, context use, helpfulness
Failure categories: none, RAI refusal, safety block, tool error, hallucination, context gap, persona drift
RAI refusals shown as "Expected Behavior" (green) — not real failures
Company filter + "My Companies Only" toggle
Clickable trace rows with inline detail panel
Re-run eval button for updated scoring
Failure category breakdown chart in overview
Global trace search (admin-only)
View all