I got tired of AI chatbots that can only talk. So I built Spectrion — an autonomous AI agent for iPhone that actually executes tasks: sends messages, manages calendar, searches the web, creates tools, and chains it all together without you lifting a finger.
Not Just Chat. A Real Agent.
Most "AI apps" are glorified ChatGPT wrappers. Spectrion runs an agent loop — the LLM calls tools, gets results, and keeps going until the task is done:
User: "Find a pizzeria rated 4.5+ nearby and add it to my calendar"
Agent:
→ web_search("best pizzerias near me rated 4.5+")
→ Found 3 places
→ calendar("add dinner at Luigi's, 7 PM tonight")
→ Done!
One message. Multiple tools. Zero hand-holding.
57+ Built-in Tools
Eight categories covering everything your phone can do:
- Search & Web — web_search, web_fetch, URL tools
- Communication — iMessage, SMS, calls, email, contacts
- Organization — Calendar, reminders, scheduled tasks, cron
- Files — Filesystem, cloud, XLSX/DOCX/CSV/PDF parsing
- Media — Camera, vision (OCR), image generation & editing, audio
- System — Device info, brightness, location, maps, health, shortcuts
- AI & Meta — Runtime tool creation, skills, memory, sub-agents, UI rendering
- Extensions — MCP servers, plugins, community tools
Real Examples
Multi-tool task execution
"Remind me about the team standup tomorrow at 10am and add it to my calendar"
The agent calls reminders and calendar simultaneously, sets up both, and confirms — all in one turn.
Web search & summarization
"Search the web for the latest AI agent news"
Calls web_search, fetches results, and returns a structured summary.
Runtime tool creation
"Create a tool that converts temperatures between Celsius and Fahrenheit"
The agent writes JavaScript, tests it, and registers it as a new tool — usable immediately. Tools get versioning, rollback, and isolated storage.
Workflows — Visual Automation
Chain tools into multi-step workflows with:
- Triggers — manual, scheduled (cron), event-driven
- Actions — HTTP requests, LLM calls, notifications
- Logic — Conditional branching, loops, parallel execution
Build them visually or let the agent generate them from a description.
Extensions: Store, Skills, Plugins, MCP
- Store — Browse and install community tools & skills
- Skills — Reusable instruction sets (web_researcher, scheduler, etc.)
- Plugins — Hot-reload packages (Smart Summarizer, Code Sandbox, etc.)
- MCP — Model Context Protocol servers for external integrations
- Custom Tools — JS sandbox with HTTP, KV storage, SQLite, device APIs
Autonomous Agent Features
Heartbeat
The agent wakes up periodically (configurable interval) to check pending tasks, process messages, and run maintenance — even when you're not looking.
Morning Briefing
Daily briefing with weather, calendar events, news headlines — customizable topics and time.
Task Check-In
Auto-resumes unfinished tasks. Configurable active hours (default 8am–11pm).
Chat Watchdog
Auto-nudges the agent if it stops mid-task.
Evolution Engine — Self-Improvement
Every 24 hours, the agent:
- Analyzes tool usage patterns and success rates
- Refines its system prompt and persona parameters
- Auto-creates tools for repetitive tasks
- All changes versioned with instant rollback
Semantic Memory
Long-term memory with vector search. The agent decides what to remember — conversations get indexed automatically. Semantic search across all stored knowledge.
Device Mesh — Multi-Device Agent
Connect iPhone + Mac into a single agent:
- Sync conversations, tools, settings, memory
- Execute tools cross-device (Mac agent can trigger iPhone camera)
- End-to-end encrypted (Curve25519 ECDH + AES-256-GCM)
- Offline queuing with conflict resolution
Channels — Telegram, Discord, Slack, WhatsApp, Email
Connect external messaging platforms. The agent can receive and respond through Telegram bots, Discord, Slack — with full tool access.
Deep Personalization
Configure the agent's personality:
- Role — Assistant, Coder, Researcher, Writer
- Style — Friendly, Professional, Casual
- Personality Level — Pure LLM, Human (natural), Realistic (full character)
- User context — Name, address style
Choose Your AI
Works with multiple providers:
- Spectrion Pro — All-in-one, 3-day free trial, no API key needed
- Apple On-Device — Private, free, works offline
- Anthropic (Claude) — Direct API
- OpenAI (GPT-4o) — Direct API
- Ollama — Local models, fully offline
- Custom — Any OpenAI-compatible endpoint
Auto-fallback between providers. Utilization-aware load balancing.
Tech Stack
- Frontend: SwiftUI, @observable, async/await, SwiftData — zero third-party dependencies
- Concurrency: Actor-based tool executor, TaskGroup parallel execution
- Backend: Node.js, SQLite (WAL), Redis, account pooling, tier routing
- Security: Keychain storage, E2E encrypted mesh, no server-side conversation logging
- Localization: 11 languages with automatic tool activation by keywords
Links
Built as a solo dev. If you have questions about the architecture, agent loop, tool system, or anything else — ask away in the comments.
















Top comments (0)