Capabilities
Stallari is a complete AI operations platform. Here's what it does and how it works — in depth.
Core
Agents that run without you
Stallari doesn't wait for you to open a chat window. Agents fire on schedule — your inbox is triaged at 6am, your daily brief lands before coffee. They fire on events — a file lands in your vault inbox and classification starts immediately. Or you trigger them on demand from the menu bar.
Every dispatch is logged as a trace with full tool call history, token counts, cost tracking, and classification results. When something goes wrong, you see exactly what happened and why.
feat-dispatch-mixed.png
Core
A team, not a chatbot
Stallari runs a hierarchy of specialised agents. An orchestrator routes work to domain operators — one for your knowledge base, another for email, another for scheduling, another for your home. Each agent has its own state, its own scope, and its own memory.
Agents communicate through structured handoffs, not chat relay. A writer registry prevents conflicts when multiple agents need to modify the same resource. The result is a coordinated team that divides complex work without stepping on each other.
feat-agents-handoff.png
Intelligence
Intelligence that compounds — and stays yours
Every interaction teaches your agents something. Observations, decisions, corrections, commitments, preferences — each becomes a memory stored as a Markdown file in your vault and indexed in an open SQLite database.
Memories decay over time unless rehearsed. Related memories strengthen their associations through Hebbian reinforcement. Periodic consolidation merges related memories into higher-order insights. The entire system is transparent, configurable, and exportable.
feat-memory-list.png
Intelligence
Find anything, understand context
Stallari's Lens engine combines full-text search with semantic vector search. Type a query and get results ranked by both keyword relevance and conceptual similarity. The index updates in real time as your vault changes.
Lens runs as an embedded Python runtime — no external server, no cloud API, no data leaving your machine. The semantic model runs locally on Apple Silicon.
feat-search-results.png
Infrastructure
Your machines, one system
Stallari discovers your devices automatically. On the local network, Bonjour finds machines in milliseconds. Over the internet, Tailscale provides an encrypted mesh with identity built in — no custom auth, no certificates to manage, no ports to open.
Cross-host dispatch locks ensure the same job never runs twice. A Mac mini handles overnight batch work. A laptop runs on-demand requests. They share state through your vault and coordinate leases via iCloud key-value storage.
feat-fleet-status.png
Ecosystem
Install a workflow in one click
The Stallari Marketplace is a curated catalogue of Packs — pre-built agent workflows for common tasks. Inbox triage, daily briefing, email drafting, property management, infrastructure monitoring. Install with one click and the dispatch system handles the rest.
Packs use a two-axis trust model: author type (first-party, certified partner, verified developer, community) crossed with readiness level (production, beta, experimental). You always know what you're installing and who built it.
feat-marketplace-detail.png
Infrastructure
Your machine is the model server
Every Mac with Apple Silicon is a capable inference machine. A 32GB MacBook runs Nemotron 3 30B or Llama 4 Scout comfortably. Even a 16GB iMac handles Phi 4 or Gemma 3 without breaking a sweat. Stallari runs models directly on your hardware using MLX — Apple's native machine learning framework, optimised for the unified memory architecture in every M-series chip.
Multi-axis provider routing lets you send sensitive content to local models while using cloud providers for tasks that need more capability. A certified model manifest with SHA-256 verification ensures you're running the exact model you expect. LM Studio is also supported as an alternative runtime.
feat-providers-local.png
Trust
Local-first is not a marketing term
Stallari runs entirely on your hardware. There is no Stallari cloud, no telemetry server, no account system that phones home. Your vault, your models, your dispatch traces — all local. Encryption at rest protects your vault index and memory database.
The app is code-signed with a Developer ID certificate, notarised by Apple, and distributed as a verified DMG. Sealed packs undergo security inspection before installation. Provider routing gives you granular control over which data reaches which model.
Foundation
Every artefact is a file you can read
Stallari doesn't maintain a separate database you can't see. It works on a vault — a folder of plain text files on your Mac. Dispatch traces are Markdown. Memories are Markdown. Digests, decision ledger entries, agent state — all Markdown files with structured YAML frontmatter. Open them in any text editor, any time.
This isn't just a storage choice. It means you can search your agent's work with any tool you like. Browse it in Obsidian with queries and graph view. Back it up with git. Read it on your phone. Grep it from the terminal. The system is transparent by architecture, not by policy.
feat-vault-obsidian.png
10 concurrent dispatches, free forever. No credit card required.