Where Stallari is going

Swift harness, vault engine, dispatch engine, provider abstraction, trace store.

Agent memory with importance decay, recall, and consolidation. Post-dispatch memory pipeline.

macOS app — published, notarised, Developer ID signed. Setup wizard, fleet sidebar, digest views, plugin management, and privacy controls.

• ✓ Setup wizard with 9-step onboarding and vault detection
• ✓ Marketplace browser with service, tier, and category filters
• ✓ Privacy exclusion editors and content redaction
• ✓ Secure Enclave credential store for providers and tools
• ✓ Self-upgrade via GitHub Releases

Run Stallari across macOS devices in a fleet. Manage agent traces, jobs, and schedules from anywhere.

• ✓ Discover peers automatically via Tailscale or Bonjour (mDNS)
• ✓ Enroll devices with a QR scan — per-peer tokens, individually revocable
• ✓ Browse agent traces and coordinate dispatch across your fleet
• ✓ Single-writer dispatch lease prevents duplicate runs

Plugin marketplace, Packs, commerce, identity, and developer preview documentation.

• ✓ Plugin marketplace with conformance tiers — tools (MCPs) and packs (bundled agents, skills, and manifests)
• ✓ 20+ open-source Blade MCPs (Home Assistant, Shopify, Xero, Paddle, Office 365, Gmail, Xbox Live, Todoist, Mastodon, iMessage, Cloudflare, and more)
• ✓ 4 open-source community MCPs (Blender 3D, YouTube, and more)
• ✓ Multi-contract plugins — a single MCP can declare conformance to multiple service contracts
• ✓ Organisation packs — teams share private packs scoped to their org
• ✓ Setup wizard — 9-step onboarding with vault detection, provider config, and plugin install
• ✓ Developer ID signed and notarised — installs without Gatekeeper warnings
• ✓ Binary distribution and in-app self-updates
• ✓ Marketplace sign-in with GitHub and Apple
• ✓ Paddle checkout and licence management — tips and paid community packs
• ✓ Guide browser — in-app documentation and walkthroughs
• ✓ Built-in identity provider — OIDC authentication, seat-based licensing, and role-based dispatch gating
• ✓ Opt-in crash reporting and process guardian — open-source SDK with PII scrubbing and circuit breakers
• ✓ Platform observability — stale binary detection, MCP health probes, subprocess orphan cleanup, and integration health heartbeats

Daemon-routed MCP over HTTP with Tailscale HTTPS and ACME certificates, bearer token auth, iCloud KV dispatch lease, and remote traces.

• ✓ HTTPS with automatic ACME certificates (Tailscale Serve)
• ✓ MCP routed through daemon HTTP with HMAC-SHA256 request signing
• ✓ LM Studio — local LLM access over Tailscale HTTPS
• ✓ Remote trace persistence with SQLite dedup
• ✓ Cloudflare Tunnel and ACME certificate support — expose services without Tailscale (accommodate certain corporate policies and iOS single-VPN limitation)
• ✓ Networking settings — unified UI for managing exposed services and providers

Native agent workflows with multi-provider graph execution, five specialised methodologies, adaptive routing, and Pack-driven orchestration.

• ✓ Multi-provider support — run workflows locally with LM Studio, on Anthropic, xAI, OpenAI-compatible, or Apple on-device models
• ✓ 5 graph methodologies — map-reduce, critique-refine, plan-execute, gate, and consensus — with 102 skills across 12 packs
• ✓ Adaptive routing — trace-informed confidence thresholds with few-shot exemplar injection
• ✓ Provider probe caching — cold start from 12 seconds to under 3 seconds
• ✓ Memory discipline — resource limits, health monitoring, message trimming, and subprocess recycling
• ✓ Pack-driven orchestration — define agent workflows as portable YAML manifests
• ✓ Apple Foundation Models for on-device pre-classification (no API key needed)
• ✓ Webhook-triggered dispatch for inbound events (GitHub, Xero, Home Assistant) with auto-wiring
• ✓ Per-node model routing — assign different providers and tool allowlists to individual graph nodes
• ✓ Privacy enforcement — content exclusions, domain-scoped tool filtering, and per-tool allowlists
• → iOS companion app
• → Full native dispatch without CLI subprocess fallback
• → Streaming tool-use in embedded backends
• → MLX Swift — native local model execution without external servers; Nemotron 3 (Nvidia), Qwen 3.5 (Alibaba), Phi-4 (Microsoft), Llama 3.1 (Meta)
• → ISO 42001 compliance posture — alignment with EU AI Act standards

Your intelligence is yours. Full-fidelity export of every memory, association, and learning your agents have built — in open formats, with documented schemas, importable anywhere.

• → One-command memory export — Markdown, JSON, or SQLite, with full association graph
• → Documented memory schema — published spec so other tools can read your memories
• → Memory import — bring learnings from other systems into Stallari
• → Association graph visualisation — explore what your agents have learned and how concepts connect
• → Memory portability guarantee — if you leave, your intelligence leaves with you

Engage with other users' Stallari clusters. Shared private inferencing and coordination between independent deployments.

• → Multi-vault support
• → Shared private inferencing with LM Link — share idle local GPU capacity to friends and family
• → Inter-user coordination protocols