DOCUMENTATION

Understanding Spec4AI Studio

Conceptual overview of the Spec4AI Studio architecture, components, and trust model. Full API docs coming at launch.

Three-layer architecture

Layer 3Applets

Spec-driven applets: JSON specs, assets, capability manifests, tests

Layer 2Tool surface

40+ tools across 8 namespaces: app, ui, hardware, storage, package, model, debug, control

Layer 1Runtime

Flutter container, MCP server, permission engine, lease manager, receipt log

Core concepts

Applet

A self-contained app: JSON spec + assets + capability manifest + optional tests. Applets run in the Spec4AI Studio runtime with explicit, scoped permissions.

Capability Manifest

Every applet declares what it needs: network access, camera, storage, etc. The manifest is reviewed before installation via a permission card.

Permission Card

A human-readable summary of what an applet can do. Presented before installation. No hidden permissions — what you see is what it gets.

Lease

Time-limited access tokens for sensitive capabilities. Leases expire automatically. No permanent grants to sensitive hardware or data.

Receipt

An auditable record of every tool call an applet makes. Receipts include what was accessed, when, the result, and the requesting applet.

Channel

A distribution feed from a registered publisher. Subscribe to get applet updates. Like RSS for applets.

Evidence Bundle

Generated at deploy time: the spec, screenshots, test results, and capability manifest. Evidence is how trust is established.

Builder

The desktop component that generates, tests, and deploys applets. Runs locally (Claude Desktop or any MCP client). Never required at runtime.

API SURFACE

Tool namespaces

Spec4AI Studio exposes 40+ tools organized into 8 namespaces.

app.*App lifecycle, snapshots, mode switching
ui.*Deploy specs, set state, capture screenshots, run tests
hardware.*Location, camera, clipboard, network, notifications, permissions
storage.*Key-value, file system, encrypted database
package.*Install, update, rollback, remove, launch applets
model.*On-device LLM inference (Gemma 3n)
debug.*Log summaries, LLM status
control.*Ping, events, chat, context, task tracking

SECURITY

Security model

  • Sandboxed runtime — applets can only access declared capabilities
  • Time-limited leases for sensitive access
  • Receipt log for full auditability
  • Publisher registration with public key cryptography
  • Runner Safe Mode blocks unregistered publishers
  • No telemetry by default — reporting is opt-in and minimal

Want to dive deeper?

Full API documentation and integration guides will be available at launch.