THE BUILD LAYER

Your workshop for custom intelligence.

E.V.I. Canvas, Skills, Knowledge Node, Workflows, Hooks, Tool registry. Six primitives that compose into AI Workers you can actually call your own. Built on your knowledge, governed by your rules.

01 · E.V.I. CANVAS

Open the canvas → /dashboard/skills/canvas

Build AI Workers visually.

The E.V.I. canvas is the workshop. Drag persona, skill, knowledge, and workflow nodes onto an infinite grid. Wire them with conditionals, vision nodes, and tool calls. Compile, test, version, publish.

  • Drag-drop nodes for persona, model, tool, knowledge, conditional, and vision
  • Live compile + test against sample prompts
  • Versioned drafts, every save is reversible
  • Publish privately, to your org, or to the public marketplace

02 · SKILLS

Reusable bundles. Marketplace ready.

Skills are the unit of reusable intelligence. A skill bundles a prompt, the right tools, the right knowledge, and an optional workflow. Author it once, install it across every AI Worker, version it like code.

  • Three publish modes: private, org-only, public marketplace
  • Per-API-key skill allowlists (api_key_skills binding)
  • Versioned manifests with rollback
  • Discover, install, fork from the public marketplace

03 · KNOWLEDGE NODE

Zero training, ever → /privacy

Your training data, attached.

Drop PDFs, DOCX, XLSX, CSV, JSON, and Markdown directly onto a skill via the Knowledge Node. Priority-weighted, token-budgeted, injected at runtime. Stored in org-scoped object storage. Never used to train any model.

  • 10MB per file · 100 files per skill
  • 50K-token runtime budget per request
  • Priority weighting 1 to 5 (P1 gets 3× the budget share)
  • Org-scoped storage. Cross-tenant retrieval is structurally impossible.

04 · WORKFLOWS

Chain steps. Branch. Recover.

Workflows are the orchestration layer for autonomous multi-step jobs. Six step types, N-ary branching on conditions, configurable retry. Credit checkpoints at every step. 10-step ceiling, 30-min total runtime cap.

  • 6 step types: tool_call, ai_transform, condition, delay, notification, webhook
  • N-ary branching, with conditions routing to labeled targets
  • Credit checkpoint before every step; pause and resume safely
  • Webhook step fires HMAC-signed POSTs with retry + DLQ

05 · HOOKS

Autonomous triggers on domain events.

Hooks fire skills automatically when something happens. Five event patterns out of the box, plus custom patterns. Cooldown windows, daily-cap checks, and autonomous execution all built in.

  • share.accepted.welcome: personalized welcome on share-accept
  • task.overdue.summary: overdue prioritization summary
  • project.created.suggestions: next-step recommendations
  • team.member_joined.onboarding: onboarding drafts on join
  • document.analyzed.followup: action items after document analysis
  • Custom event patterns and per-hook cooldowns are also supported

06 · TOOL REGISTRY

Built-in + skill + connector. Permission gated.

Every AI Worker has access to a permission-gated tool registry. Built-ins for memory, code, research, browser, documents. Skill-declared tools for domain actions. Connector-declared tools for external systems. Every call audited.

  • Four permission tiers: user, org_member, org_admin, system
  • Tool execution audited with arguments + output
  • Skills declare tools in their manifest; connectors via capability descriptors
  • Resolved per-request based on caller scope + skill bindings
ZERO TRAINING, EVER

Your workshop. Your data. Not our training set.

Every prompt, knowledge file, skill, and workflow you build stays yours. We persist them so your AI Workers can run, not to improve our models. The memory graph is exportable on demand and deletable on request. The SDK is open source if you want to read the code.

Read the privacy stack

Open the canvas. Ship a Worker.

Get your API key in 30 seconds. Drag your first persona node, attach a skill, drop in a Knowledge file, hit publish.