Verloray AI Command Center

A private, production-structured multi-model orchestration command center. Phase 1 is a working local vertical slice: dimensional provider selection, Direct and concurrent Council modes, normalized SSE streaming, server-only OpenAI/Anthropic adapters, an explicit simulation adapter, and persistent project conversations.

Launch

Requirements: Node 20+ and pnpm.

Copy-Item .env.example .env.local
pnpm install
pnpm dev

Open http://localhost:3000. The simulation provider works without credentials. To enable a real provider, add OPENAI_API_KEY or ANTHROPIC_API_KEY to .env.local, restart the server, open Integrations, and select that connected provider. Never put keys in browser code or commit .env.local.

Checks

pnpm typecheck
pnpm lint
pnpm test
pnpm build

Architecture

See ARCHITECTURE.md. Browser requests cross a same-origin route boundary and enter a provider registry implementing AIProviderAdapter. All streams become normalized SSE events. Council calls fan out concurrently with Promise.allSettled, isolating failures. Local MVP projects use an atomic, serialized JSON store under data/; the complete PostgreSQL target schema lives in prisma/schema.prisma.

Adding a provider

1. Add non-secret metadata to src/lib/providers/catalog.ts.
2. Implement AIProviderAdapter in a separate server-only module.
3. Read credentials only from server environment or a vault reference.
4. Register the adapter in registry.ts.
5. Normalize deltas/errors; preserve raw metadata server-side when needed.
6. Add adapter tests and disclose provider data-retention terms in Integrations.

Only official APIs, SDKs, OAuth, MCP, or authorized local brokers are acceptable integration paths.

Database and deployment

The checked-in Prisma schema is the production PostgreSQL design. Before a public deployment, install Prisma, configure DATABASE_URL, create migrations, replace src/lib/store.ts, and test backup/deletion paths. Deploy behind HTTPS with managed Postgres, a secrets vault, durable rate limiting, file malware scanning, monitoring, and an identity provider. Set AUTH_SECRET in the vault—not Git.

Security notes

Phase 1 includes server-only keys, input validation, prompt limits, local rate limiting, normalized errors, and secure headers. Authentication is structurally anticipated but not enabled, so this build must not be exposed publicly. External content must be treated as untrusted instructions. The execution broker is a safe mock; no browser-triggered shell execution exists.

Current limitations / roadmap

• Live OpenAI and Anthropic paths require user-supplied credentials and have not been claimed operational without them.
• Authentication, PostgreSQL/Prisma runtime, Redis-backed limits, file uploads, Compare evaluation, Wisdom orchestration, Guide editing, approval execution, and usage accounting are later phases.
• Codex, Claude Code, Copilot, and avatar integrations are abstractions only. Local workers must use isolated worktrees, command logs, diff review, and approval gates.
• Cost is intentionally unavailable until a dated pricing table is configured.

Next recommended step: add authentication plus the PostgreSQL repository implementation, then validate one credentialed provider end-to-end before expanding Phase 2 adapters.