Architecture overview

Carina follows a gateway + agent + tools architecture with multi-tenant SaaS and Scout security as differentiators.

flowchart LR
  subgraph channels [Gateways - 19 adapters]
    CLI[CLI]
    Web[Web UI]
    TG[Telegram]
    DC[Discord]
    More[Slack WhatsApp SMS ...]
  end
  subgraph core [Core]
    Router[Gateway router]
    Orch[Orchestrator]
    Agent[CarinaAgent.turn]
    LLM[LLM providers]
    Mem[Memory backends]
  end
  subgraph exec [Execution]
    Tools[Tool registry]
    Skills[Skill packs]
    Scout[Scout client]
  end
  channels --> Router --> Orch --> Agent
  Agent --> LLM
  Agent --> Tools
  Agent --> Mem
  Tools --> Scout

Main modules (core.carinaai.uk/src/)

Module	Role
agents/	CarinaAgent, orchestrator, sub-agents, cron
brain/	LLM providers, credential pools, prompt cache
gateway/	Channel adapters, web UI, router
tools/	Tool registry and implementations
memory/	Episodic store, RAG, pluggable backends
skills/	Packs, curator, auto-writer
security/	Prompt guard, Scout reporting, JIT credentials
cli/	49 CLI commands
commitments/	Inferred follow-up memories (opt-in)
workspaces/	Named agent workspaces and channel routing
ops/	Sandbox, DNS, proxy, backup, secrets reload, audit
infrastructure/	Host service, device pairing, control auth

Request flow

1. Inbound message arrives on a channel adapter
2. GatewayRouter maps to a session and CarinaOrchestrator
3. Context is built: memory search, RAG, skills, system prompt
4. LLM may return tool calls; tool-runner executes with Scout policy
5. Response is formatted per platform and sent outbound

Deployment modes

• CLI / self-hosted: ~/.carina/ config, local Postgres + Redis
• Carina Cloud: multi-tenant Postgres schemas, Stripe billing, hosted web app
• Docker: docker-compose.yml for Postgres, Redis, and the agent container