Intelligence

AI Engine

The Engines page shows every AI engine detected on your installation. CorvinOS is engine-agnostic — it can run conversations and sub-tasks across five different engines, each with different capabilities, cost models, and data-locality guarantees.

Detected engines

The Detected Engines section shows all engines found on the current host. A badge in the top-right corner indicates the total number of ready engines (e.g. 4 ready). Each card shows real-time status, version, and authentication method:

Engine architecture

CorvinOS distinguishes between OS engines (which can host the main conversation turn and delegate sub-tasks) and worker engines (which receive delegated tasks from an OS engine). The five engines and their roles:

Engine	Role	Locality	Cost model
Claude Code	OS → Worker	Cloud	Per token
OpenCode	OS → Worker	Cloud / On-premise	Per token / free (Ollama)
Hermes	OS → Worker	On-premise (CONFIDENTIAL-capable)	Zero (after model download)
Codex CLI	Worker only	Cloud	Per token
GitHub Copilot	Worker only	GitHub	Zero (Copilot Business/Enterprise)

Switching engines

The active engine for a conversation can be set per-chat from the engine selector in the Chat input bar. For a tenant-level default, use System → Engine Control or the REST API at PUT /v1/console/settings/engine. Individual personas can also pin a specific engine — see the Personas section below.

Personas

Personas define how the AI behaves in a specific context — which tools it can access, which engine it runs on, what its system prompt says, and what permissions it operates under. Every chat is driven by exactly one persona.

Bundle personas

Bundle personas ship with CorvinOS and are maintained by the operator. They cover the most common use cases out of the box:

• Assistant — generalist persona; the default fallback when no other persona is matched by the auto-router.
• Coder — optimized for software engineering tasks; mirrors the legacy voice plugin behavior.
• Copilot Worker — wraps GitHub Copilot CLI as a worker engine for shell, git, and gh tasks.
• Forge — runtime tool and skill generation; can create and register new MCP tools on the fly.
• Homeassistant — smart-home automation via the Home Assistant MCP server.
• Inbox — Gmail triage; reads, labels, and drafts email via MCP.
• Orchestrator — OS-mode Layer-29 delegation; spawns and manages worker engines for complex multi-step tasks.
• Os — system administrator persona; broad filesystem and shell access.
• Research — web search and Playwright-driven browsing for research tasks.

Persona cards

Each persona card shows:

• Name and description
• Capability badges — e.g. Forge (can create tools), SkillForge (can create skills), bypassPermissions (sandbox boundary enforced by Layer 10 path-gate rather than permission prompts), MCP server count

Persona detail

Clicking a persona card opens the detail view, which shows all configurable fields:

• Engine pinning — lock the persona to a specific engine (e.g. always use Hermes). Defined per ADR-0123.
• Model selection — override the default model for this persona (e.g. use a specific Ollama model for Hermes).
• Permission mode — controls whether the AI can execute tools without confirmation. Bundle personas with bypassPermissions rely on the Layer 10 path-gate as the sandbox boundary.
• MCP servers — which MCP tool servers are available to this persona (e.g. imagegen, forge, skill-forge).
• System prompt additions — persona-specific instructions appended to the base system prompt.
• LDD preset — Loss-Driven Development layer configuration for this persona.

User personas

You can create user-scoped copies of any bundle persona. User personas are stored separately and can be edited freely without modifying the system default. They appear in the same card grid, distinguished by a "User" label.

Auto-routing

When no persona is pinned for a chat, CorvinOS uses the auto-router to select the best persona based on the message content. The default routing mode is heuristic — keyword pattern matching with zero latency, no API call. Low-confidence routing falls back to the assistant persona. You can override auto-routing by pinning a persona to a chat profile in the bridge settings.json.

Files

The Files page provides a browser-based interface to the tenant's file storage. Files stored here are accessible to AI engines during tasks — they can be referenced, read, and written as part of any workflow.

File operations

The file browser supports the following operations:

Storage location

Files are stored under the tenant's workspace at ~/.corvin/tenants/<tenant_id>/. Session artifacts (PDFs, images, CSVs, exports generated during a task) have a dedicated storage path at sessions/<bridge>:<chat>/artifacts/ and are managed by the Layer 33 artifact memory system — they are registered in an append-only manifest with SHA-256 integrity hashes and can be searched and retrieved by the AI in future sessions.

AI access to files

The AI can read and write files in the tenant workspace during tasks, subject to the Layer 10 path-gate. The path-gate blocks writes to protected directories (audit logs, policy files, forge workspaces, license trees) regardless of which engine is running. Files you upload here are immediately available to be referenced in any conversation or task.