Agents

Autonomous coding on your machine — with brakes.

ORQVION agents plan and execute multi-step work in a Rust ReAct loop on your hardware. They read, search, shell, and edit with the same ONLINE · OFFLINE · AUTO routing as chat — cloud inference only when your mode and keys send it there (BYOK or configured online path). Writes need approval, Docker can be mandatory, and audit events never carry secrets.

Modes

Pick how much autonomy the agent gets.

Same engine, different guardrails. Org RBAC can lock Viewers to Ask only. Full Agent mode is on Pro+; Free includes Ask with limited online entitlements.

Ask

Read-only answers about your codebase. No file writes, no shell. Safe for viewers and quick questions.

Plan

Produces a step-by-step plan you approve before any edit lands. Good for refactors and unfamiliar repos.

Debug

Leans on logs, test output, and DAP sessions — not guess-and-patch. Tied to the integrated debugger.

Agent

Full tool loop: read, write, search, shell, Artisan, and MCP. Checkpoints recover long tasks; you still approve writes. Pro+ entitlement.

Control

Nothing lands in your tree without a decision.

Diff review, write approval, and optional container isolation are defaults — not enterprise upsells.

Unified diff review

Per-hunk accept or reject for agent edits. Pending changes stay in a Rust diff model until you accept.

Write approval

File writes pause for explicit approval. Hooks can block PreToolUse events and the block is audited.

Docker sandbox

Agent shell and file tools can run inside a container. Org policy can require it for every developer.

Checkpoints

Long agent runs save recoverable checkpoints in local SQLite — survive restarts without losing context.

Scale-out

Subagents and architect workflows for bigger builds.

Parallel child agents with isolated context — concurrency caps keep the desktop responsive whether routing is online or offline.

Subagents

Research, Shell, and Review children with separate context windows. Custom agents live in `.orqvion/agents/*.md`.

Architect path

Architect → Developer → QA mapping for longer features, with clarification questions before work starts.

MCP & hooks

Model Context Protocol tools and `.orqvion/hooks.json` gate what the agent can call. Policy can disable tools org-wide.

Install the desktop app and try Agent mode.

Start with Ask on a real repo. Upgrade to Pro for full Agent mode, or enable offline models under Settings → Advanced Local.