Chat is the front door
Users speak normally to an LLM-backed chat. Browser, workspace, and patch work becomes governed action cards instead of hidden local execution.
Product
Local AI action control, end to end.
Dvara is not another browser or another model. It is the desktop control plane that lets any approved AI client request local capabilities through grants, policy, artifacts, and audit.
The complete loop
From LLM intent to visible local result, every step remains inspectable.
1LLM chat or MCP client
2Dvara token and grant check
3Managed browser or workspace tool
4Artifact, patch, or structured result
5Local audit and emergency stop
Capabilities
Managed browser control
Dvara opens an isolated Chrome-compatible browser profile, navigates, searches, inspects, clicks, fills, types, presses, screenshots, and records artifacts.
Read-only workspace context
Agents can list, search, and read granted workspaces with path containment, blocked secret files, redaction, size limits, and audit-safe summaries.
Human-applied patch proposals
External clients propose patches as artifacts. The desktop UI validates paths and blocked files, previews diffs, and lets the user apply selected hunks.
One permission layer
Per-client tokens, durable grants, domain policy, emergency stop, and audit events keep Claude, Codex-style tools, Qwen, DeepSeek, and MCP clients inside the same control model.
Local audit trail
Tool requests, approvals, denials, failures, artifacts, and cleanup decisions are recorded locally so users can inspect what happened after the fact.