The governance plane for coding agents.

Generation is no longer the bottleneck. Authority is.

Coding agents can already write and change production software. The open question in serious environments is control: who approved which action, on what scope, with what blast radius — and can you prove it afterwards. Editors optimize for speed of generation. Keviq Code is built for accountability of action.

Agents propose. The workspace decides.

1. 01

   Propose

   The agent requests an action — run a command, edit files, touch git. It never acts directly.

2. 02

   Authorize

   Each request becomes a first-class approval with action, risk level and scope. Routine steps are one click; destructive ones require typing the exact command.

3. 03

   Execute in isolation

   Approved work runs in an isolated worktree. Your main branch cannot change as a side effect.

4. 04

   Review, test, promote

   Inspect the real diff, run tests, then promote through an audited, revertable merge.

Running today — verified on the real app

Every item below is exercised end-to-end on the desktop app and covered by the test suite (279 backend, 207 frontend). No mockups.

• Workspace-first binding

  Terminal, git and files agree on one project, and state survives restarts.

• Live agent sessions

  Claude Code runs headless with real-time lifecycle status — no polling.

• First-class approvals

  Approve-once semantics with expiry, replay protection and consume-exactly-once, enforced at write time.

• Typed confirmation for critical ops

  Force resets and other destructive commands require typing the exact command; a structural deny-list refuses the worst outright.

• Isolated worktrees, audited promotion

  Agent edits land isolated; promotion to main is a two-parent merge you can revert.

• Context egress preview

  Before anything is sent to a model you see what is included, blocked and masked — secrets are redacted, and the audit log stores categories, never content.