Kavi coordinates Codex and Claude as specialized agents with intelligent routing, DAG-based planning, structured code reviews, and approval gates — all from a keyboard-driven TUI.
Built for control, designed for speed
Every feature is designed to keep the operator in the loop without slowing down the agents.
Intelligent Routing
Route tasks to the right agent using path ownership rules, keyword heuristics, or AI-powered classification. Every decision includes a confidence score.
DAG Planning
Complex prompts are automatically decomposed into dependency-aware task graphs. Agents work in parallel across the execution plan.
Dual Worktrees
Each agent operates in an isolated git worktree. No merge conflicts during work — integration happens only when you're ready to land.
Structured Reviews
Attach review notes to specific diffs with dispositions like approve, concern, or question. Convert concerns into follow-up tasks automatically.
Approval Gateway
Control what each agent can do with per-tool approval rules. Remember decisions, or toggle full-access mode for high-confidence sessions.
Smart Recommendations
Kavi detects when agents should hand off work, when changes need integration, or when ownership rules should be updated.
Get the best of both worlds
When Claude Code and Codex get better, Kavi automatically gets better. No wrapper tax. No translation layer. Just the real thing.
Why Kavi
Every other tool builds their own agent from scratch. We orchestrate the ones that are already best-in-class.
Direct orchestration of first-party agents
- Runs the real Claude Code CLI and the real Codex CLI — the actual agents built by Anthropic and OpenAI.
- Zero translation layer. When Claude Code gets a new capability overnight, you have it in the morning.
- Two best-in-class agents, each doing what they're best at, coordinated by an operator who stays in control.
Not a wrapper. Not a fork. The actual tools, orchestrated. That's the difference.
How it works
From goal to landed code in four steps.
Initialize
Bootstrap a managed session with isolated worktrees.
Route & Plan
Kavi decomposes goals into a dependency graph and routes tasks to the best agent.
Review & Approve
Review diffs, add dispositions, approve tool calls — stay in control at every gate.
Land
Merge validated work through an integration worktree with optional test validation.
Orchestration architecture
Kavi runs a background daemon that communicates with the TUI over a local RPC socket. Each agent operates in its own git worktree, ensuring complete isolation during execution.
The daemon schedules tasks respecting dependency graphs, manages approval gates, tracks path claims, and generates actionable recommendations — all while streaming real-time updates to the operator console.
Session state is persisted locally under .kavi/ with JSONL event logs, task artifacts, and landing reports.
9 TUI Views
Activity, Results, Tasks, Recommendations, Approvals, Claims, Decisions, Messages, and Worktrees.
Peer Messaging
Agents communicate with structured intents: handoff, blocked, review request, and context sharing.
Full Audit Trail
Every routing decision, approval, and review is persisted in JSONL events with optional SQLite mirroring.
36+ CLI Commands
Full session control from any terminal — inspect tasks, claims, reviews, decisions, and land work.
Path Claims
Automatic ownership tracking with overlap hotspot detection and conflict resolution.
Smart Landing
Merge agent worktrees through an integration branch with validation commands and landing reports.
Start orchestrating
Install Kavi globally and open your first managed session in under a minute.
npm install -g @mandipadk7/kavi