Kanna
Run agents like documents, not log files.
A web UI for the Claude Code and Codex CLIs, built for long sessions across many projects. OAuth pool rotation, kanban-driven worktrees, subagent orchestration, durable approvals. No dashboard chrome.
$ bun install -g @cuongtran001/kanna
Your plan, across as many tokens as you have.
Register one Claude OAuth token, or several. Kanna picks the least-loaded one per chat and falls over automatically when a token hits a rate limit, so a long session does not stop at the first limit you reach.
Codex and OpenRouter sit beside Claude in the same composer. Switch provider mid-chat; previous turns stay where they are.
Subagents that report back.
Configure named subagents with their own system prompts and tool
surfaces. The main agent decides when to delegate, calls
mcp__kanna__delegate_subagent, and synthesizes the reply back into
its own turn. Cycle detection and depth limits run on every
delegation. Live progress bubbles up to the transcript without a
separate window.
A transcript you can audit.
Tool calls render hydrated with collapsible groups. File diffs show
inline. Plan-mode and AskUserQuestion prompts route through a
durable approval protocol, so pending requests survive server
restarts and replay on reconnect. Context compaction happens
proactively, before the window closes around you.
Projects up front, chats below.
Chats live under projects with live status (idle, running, waiting, failed). Kanna auto-discovers projects from your local Claude and Codex history, imports existing sessions with one click, and resumes them with the full transcript intact.
Each project also gets a kanban board. Starting work on a card cuts a branch, a git worktree and a chat in one gesture — which is what makes three agents working at once safe from each other.
Kanna is a web UI for the Claude Code and Codex CLIs, with OAuth pool rotation, kanban-driven worktrees, durable approvals, and subagent orchestration. macOS and Linux. Self-hosted.