yume

claude code, supercharged orchestration flow. live interleaved thinking. background agents. plugins & skills.4 built-in agents. vim mode. 4 providers. 7 languages. mac, windows, linux.

Download

---

why yume?

Claude Code is transformative for development. Terminal interfaces hold it back.

terminal limitations

• flickering — terminal redraw artifacts, a long-standing cli complaint
• input lag — terminal rendering slows down in extended sessions
• single session per window — parallel work means juggling terminals
• rate limits on demand only — requires manual /usage checks
• no gui affordances — no visual diffs, clickable rollback, or drag-and-drop

yume's approach

• ✅ native rendering - hardware-accelerated UI, zero terminal artifacts
• ✅ responsive input - typically <50ms, independent of session length
• ✅ standalone app - your session doesn't die with your terminal or IDE
• ✅ persistent rate limit display - 5-hour and 7-day metrics always visible
• ✅ automatic session recovery - 5-minute snapshots with 24-hour retention

---

how it works

yume is not an alternate harness or api wrapper. we spawn the official claude code cli as a subprocess.

yume app → spawns real cli binary → cli handles auth → anthropic servers

this is fundamentally different from tools that spoof headers to bypass api pricing. yume:

• runs the actual claude binary installed on your machine
• uses your existing claude pro/max subscription
• never touches anthropic's api directly
• maintains 100% cli compatibility (subagents, mcp, hooks, skills, claude.md all work)

think of it like iterm vs terminal.app - a different frontend for the same tool.

---

features

orchestration flow (unique to yume)

• automatic task decomposition - understand → decompose → act → verify workflow
• complex task handling - 3+ step tasks get architect planning, implementer execution, guardian verification
• trivial task passthrough - simple tasks proceed directly without overhead
• no commands needed - baked into every session via --append-system-prompt

agent system

• 4 built-in agents - synced to ~/.claude/agents/yume-*.md
  • architect - plans, decomposes tasks, identifies risks, todowrite tracking
  • explorer - read-only codebase search, glob/grep/read, never modifies
  • implementer - focused code changes, edit/write, minimal diffs
  • guardian - reviews for bugs, security, performance after changes
• 4 concurrent background agents - async execution with git branch isolation (yume-async-{type}-{id})
• 30-minute timeout per agent - output to ~/.yume/agent-output/, 24h retention
• streaming isolation - background agents don't interfere with main session
• merge conflict detection - pre-merge checks before integrating agent work
• worktree isolation - full filesystem isolation via git worktree
• agent coordination - inter-agent messaging to hand off work between background agents