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

multi-provider support

• 15+ models, 4 providers - non-claude providers via yume-cli shim
  • claude fable 5 (default flagship), sonnet 5 (1M context), sonnet 4.6, opus 4.8, haiku 4.5
  • gemini 3.1 pro preview, gemini 3 flash preview
  • gpt-5.5, gpt-5.4 mini
  • kiro (auto, claude sonnet 4.5, claude haiku 4.5, glm-5)
• protocol normalization - stream-json output for all providers
• same interface - switch providers without changing workflow

core features

• persistent rate limit visibility - 5h and 7d limits always visible, no /usage needed
• zero visual artifacts - native rendering eliminates terminal flickering
• interleaved thinking toggle - watch claude reason live, word-by-word, with on/off control
• keyboard-first - 50+ shortcuts, full mouse-free navigation
• <50ms response time - instant UI regardless of session length
• standalone reliability - independent process, immune to IDE crashes
• auto-compaction - dynamic threshold (default 85%), user configurable or disable
• effort + budget controls - low/medium/high/xhigh effort and thinking budget per model
• usage-limit guard - optionally stops new claude turns before the 5-hour window hits 100%, so a subscription never spills into metered overage
• schedule system - /schedule 5m, /schedule 2pm, /schedule done for timed and event-based tasks
• askuser ipc - claude asks structured questions mid-session with choice popover
• secret censoring - api keys and tokens automatically redacted in all output
• 7 languages - english, spanish, french, german, polish, chinese, japanese

plugin & skills system

• complete extensibility - commands, agents, hooks, skills, MCP servers, monitors
• 23 hook events - pre-tool-use, post-tool-use, context warnings, compaction, subagent lifecycle, and more
• skills auto-injection - context based on file extensions, keywords, regex
• plugin directory - install/enable/disable per component
• bundled yume plugin - default commands and agents

additional features

• crash recovery - 5-min snapshots, 24hr retention, complete session restore
• history rollback - restore points per message, conflict detection, undo edits
• mid-stream context - real-time token updates during streaming
• analytics dashboard - usage by project/model/date, cost tracking, export
• voice dictation - F5 to toggle, native speech-to-text
• @ mention system - @ shows recent/modified files with autocomplete, folder navigation
• vim mode - full normal/insert/visual/command modes for chat navigation
• row split panels - up to 6 panels in 3x2 grid with F7/F8 split/combine
• rules management - .claude/rules/ CRUD in settings
• image thumbnails - drag-and-drop images with compression pipeline
• slash commands - /schedule, /loop, /compact, /clear, /model, /resume, /usage, plus custom and cli comman