Skip to content
kanade — 奏 kanade docs

Feature availability

Kanade has two layers. The desk itself is a Claude Code workstation. The local LLM is an optional narrator and prose engine on top of the local ledger.

Works with Claude Code alone

Section titled “Works with Claude Code alone”

These features only need Kanade, the daemon, and your Claude Code login:

  • Agent-mode chat. Chat-view and terminal drawer are the same Claude Code session, with markdown turns, thinking blocks, and tool cards rendered live.
  • Tools, MCP, skills, shell, and file edits. Kanade launches Claude Code and keeps the normal agent loop intact.
  • Provider lanes that speak the Claude Code contract. Claude, DeepSeek, Gemini, Codex, and other CLI-backed lanes can appear as account pills.
  • Project picker, file tree, paste and drag/drop. Attachments stay local and are sent as native @<path> references.
  • Live Floor, Sessions, Catalog, and usage panels. Hook events are captured into SQLite, streamed locally, and shown without a model call.
  • Local chronicle data. Kanade records recent sessions, tool activity, file touches, skill signals, and summaries locally. Chat questions still go to the active lane until Kanade has a durable routing policy instead of brittle shortcut matching.
  • Deterministic summaries and diagnostics. Counts, paths, tool usage, skill signals, cost estimates, update-channel diagnostics, and safety prompts are computed locally.

Needs a local LLM for the voice layer

Section titled “Needs a local LLM for the voice layer”

These features still have deterministic fallback data, but the polished prose requires LM Studio, Ollama, or another OpenAI-compatible endpoint:

  • Yume / Nightly Read. The daily reading of recent activity in Kanade’s voice.
  • Optimize prose. The smith’s written appraisal around skill usefulness and retirement candidates.
  • Per-session chronicle narration. The ledger can always show counts and events; the restrained narrative paragraph is generated by the local LLM.

If the local LLM is offline, Kanade should degrade quietly: the workstation still works, Claude Code lanes still work, and local summaries remain available.

Direct mode is separate

Section titled “Direct mode is separate”

Direct mode calls a provider API from Kanade for fast plain chat and image input. It is useful for lightweight conversations, but it is not Claude Code agent mode: no tools, no MCP loop, no shell, no skill execution, and no file edits. Switch a lane back to agent mode before asking it to work in the repo.