Kucedr is built around a conversational assistant that takes actions through tools.
Assistant Chat
Assistant output streams into the conversation as GitHub-flavored Markdown: headings, lists, tables, blockquotes, inline code, and syntax-highlighted code blocks. External links open outside the app.
- Tool calls are grouped into collapsible activity summaries showing running, completed, or error state, input, output, duration, and the tool-call identifier.
- Generated images, video, and audio appear inline and play without leaving the conversation. Right-click opens a native menu to open, reveal, copy, or save a copy.
- Copy, read-aloud, and reply actions sit on each assistant message.
- Multiple files can be attached to one request. Attachment chips show filenames, sizes, and removal actions; queued files remain as session drafts. The composer does not validate count, type, size, or model support before Send, so an unsupported file may fail later.
/planswitches the session into Plan mode, while/goal <objective>creates a durable conversation goal. Task commands are expanded into agent instructions. Home does not currently mount a general slash-command menu or searchable skill picker.Cmd/Ctrl+/focuses the prompt editor. The send button becomes a stop button during a run, and stopping aborts the run and rejects any pending tool approval.
Sessions are stored locally, listed newest first, and titled from the first user message. Switching sessions restores its stored transcript; the chat view loads at most the last fifty stored messages. Settings includes a Chat History screen with per-session deletion.
Agent Tools
| Area | What it does |
|---|---|
| Files | Read a complete UTF-8 file, create or overwrite a text file, replace one exact unique match, and apply structured multi-file patches. |
| Commands | Run a shell command with working directory, environment, timeout, background behavior, and optional PTY. |
| Processes | List, poll, page through logs, write or paste text, send special keys, kill, clear, or remove retained long-running sessions. |
| Web search | Query Brave or Tavily for 1–20 results using the key from Settings. |
| Web fetch | Fetch public HTTP(S) pages or JSON, follow up to three redirects, convert HTML to text, and truncate long output. Private, loopback, and link-local targets are blocked. |
| Browser | Start or stop a persistent visible Chrome profile; manage tabs; navigate; take DOM, text, screenshot, or PDF snapshots; read console output; click, type, hover, drag, select, fill, wait, or evaluate. |
| Media | Generate an image, video, music track, or sound effect and save it in the agent workspace by default. |
| Skills | Load an enabled skill's SKILL.md and return its directory path to the run. |
| MCP | Search the saved catalog of enabled server tools and connect to a selected server when a tool executes. |
| Schedules | Create, update, pause, resume, delete, inspect, list, or trigger schedule records. |
| Health | Replace the HEALTH.md checklist or update health-run settings. |
| Bootstrap | Complete the one-time conversational bootstrap after the profile files are written. |
| Subagents | Run one independent subagent with a fresh conversation and the same tools, minus further subagent spawning. |
Subagents inherit the parent run type and filtered tool set, but cannot spawn nested subagents. Background runs are non-interactive, so an operation that requires approval is denied when no permission card can be shown.
Normal commands run in the operating-system command sandbox. Host execution with elevated: true always requires interactive approval; gateway and remote-node execution are not implemented in this runtime.
Voice
With a streaming speech-to-text model, Kucedr captures mono PCM audio and appends partial and final transcript events live. With a batch-only model it records locally, submits when you stop, and appends the returned transcript. Dictation includes microphone permission checks, elapsed time, waveform, confirm, cancel, and error states. Assistant replies can be read aloud through the configured text-to-speech provider.
With a supported OpenAI or xAI realtime model, the empty-composer voice action starts a full-duplex session with microphone capture, assistant audio playback, transcript messages, interruption, tool activity, permissions, and generated media in the normal conversation.
Image, Video, And Audio Generation
Describe an image, video clip, music track, or sound effect in chat. Kucedr uses the configured provider and model through an agent tool. A model appearing in the catalog does not guarantee that its generation adapter works.
Agent-created media is saved in the agent workspace by default and displayed inline in chat. Library manages files imported into it, with Collections and List views, upload, preview, move, download, and delete actions.
See Providers for which media adapters actually execute.
Memory And Personalization
Kucedr keeps profile Markdown in its agent workspace and manages durable memory separately:
| File | Purpose |
|---|---|
AGENTS.md | Generated context from completed identity, soul, and user profiles. |
BOOTSTRAP.md | One-time conversational setup for a fresh profile. |
IDENTITY.md | Assistant identity and presentation. |
SOUL.md | Personality and behavioral guidance. |
USER.md | Your profile and preferences. |
~/.kucedr/memory/MEMORY.md | Consolidated durable memory, managed by the memory module. |
HEALTH.md | Checklist used by periodic health runs. |
Save the identity, soul, and user profiles through their update tools; Kucedr then generates AGENTS.md. When enabled, the separate memory module processes saved conversations asynchronously, including requests to correct or forget information. Relevant memories are supplied to main chat and realtime voice. Memory, conversation history, and folder retrieval are separate stores. See Memory And Workspaces for details.
Apps
Apps are standalone mini-app windows discovered from installed app folders or registered external development folders. Settings supports import, Debug registration, refresh, Details, Open, and Delete. Cards themselves are not navigation links. The built-in Workspace is a main-window view, not an installed app. See Apps for import and data behavior.
Plugins, Channels, And Remote Agents
The Plugins guide distinguishes built-in MCP, database, and storage integrations from CLI-installed packages, whose contributions are not yet activated by the desktop runtime. Channels currently supports a Telegram bot with sender policies and a separate reply model. Remote Agents covers validated Agent2Agent connections and the data boundary for delegated work.
Cloud Backup
Cloud settings select a configured storage provider, local folders, schedule, backup, and restore. Restore requires confirmation, replaces matching files, and retains unrelated local files. Account authentication is separate from provider-backed folder backup, and the current app does not synchronize provider keys. This is folder backup, not bidirectional device sync.