overview-navigation
DENOJU Overview & Navigation
DENOJU is a local-first AI generation OS connecting on-Mac local models, conversations, creation tools, external agents, and memory.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
The core functionality spans 13 areas. Chat manages everything from inquiries to tasks and finished artifacts within the same conversation. Harness manages tools and approvals, while Addons provide surfaces for creation, configuration, and history. Models handles MLX/GGUF installation, loading, and fit assessment; Meeting Room coordinates roles between humans and multiple AIs; and External Agents connects with separately installed and authenticated CLIs. Marketplace delivers modular features, Projects isolates workspaces and context, Login separates DENOJU accounts from third-party providers, Settings manages connections and permissions, and Memory governs conversation, project, and durable memory. Generation covers documents, code, web, and supported media, while automation manages Bots, Routines, and Schedules. This overview presents capabilities first, followed by prerequisites and boundaries.
chat-basics
Chat Basics & Routing
Handle questions, task requests, local and external LLM responses, and artifacts in a unified chat.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Load an installed MLX or GGUF model and ask questions directly from the shared input field. You can switch the active responder between local models and enabled external LLMs, and all conversations initiated in DENOJU appear in a shared history. The 'External' history filter refers to conversations imported from external tools, not a separate composer for external LLMs. When transferring previous context to an external LLM for the first time, choose between full text, a locally generated summary, or cancellation; when only summary is approved, unreviewed full text is never sent. Returning to a local model never triggers external network transmission. Response modes include Quality, Balanced, and Speed, distinguishing progress states, pending approvals, artifacts, and errors from conversational text. Physical interaction with external CLIs depends on installation, authentication, and CLI versions.
chat-context
Chat Context, Continuity & Compaction
Prioritizes the current request and assembles only relevant history, attachments, search results, and memory into a bounded context.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
Independent new topics never inherit unrelated historical context. In continuous tasks, DENOJU succinctly preserves objectives, confirmed facts, pending items, and recent dialogue. Long histories are compacted via ChatContextCompactor; upon token limits or interruptions, tasks resume without resending previously transmitted full text. Content from separate chats is considered only upon explicit user inquiry. Official Manual Memory is added strictly for product inquiries and operates independently of personal Conversation Memory settings.
chat-tools
Chat Tools, Artifacts & Approvals
Invokes corresponding Harnesses or Host tools strictly upon detecting verified task intent.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Tool names mentioned in quotes, logs, or code snippets do not trigger execution. File edits, shell commands, external transmission, deletion, publishing, and Computer Use strictly adhere to scopes and approval boundaries. Generated Documents, Media, Websites, and Code are displayed as artifacts only after verifying their physical existence and validation signals. Unexecuted commands or non-existent files are never marked as completed.
tool-capabilities
Available Tools in DENOJU
Distinguishes between functional classifications in the manual and currently active Host capability summaries.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Questions answered purely through conversational knowledge require no Host tools. The default capability directory includes Plan, Browser/Web research, local image generation, and Documents authoring; Mac App bridges and other tools are added based on active Parts in MYCORE. Broad Chat task paths include CodeWiki/Memory search, Media, Photos, Contacts, Notes, Reminders, Calendar, Editor/Patch, Shell/FileOps, WorkGraph, Website, Coding, Mail, and more. Each capability's execution depends on active Parts, addons, runtimes, projects, connections, macOS permissions, and approvals. The `[DENOJU capability summary]` accompanying your prompt serves as the canonical record of active executable candidates. Web searches are not needed merely to answer manual inquiries. When capabilities are present, responses explain available items, prerequisites, and execution locations rather than simply stating 'unconfirmed'.
harness-core
Harness Core Specifications
Harness acts as an execution contract linking model decisions to the UI, addons, execution processes, and artifact verification.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
A Harness is an execution contract rather than a prompt asset like a Skill or Rule. Built-in contracts exist for Conversation, Plan, Research, Code, Document, Image, Video, and Evaluation, declaring required Parts, permissions (file, network, process, model loading), output formats, approval requirements, and validation signals. Execution states are queued, running, waiting-approval, succeeded, failed, and cancelled. Operations cannot proceed if required Parts or permissions are absent.
quality-harness
Quality Harness & Model Evaluation
Evaluates local model comprehension, routing, responses, continuity, latency, and stability across test fixtures.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
Quality evaluation isolates whether the Host assembled proper context, the model utilized it, the runtime completed execution, and the UI displayed results accurately. It utilizes golden fixtures, training repair sets, holdouts, counterfactuals, mutations, durability tests, safety suites, presentation checks, and actual model matrices. Results are stored per model/profile/runtime fingerprint, feeding failures back into repair sets or Model Operating Memory. Enabling evaluation never grants arbitrary external transmission or destructive execution.
addons-architecture
Addon Architecture & Security Boundaries
Governs Addon Core, Mac App bridges, and Marketplace packages under distinct origins, permissions, and runtimes.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
An addon package declares its version, capabilities, providers, UI surfaces, policies, schemas, fixtures, permissions, filesHash, signatures, update channel, and revocation in addon.json. Declarative addons cannot hold execution providers, while isolated-providers require explicit provider declarations and permissions. The current validator checks a 64KB manifest limit, ID/version/path formats, filesHash and signature field structures, and self-reported revocation; it does not yet guarantee cryptographic publisher verification via trusted public keys or live external revocation feeds. Users must independently verify the source and contents of external packages. Displaying a panel grants neither OS permissions nor external transmission authority.
addons-builtin
Built-in DENOJU Addons
Handles documents, web, graphics, media, monitoring, maps, and more via addons, routing AI instructions through the shared chat.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
Built-in Addon Core features include Graphics, Media, Documents, Website Builder, Password Generator, Agent Plans, LIVE Monitor, Activity Monitor, Prompt Pool, Knowledge Learning, Local Chat Durability, Cyber Tree, WorkGraph, Weather, Map, Siri, and others. Each panel provides configuration, progress tracking, artifacts, and history. AI commands and generation proceed from local chat, shared chat targeting external LLMs, or meetings into Host capability verification and execution, returning outcomes to originating chats and addon panels. The living floating chat in the bottom right uses the shared draft, history, destination, and delivery pipeline. When the chat column is visible, it focuses the existing input, preserving drafts without auto-submitting. Panels do not contain standalone prompt inputs or generate buttons. Manual editing of existing artifacts and recovery of submitted results are supported. Availability depends on Parts, setup, entitlements, runtimes, and OS permissions.
addons-mac-apps
Mac App Bridge Addons
Connects supported on-Mac apps as App Core bridges to the right panel and Chat.
- Availability
- Requires additional installation
- Execution Location
- On this Mac
Supported app registries include Photos, Notes, Contacts, Calendar, Mail, Reminders, Music, iPhone Mirroring, Meeting Link, Pencil, Safari, Shortcuts, Home, Messages, FaceTime, and Passwords. They require the app to be installed on the Mac, macOS permissions granted, and the target app in a responsive state. Disconnecting a bridge never deletes source app data. Side effects like reading, writing, sending, or placing calls adhere to explicit bridge actions and user approvals.
model-types
Model Types, Storage Sources & Executability
Distinguishes between conversational LLMs/VLMs and image/video/audio generation models, showing source storage and execution status.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
LLMs power text conversations, while VLMs handle multimodal conversations with image understanding; VLMs themselves do not synthesize image pixels. Image, Video, and Music models are dedicated to media generation. Supported formats include MLX, GGUF, and unknown, sourced from DENOJU, LM Studio, Ollama, and local directories. MLX packages can be linked when criteria are met, and specific GGUF paths execute via an embedded llama-server. Candidates lacking an executable path or companion models may serve solely as references rather than standalone chat models.
model-lifecycle
Model Manager & Lifecycle
Manages discovery, installation, verification, loading, unloading, and deletion of local models.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Choose models from catalog items, Hugging Face search, or local file imports. Model installation and loading depend on available system memory and format compatibility. Compatibility ratings serve as guidelines relative to your Mac's unified memory. Switching models automatically unloads previous active models. Deletion targets downloaded files only, leaving manually created assets untouched.
model-runtime-profiles
Runtime Profiles, AutoTune & Acceleration
Optimizes local execution via context length, quantization, GPU/CPU allocation, and AutoTune.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
Select profiles tailored for Quality, Balanced, Low Memory, or Maximum Speed. AutoTune dynamically adjusts inference parameters based on your Mac's hardware and current memory pressure. GGUF models leverage Metal acceleration, while MLX utilizes Apple Silicon unified memory directly. System warnings or limits activate if aggressive settings threaten memory stability.
external-cloud-models
External Cloud Models & Execution Boundaries
Connects with OpenAI, Anthropic, Google, and other cloud APIs for specialized capabilities.
- Availability
- Cloud / Network enabled
- Execution Location
- Cloud
Using cloud models requires configuring valid API keys for each provider. Execution occurs on the respective provider's remote servers, incurring network traffic and API usage fees. Conversation history and context transmission strictly adhere to user configuration. When working with sensitive data, prefer local models or review context sharing settings carefully.
meeting-room
Meeting Room Core Specifications
A collaborative workspace where multiple AI models and human participants deliberate and execute with distinct roles.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Participants can be assigned roles such as Facilitator, Researcher, Coder, or Critic. Meeting rooms can mix local models, external agents, and human collaborators. Speaking order, agenda progression, and consensus building are guided by the Meeting Coordinator. Minutes, decisions, and action items are automatically summarized.
meeting-workflow
Meeting Workflow, Verification & Integration
Executes tasks decided in meetings through Harness and Coding engines, verifying and integrating results.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Agreed meeting outcomes convert directly into Action Items or WorkGraph tasks. Assigned AIs handle code modifications and document drafts, while reviewing AIs evaluate quality. Changes requiring authorization are applied to projects only after explicit user approval. All steps are recorded in structured audit logs.
external-agents
External Agents & CLI Models
Connects with external agent CLIs such as Codex, Claude Code, and Cursor, invoking them directly from DENOJU.
- Availability
- External service integration
- Execution Location
- External
Requires compatible CLI tools to be installed on macOS and authenticated with their providers. DENOJU communicates via standard I/O, MCP, or local socket protocols. Autonomous actions by external agents remain guarded by approval gates, preventing unintended file modifications or data egress. Execution results integrate into DENOJU history and artifact viewers.
marketplace
Marketplace Catalog & Extensions
A directory to discover, install, and manage addons, Brain Packs, Skills, and themes.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Browse verified catalog items and community packages. Review required permissions, execution models, and author metadata prior to installation. Packages failing security standards are restricted or blocked. Installed extensions can be enabled, configured, or uninstalled at any time in Settings or Addons.
brain-packs
Brain Packs, Skills, Rules & Agents
Add specialized bundles of knowledge, rules, and procedures tailored for specific industries or programming languages.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Brain Packs package prompt instructions, domain knowledge (Memory), operational procedures (Skills), and quality standards (Rules). Once added, they are referenced automatically or explicitly during relevant tasks. In case of conflicts with user settings or existing rules, precedence orders can be customized.
projects
Projects & Workspace Boundaries
Isolates and manages target directories, files, context, and memory on a per-project basis.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
Opening a project limits file operations strictly to the selected directory tree. Project-specific Memory, rules, and Git repositories are preserved independently. Access to files outside the project boundary requires explicit user authorization. You can switch smoothly between multiple active projects.
account-login
DENOJU Account & Authentication
Authenticates DENOJU accounts for license management, cloud sync, and device inventory.
- Availability
- Cloud / Network enabled
- Execution Location
- Cloud
Sign in via magic email links or OAuth (Google, GitHub, etc.). Credentials are encrypted and securely stored in the macOS Keychain. Unauthenticated use of local capabilities (inference, chat, document creation) remains fully supported. Account features manage cross-device licensing, shared meetings, and remote sync only.
login-types
Authentication Types & Provider Separation
Strictly separates DENOJU platform accounts from third-party service credentials (GitHub, API providers).
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
DENOJU accounts manage product entitlements and device pairings. Using external LLMs or agents requires separate API keys or tokens from their respective providers. DENOJU never transmits third-party credentials to its own servers without permission. All API tokens reside securely in local Mac storage.
settings-overview
Settings Overview
Centralized control for general preferences, appearance, models, connections, permissions, memory, and safety.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Access Settings from the sidebar or system menu. Preferences take effect immediately and are persisted safely in local configuration stores. Settings support export, import, and resetting to factory defaults.
settings-connections
Connections, Search, Remote & MCP
Configure external APIs, search providers, remote device pairing, and Model Context Protocol (MCP) servers.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Register API keys, configure proxy settings, and manage local or remote MCP servers. Built-in connection testing verifies reachability prior to enabling integrations. Unused connections can be deactivated or deleted instantly.
settings-intelligence-autonomy
Intelligence, Autonomy & Safety Settings
Fine-tune autonomous decision limits, approval thresholds, guardrails, and execution budgets.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
Set confirmation policies per action (always prompt, prompt on sensitive actions, or auto-run). High-risk actions such as file deletions, shell executions, external data transmission, or financial transactions strictly mandate explicit confirmation. Safety filter sensitivities can be adjusted.
memory
Conversation, Project, Durable & Official Memory
Hierarchical memory management covering ephemeral dialogue, project context, durable user knowledge, and official manual memory.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
Conversation Memory preserves immediate interaction context. Project Memory records requirements and architectural decisions for specific repositories. Durable Memory stores persistent user preferences and domain insights. Official Manual Memory delivers grounded facts about DENOJU capabilities. Users maintain full control to inspect, edit, or delete stored memories.
documents-artifacts
Documents, Attachments & Artifacts
Generates PDF, Word, Excel, and PowerPoint files while reading and modifying existing documents.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
Specify document format, content, and save path directly in chat. Writers for PDF, Word (DOCX), Excel (XLSX), and PowerPoint (PPTX) operate through Host document pipelines. Attachment parsing depends on format and extractor compatibility. Artifacts return to the conversation and addon panels for review and export. Generated documents do not guarantee legal or business validity; always verify final content.
media-website-canvas
Media, Website & Canvas Generation
Produce images, videos, music, websites, and interactive canvases from shared chat. Media requires dedicated models and runtimes.
- Availability
- Requires additional installation
- Execution Location
- Mac or External depending on config
Website Builder handles site listings, Edit STUDIO, design/SEO/block modifications, and static site export to designated folders. Prompts and edit requests are issued through shared chat without standalone generation inputs in addon panels. Image, video, and audio generation require compatible models and local runtimes such as media-gen, which are not preinstalled by default. ComfyUI integration requires connecting to a running instance on Mac. The conversation model may be cloud-based while the media engine runs locally, or vice versa. Canvas provides an interactive board for widgets and artifacts. VLM image understanding does not generate images.
coding-computer-use
Coding, Editor, Terminal & Computer Use
Request code editing, build/test validation in selected projects, and governed Mac UI interactions through chat.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
Coding operates using project roots, contracts, target files, and build/test observations, executing steps across reading, editing, building, testing, and terminal tasks. Subsequent requests reference existing state and test results. Computer Use utilizes Accessibility and Screen Recording permissions along with allowlisted app configurations to inspect screens, click, type, and verify outcomes. Model suggestions pass through Host parsers and approval gates before execution. Actions such as sending messages, deleting files, purchasing, authenticating, or accessing arbitrary URLs require explicit confirmation. System permissions alone do not bypass approval gates. Incomplete setups, missing tools, or unsupported apps may prevent task completion.
bot-automation
Auto Work Bots, Routines, Schedules & WorkGraph
Executes recurring automated workflows bounded by declared capabilities, models, budgets, tools, memory, and schedules.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Bots define capabilities, inputs, model routing, Harnesses, audits, projections, group chats, and post-work routines. Routines and Schedules provide contracts to reuse and run workflows periodically. WorkGraph coordinates dependencies and live execution status. Automation is strictly bounded by budgets, tool permissions, project scopes, network rules, approvals, and stop conditions. Scheduling a task never eliminates the requirement for sensitive action confirmations.
growth-mycore-selfmod
MYCORE, Growth, SelfMod & Learning
Accumulates memory, Skills, Rules, runtime configurations, and evaluation benchmarks to enhance future workflows.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
MYCORE configures functional Parts and user panels. Memory preserves explicit knowledge and search boundaries, Skills capture operational procedures, and Rules enforce quality gates and restrictions. Growth synthesizes successes, failures, and evaluations into Growth Memory and operating wisdom. This growth does not mean model weights update automatically through conversation alone. SelfMod operates under strict boundaries of proposal, isolated patching, verification, and Host handoff, never replacing core binaries silently. LoRA fine-tuning requires dedicated training runtimes and explicit user actions. Proprietary, sensitive, or restricted data is never ingested into training corpuses without authorization.
other-features
Additional DENOJU Capabilities
Integrates browser automation, web search, Siri, notifications, audio, graphics, canvas, and remote control into core workflows.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
Browser/Web handles search and web inspection, Siri/Voice manages speech input and commands, Notifications surfaces task status, Graphics delivers visual outputs like Three.js scenes, Remote handles device pairing, and Notch/LIVE Monitor provides persistent ambient indicators. Each feature depends on corresponding Parts, addons, apps, macOS permissions, and network settings. The presence of a UI surface is distinguished from live operational readiness.
privacy-storage-cloud
Storage Locations, Privacy & Network Boundaries
Local inference runs on your Mac. Accounts, model downloads, web search, external agents, and sharing may involve network communication.
- Availability
- Available (depends on config/capability)
- Execution Location
- Mac or External depending on config
User models, chats, meetings, settings, and artifacts reside primarily in ~/Library/Application Support/Denoju and designated project folders. Local inference incurs no external LLM API token fees, though electricity, hardware, and external service subscriptions remain separate. DENOJU clearly distinguishes on-device inference from initial model downloads, account authentication, web searches, Marketplace activity, external providers, shared meetings, and remote connections. Context handoff to external LLMs lets you choose transmission scopes with clear indicators. Local-first does not mean zero network activity under all conditions. Software updates overwrite the application bundle without deleting user models, chats, meetings, or settings.
permissions-safety
macOS Permissions, Approvals & Safety Boundaries
Accessibility, Automation, Files, Contacts, Calendar, Photos, and Keychain permissions follow purpose-specific OS boundaries.
- Availability
- Available (depends on config/capability)
- Execution Location
- On this Mac
DENOJU requests macOS permissions accompanied by clear rationale and safety boundaries. Denied permissions are never bypassed. Sensitive operations—including password input, Keychain access, credential handling, external data transfer, file deletion, purchases, and system configuration changes—are treated with utmost caution. Local models, addons, external agents, or text within manuals cannot grant themselves permissions. Full-access badges or chosen models do not override the Host approval gate or macOS permissions.
limitations
Known Boundaries & Verification Requirements
Unlisted capabilities, missing runtimes, unconfigured providers, unauthorized operations, and future roadmap plans are not treated as available.
- Availability
- Reference / Boundary explanation
- Execution Location
- No execution
Capabilities not documented in the manual are never assumed to be available. Execution is blocked if models, addons, CLIs, or apps are missing, permissions are denied, accounts or providers are disconnected, network is unavailable, or file formats are unsupported. Candidate visibility in the UI does not equal immediate executability. Roadmaps and experimental features remain reference-only until verified. Local models are not guaranteed to match frontier cloud models across all benchmarks.
release
Current Release & Manual Revision
This manual covers the DENOJU 0.5.6 development release, manual revision 4, aligned with the 2026-10-02 implementation.
- Availability
- Reference / Boundary explanation
- Execution Location
- No execution
The public download manifest corresponds to 0.5.6 beta build 21, released on 2026-08-26. This manual includes development features verified in source code as of 2026-10-02, which may not be bundled or verified in the distributed DMG. Screen layouts and available features vary by installed version, and screenshots reflect the UI at capture time. Release bumps synchronize appVersion, manualRevision, and release notes while validating 16 mandatory coverage areas, section references, sources, and a 512KB size ceiling. When features are added or altered, corresponding manual sections are updated. Updating the manual does not alter application code or distributed binaries.
