Yes: the README explicitly provides a local HTTP API, an MCP Server, and an Agent Skill for Claude Code and Codex, covering the integration and output requirements you described.

• “Local HTTP API + MCP Server + AI Agent Skill” lists hybrid search, file reading, graph traversal, and source rescanning, so the client does not need to reimplement the core retrieval layer.
• The same section exposes the built-in API at 127.0.0.1:19828 and states that a bundled MCP server is available.
• “Rust Backend Chat Agent & Skills” says the Agent can use wiki/source/graph/web retrieval, workspace file tools, and approved shell commands.
• “Generated Outputs Preview” supports previewing Markdown, HTML, images, and other workspace files created by the Agent; external shell commands still require explicit approval.

This makes it a good fit as a local knowledge service for Claude Code or Codex. However, the README does not document remote multi-user access, an MCP client compatibility matrix, or recovery behavior after an abnormal API shutdown.

No: it should not be used directly as an authoritative knowledge base, although it is suitable as a source-traceable layer for understanding and maintaining technical materials. Read Sources Only narrows the evidence scope but cannot guarantee that the source or the model’s interpretation is always correct.

• “Source-grounded Retrieval” provides Read Sources Only, requiring answers to rely exclusively on imported original material and reducing unsupported expansion.
• “Async Review System” lets the LLM flag items requiring human judgment and provides predefined actions and pre-generated search queries; Quick Start also tells users to check Review.
• Quick Start recommends running Lint periodically to maintain Wiki health, indicating an ongoing maintenance loop rather than one-time page generation.
• Project insights explicitly state that generated knowledge may contain omissions, misclassification, or outdated information. Source tracking enables evidence recovery but does not eliminate hallucinations. The usage limitations also say the system cannot guarantee reliable conclusions for professional, legal, medical, or high-risk technical content.

It is therefore suitable for organizing, retrieving, and exploring materials, but not as the sole source of truth, a compliance database, or a high-risk decision authority.

Yes: the persistent ingestion queue, incremental cache, and source-folder watcher directly address interruption recovery and ongoing synchronization, rather than rebuilding context only when a question is asked.

• “Persistent Ingest Queue” explicitly supports serial processing, crash recovery, cancellation, retry, and progress visualization, which fits imports that may be interrupted.
• “Folder Import” supports recursive import while preserving directory structure and uses folder context as an LLM classification hint.
• “Source Folder Auto-Watch” monitors external changes under raw/sources/ and keeps ingest and delete cleanup synchronized.
• The project overview and “What is this?” emphasize incremental construction, persistent storage, and continuous maintenance. Complete project archives and rebuilding the Wiki index from existing pages are also supported.

It is therefore a better match than one-shot RAG when the goal is a long-lived, maintained Wiki. The README does not, however, specify an ingestion-size limit, single-machine performance, the queue’s persistence location, or parsing success rates for complex scanned PDFs and Office layouts.

Yes: the README provides knowledge-gap discovery, Deep Research, multi-query Web Search, and automatic ingestion of results, covering the workflow you described.

• “Graph Insights” lists surprising connections and knowledge gaps and supports one-click Deep Research, using the persistent knowledge graph as the entry point.
• “Deep Research” generates LLM-oriented search topics and performs multi-query searches through Tavily, SerpApi, or SearXNG.
• The same feature states that search results are automatically ingested into the Wiki, so research output becomes available for later retrieval rather than remaining only in the current conversation.
• “Rust Backend Chat Agent & Skills” also exposes web search as an Agent tool alongside wiki, source, and graph retrieval.

This fits a research workflow that expands a personal Wiki with external information. However, the README does not explain deduplication, copyright, or quality filtering for search results. It also does not say whether existing source text is sent to external search providers by default, and network availability and third-party API quotas are not defined.

Yes: the README explicitly connects embedded-image extraction from PDFs, factual descriptions from a vision model, image-aware search, and source navigation in one workflow.

• “Multimodal Image Ingestion” says the system extracts embedded images from PDFs and uses a vision LLM to generate factual captions.
• The same feature lists image-aware search results, lightbox preview, and jump-to-source, directly matching your requirements for image retrieval and page-level traceability.
• “Multi-format Document Parsing” also covers PDFs, Office documents, EPUB/MOBI, images, and media, so the input is not limited to plain text.
• Project insights state that original sources, Wiki pages, embeddings, and graph relationships are persisted, while source tracking provides an evidence path during retrieval.

This does not guarantee that image captions are correct. The README does not specify supported vision models, whether scanned PDFs follow the same pipeline, the recognition boundary for tables and charts, or the storage cost and ranking details of image indexing.

It depends: the project fits a multi-format, desktop-first personal knowledge base, but the README does not prove that every processing step stays local.

• The “Features” section lists PDF, Office, EPUB/MOBI, images, media, web clips, URL batches, and recursive folder import. PDF processing can use built-in, cloud, or local MinerU.
• “Flexible Model Configuration” allows per-project model configuration and independent Chat/Ingest routing. “Vector Semantic Search” supports LanceDB and OpenAI-compatible embedding endpoints.
• “Quick Start” requires configuring an LLM provider, API key, and model, so an external model service may be a prerequisite.
• Source traceability, Read Sources Only, and complete project archive import/export support long-term maintenance and evidence recovery.

However, the README does not specify which source text, images, or embeddings are sent to cloud providers under each configuration. It also does not document the range of supported local models, encryption, or enterprise-grade access controls.