Skip to content

MCP tools

KAOS is MCP-native: every tool is typed, annotated, and exposable over the Model Context Protocol. Each package ships a kaos-*-serve entry point that serves its tools to any MCP client (Claude Code, Codex, …); kaos-mcp bridges any runtime to MCP over stdio or streamable HTTP.

Approximate counts (the authoritative list is generated per package — see below):

PackageMCP toolsServe command
kaos-core~10— (use kaos-mcp)
kaos-content~17
kaos-llm-client~7kaos-llm-serve
kaos-llm-core~35kaos-llm-core-serve
kaos-agents~14 (+7 retrieval)kaos-agents-serve
kaos-pdf~8kaos-pdf-serve
kaos-office~18kaos-office-serve
kaos-tabular~17kaos-tabular-serve
kaos-source~30kaos-source-serve
kaos-web~45kaos-web-serve
kaos-nlp-core~17kaos-nlp-serve
kaos-nlp-transformers~7kaos-nlp-transformers-serve
kaos-ml-core~11kaos-ml-serve
kaos-graph~17kaos-graph-serve
kaos-citations3kaos-citations-serve
kaos-ui4 (read-only)
kaos-mcp0 (pure bridge)kaos-mcp serve
kaos-names0

kaos-llm-core exposes its vision-language-model page programs as three MCP tools, so any MCP client can run a model over a page image: kaos-llm-core-vision-ocr (transcribe every word), kaos-llm-core-vision-describe (page structure — headings, tables, stamps, signatures, redactions), and kaos-llm-core-vision-classify (one of ten page types with a confidence). Each takes a file path or base64 image_base64 and an optional model. These are the escalation path behind garbled-scan recovery when a deterministic OCR engine isn’t enough.

  • Every tool is a KaosTool with typed inputs (schema derived from type hints) and annotations (readOnlyHint, destructiveHint, …) that let an agent’s permission policy decide what to auto-allow.
  • kaos-mcp mounts a whole runtime: kaos-mcp serve (stdio) or --http (streamable HTTP, which requires an explicit auth token).
  • Connect a serve command to an MCP client to make KAOS tools available inside it.
Terminal window
uv run --with kaos-pdf kaos-pdf-serve --help # or list-tools via kaos-mcp