Extending ToolStorePy

ToolStorePy is deliberately modular. Most extension work maps cleanly onto one file or subsystem.

Common extension points

What you want to change	Where to look
Add a built-in index	index/registry.py
Change embedding or reranking models	orchestrator.py constructor
Change index download logic	index/downloader.py
Adjust repo caching behavior	loader/cache.py
Add or tighten AST security rules	utils/security_scanner.py
Tune the LLM scan system prompt	utils/llm_scanner.py
Change LLM scan token budget	utils/llm_scanner.py (MAX_REPO_CHARS, MAX_FILE_CHARS)
Change env merge behavior	utils/env_merger.py
Change decorator detection	builder/parser.py
Change generated server structure	builder/mcp_builder.py
Change default host/port	orchestrator.py constructor defaults

LLM scanner extension

utils/llm_scanner.py uses LangChain's init_chat_model and with_structured_output. To change the verdict schema, update the LLMVerdict and LLMFinding Pydantic models. To tune what the model looks for, edit _SYSTEM_PROMPT.

The scanner is provider-agnostic. Any LangChain-supported model string works.

Adding a built-in index

In index/registry.py, add an entry to BUILTIN_INDEXES:

BUILTIN_INDEXES = {
    "core-tools": {
        "url": "https://...",
        "sha256": "abc123...",  # fill in after publishing
    },
    "my-index": {
        "url": "https://example.com/my-index.zip",
        "sha256": None,
    },
}

The sha256 field enables integrity verification in the downloader. Set it once you have a stable release.

Good extension patterns

• keep retrieval changes separate from build synthesis changes
• preserve auditability in generated outputs
• add evaluation coverage when you change parser or build behavior
• treat security rules as user-facing policy, not only internal code

Related contributor workflow

If you are extending the ecosystem rather than the core codebase, see Publishing a Tool Index.