MCP Evidence Tools | Synthetic Data Demo

MCP Evidence Tools | Synthetic Data Demo

Try the live demo with no installation: https://tribloc-mcp-tools.pages.dev/

Hosted version The public version connects through the official SDK to a real HTTP MCP endpoint. The same three read-only synthetic tools remain available locally over stdio. The combined suite has 36 passing automated tests, and all six public browser scenarios passed.

Project overview An independent, locally runnable Model Context Protocol demonstration using the official MCP SDK. A real stdio client/server connection exposes three read-only tools over two entirely fictional organizations. No internal project code, customer data, credentials or private architecture is included.

The problem AI integrations need an explicit boundary between available evidence, missing information and hypotheses. This demo makes that boundary visible while showing actual MCP tool discovery, structured responses and input validation.

My role I defined the architecture, bounded tool interface, synthetic scenarios and acceptance criteria. Implementation was assisted by AI. This is a public portfolio demonstration, not a client engagement or a claimed production outcome.

What was implemented A Node.js MCP server provides list_demo_accounts, get_demo_brief and check_demo_claim. A separate SDK client establishes the connection, discovers tools and executes calls. An optional local browser interface shows fresh protocol results for six scenarios. Inputs are restricted to explicit fixture IDs, claim enums and bounded limits; no arbitrary URL, file or write-action tools are exposed.

Validation and evidence 36 automated tests passed on 17 October 2026, including real stdio integration, output provenance, uncertainty handling, invalid-input refusals and local HTTP restrictions. All six browser scenarios were exercised. The public repository contains source code, pinned dependencies, tests, reports, a sample response and local setup instructions.

Scope and limitations All organizations, evidence and statements are invented. A supported claim is supported only by a synthetic fixture; the demo performs no real research or AI reasoning. No LLM, API key, external CRM or outreach is used. SDK-client integration was tested; an interactive Claude Code host session was not tested. This is a public synthetic demonstration, not a production service or a security-audit result.

Source and local setup: https://github.com/truSwen/synthetic-mcp-tool-demo