AI MCP Guide
Connect your AI client to search TrustDecision documentation.
Connect your AI client to search TrustDecision documentation.
The TrustDecision documentation MCP service provides read-only access to public documentation through the Model Context Protocol (MCP). AI agents can use it to search and read documentation.
https://docs.trustdecision.com/mcptrustdecision_docs is an example name; you can choose another. Use the same service URL in every client. No language path is required.
Choose one configuration method for your client and workflow. Desktop clients can use their settings or configuration files without a separate command-line interface (CLI) installation.
Menu labels may vary by client version. When editing an existing configuration file, merge the new server entry and keep the existing settings. In the paths below, ~ refers to your home directory.
Desktop (local projects)
Add the following configuration to .mcp.json in your project root:
{
"mcpServers": {
"trustdecision_docs": {
"type": "http",
"url": "https://docs.trustdecision.com/mcp"
}
}
}Open the project in Claude Code on desktop, start a local session, and approve the project MCP configuration when prompted. This configuration also applies to the CLI when launched from the same project directory.
CLI
Run this command to add the server to your user configuration:
claude mcp add --transport http --scope user trustdecision_docs https://docs.trustdecision.com/mcpRun claude mcp list to inspect the configuration. In a Claude Code CLI session, enter /mcp to check the connection status. To limit the configuration to the current project, replace --scope user with --scope project.
References: Claude Code MCP configuration, shared desktop configuration.
Use an MCP client that supports Streamable HTTP and configure the service URL above. Follow the client's documentation for its configuration format.
For a custom agent, use an MCP client SDK to initialize the connection, discover tools, and call them. Handle results as described in the agent reference below.
Send this prompt to your AI agent:
Use trustdecision_docs to find the Device Information Query API documentation.
Read the relevant page, explain the endpoint and required parameters,
and include source links. If the documentation is insufficient,
state which information is missing.Confirm that the agent actually calls the documentation tools and returns relevant content with source links. If you chose a different server name, replace trustdecision_docs in the examples.
The service exposes three tools:
Include the product, platform, and use case in your question. If a particular version or deployment region matters, include it as well.
Use trustdecision_docs to find the Device Fingerprint Android SDK documentation.
Outline the steps and prerequisites for initialization, obtaining blackBox,
and querying device information from the backend.
Cite the relevant documentation and distinguish SDK operations from server-side API calls.Use trustdecision_docs to explain the black_box field in the Device Information Query API.
Where does its value come from, when is it passed to the API,
and what requirements apply? Include source links.I receive an empty blackBox when integrating the Device Fingerprint Android SDK.
Use trustdecision_docs to find initialization requirements, usage notes,
and troubleshooting guidance. If you need the SDK version or other details,
list the information needed before drawing a conclusion.This section is for agent developers. The JSON examples below are tool arguments, which the MCP client wraps in a tool-call request.
First, call search_reference to locate a page:
{
"query": "Device Information Query API black_box",
"locale": "en",
"limit": 3
}Take the page identifier from data.results[].document.doc_id and call get_reference_page. Also pass the top-level content_revision from the search result to verify that both calls use the same content version. If that revision is no longer available, search again. Search snippets help locate a page; read the relevant source before answering questions about parameter requirements or other details.
get_reference_page uses mode: "full" by default to read the page content. Use mode: "outline" for the section outline, or mode: "sections" for specific sections and their required prerequisites. Section reads require section_ids and the corresponding content_revision. Always use page and section identifiers returned by the tools.
For a specific technical question, call get_reference_context:
{
"question": "How do I initialize the Device Fingerprint Android SDK and obtain blackBox?",
"locale": "en",
"intent": "integration",
"top_k": 5
}The tool returns data.evidence, data.sources, data.missing_requirements, and data.follow_up_reads. The calling AI agent uses these results to compose its answer. If the sources are insufficient, make additional search_reference and get_reference_page calls. Do not infer API behavior that the documentation does not establish.
For search_reference and get_reference_context, set locale to cn (Chinese), en (English), or all (both languages, the default). For page reads, pass the doc_id returned by search to get_reference_page.
Documentation language does not determine the deployment region. The search and context tools accept filters for product, platform, region, and other criteria. Apply filters only when the relevant conditions are known, using the values listed in the documentation index.
Use get_reference_page to read td:overview:en:v3 or td:overview:cn:v3 for available products and filter values. filter_match: "unknown" means applicability has not been established for the supplied filters. source_facts records versions, regions, and other information found in the source; it does not define the full supported range.
get_reference_context assembles evidence from complete sections and their required prerequisites. Its max_chars parameter limits evidence text in Unicode code points. It excludes response metadata and is not a model token count. If omitted, the default budget applies. Check the discovered tool definitions for each tool's parameter limits.
If the budget is insufficient, the tool includes only complete sections and their prerequisites that fit. It lists additional sections in data.follow_up_reads. Call get_reference_page with mode: "sections", passing the doc_id, section_ids, and content_revision from the relevant entry. With a very small budget, data.evidence may be empty; use the follow-up entries to read the source.
If a search or page-read result has a non-null top-level next_cursor, pass it as cursor in the next call. Keep the original query or read parameters and the content_revision. To retrieve the full result, continue until next_cursor is null. Do not change limit, max_chars, or filters between pages.
data.content contains the document fragment returned by the current page-read call. For display, combine data.presentation.prefix, data.content, and data.presentation.suffix to include reading aids such as code fences and table headers. To assemble content across pages, concatenate only data.content in offset order. An incomplete code block or table must not be treated as a complete example.
Prefer the MCP response's structuredContent. If the client exposes only text content, parse the JSON in the relevant content text block. The result payload is in data. Use the top-level status, issues, truncated, next_cursor, and error fields to determine the next action.
Include source links in answers and identify any conditions that still need confirmation.
No. This MCP service provides public documentation and does not require TrustDecision business API keys or customer credentials. Business API authentication remains subject to the relevant product documentation.
/mcp is a protocol endpoint for MCP clients, not a web page. Opening it directly in a browser may return 405. Verify the service through client configuration and an actual tool call.
Check that the server and its tools are enabled and that the current session has loaded the configuration. Explicitly ask the agent to "use trustdecision_docs." In a desktop app, check the current session's connector or MCP list. For a CLI, check its working directory and configuration scope. If the client requests permission for a tool call, follow its prompt.
Confirm that your client version supports custom remote MCP servers, then consult its official setup instructions. An enterprise workspace may require an administrator to add or approve the service. Desktop, CLI, and cloud sessions can have different configuration scopes; check which tools the current session actually discovers.
Some clients connect to remote MCP servers from the cloud. Use the public HTTPS URL on this page and check that the agent's environment can reach it.
For HTTP 429, wait as indicated by Retry-After and reduce concurrent or repeated requests. For CONTENT_REVISION_UNAVAILABLE, discard old cursors and search again to obtain the current revision. For INVALID_CURSOR, restart from the first page. Do not edit cursors or reuse them across unrelated requests.
No. Documentation updates do not require a change to the service URL. The MCP service and documentation site are updated together during build and deployment, so the service provides the documentation included in that release.
This service searches and reads documentation. It does not execute business API requests, run SDKs, or make risk decisions on behalf of a business system. An agent's integration guidance should be supported by the documentation. For products, environments, and permissions enabled for a particular customer, use the formally agreed integration details.