Use PRISM research with an AI agent

*AI-generated recommendations for discussion. These are not Nick’s recommendations or approved JLL/PRISM guidance.

Updated September 29, 2026 (MCP 1.1).

Getting started

Connect this MCP when you want your AI tool to answer product questions or prepare a competitor brief using the library’s sources. It retrieves the evidence; your AI tool writes the answer. Product and design can inspect workflows and screens, sales can prepare discovery questions, and GTM teams can explore weekly changes and market context.

After connecting below, try: “Compare Cove and Visitt amenity reservations. Show the steps, screenshots, source dates and unknowns, then suggest questions to check against PRISM.” Your authorized internal tools are needed to verify PRISM itself.

Prefer to start in the browser? Open Getting started. No MCP is needed to read the library or export a page as Markdown.

Endpoint: https://crefieldnotes.com/mcp Web guide: Use with AI

The public MCP provides read-only evidence for product questions and deal preparation. Your connected agent performs the analysis using its own model. MCP queries do not trigger a paid search, collection run, or another LLM call. Do not send private deal notes or code to this public endpoint; combine private context in your own authorized tools.

What has been checked

Live JSON-RPC and official TypeScript SDK checks cover connection, tool discovery, search, workflow retrieval and image retrieval. The local suite includes citation and mixed static/live pagination regression tests. These checks do not certify every desktop client or guarantee correct AI answers.

Client setup depends on application version and company policy. The instructions below distinguish a tested SDK path from client-specific guidance.

Claude

For a remote MCP, use Settings → Connectors → Add custom connector where available, and enter the endpoint above. Organization settings may control availability. See Claude’s official remote connector instructions.

This route has not been interactively verified in your colleagues’ accounts. Do not use a remote-URL-only entry in claude_desktop_config.json as a universal setup instruction.

Cursor and other clients

In a client that supports Streamable HTTP MCP servers, add the endpoint using its remote-server settings. For Cursor, the project configuration example is:

{
  "mcpServers": {
    "prism-research": {
      "url": "https://crefieldnotes.com/mcp"
    }
  }
}

This is a configuration example, not a claim that Cursor, Windsurf, Cline, Goose or other applications were all tested. Choose HTTP/Streamable HTTP, not a legacy SSE transport. Follow your client’s current documentation and your organization’s access rules.

Tested TypeScript SDK path

The independent smoke check uses @modelcontextprotocol/sdk version 1.30.1 and StreamableHTTPClientTransport. In a Node project with that package installed:

import { Client } from '@modelcontextprotocol/sdk/client/index.js';
import { StreamableHTTPClientTransport } from '@modelcontextprotocol/sdk/client/streamableHttp.js';

const client = new Client({name: 'prism-research-client', version: '1.0.0'});
try {
  await client.connect(new StreamableHTTPClientTransport(
    new URL('https://crefieldnotes.com/mcp')
  ));
  const result = await client.callTool({
    name: 'get_workflow',
    arguments: {workflow_id: 'visitt-billing'}
  });
  console.log(result.content);
} finally {
  await client.close();
}

Official SDK client documentation

Available tools

Tool	Use
`search_research`	Find static research and matching live source evidence; use the returned cursor for more results.
`get_record`	Retrieve a record by stable ID, including source metadata.
`get_workflow`	Retrieve documented steps, reviewed or AI-extracted details, visual evidence and unknowns. `detail_level` says which; AI-extracted details cite their source articles and have not been reviewed by a person.
`compare_capability`	Assemble evidence for the same capability across companies.
`get_asset`	Retrieve an asset URL and metadata; request an inline image when supported. Formats a client cannot display, such as AVIF, arrive as the JPEG thumbnail.
`get_related_evidence`	Follow explicit research relationships one or two hops.
`get_customer_feedback`	Retrieve reviews and historical audience comments, preserving cohort and sample limits.
`get_changes`	Inspect source changes; unreviewed candidates require an explicit option, and the response says how many are waiting.
`get_market_context`	Retrieve market signals, macroeconomic observations and digest context.
`get_coverage`	Inspect freshness, failures grouped by cause, and known collection gaps; `health: "error"` lists only failing sources.
`list_research`	Browse all records by company and kind, with totals and pagination.
`get_document`	List and read the complete published Markdown reports, playbooks and flow guides.
`get_social_posts`	Retrieve original company posts and attributed AI analysis; distinct from customer sentiment.
`get_digest`	Read a complete weekly GTM edition, including recommendations, WHY, source citations and gaps.
`get_macro_series`	Page through stored FRED observation histories with units and retrieval dates.

Useful first questions

“How does Visitt handle billing verification? Show the sources and screenshots.”
“Compare Cove and Visitt amenity reservations for a tenant. Separate reviewed details from AI-extracted details and outlines.”
“What customer feedback would help us prepare discovery questions for an office operator?”
“Which sources have failed or become stale?”

For a product answer, ask for: direct answer; documented sequence; roles and exceptions; source links and visual evidence; remaining unknowns.

For PRISM verification, the agent must use its separately authorized codebase or documentation tools. This MCP does not establish current internal PRISM implementation, entitlement or production availability.

Evidence and freshness

Current system and monitoring coverage describes the deployed collectors, models and budget limits. The dated JSON snapshot is useful for agents; live get_coverage results take precedence for subsequent health.

Static research and images retain their original dates. Customer-review collection is scheduled daily; broader collection and the digest weekly. Audience comments are a historical unscheduled pilot. Check get_coverage for actual success rather than assuming a schedule completed.

Public documentation describes behavior; it is not proof from a tested customer account. Missing public evidence does not prove a missing feature. Reviews and comments are opinions and allegations, not verified defects. Images include crops, illustrations and historical training frames. Inline image support varies by client.

Recommended PRISM actions are AI-generated for discussion, not Nick’s recommendations or approved JLL/PRISM guidance.

Finding a complete answer

Start with search_research for a question, or list_research to browse without guessing keywords. For example, kind: "workflow_answer" lists the 58 documented workflows; kind: "asset" browses the screenshot catalog.
Use get_workflow for steps and exceptions, get_asset for the actual image, and get_related_evidence to follow source/claim relationships. A public outline is not a tested customer journey.
Use get_document for the full strategy, messaging, GTM, review analysis, experience maps or Confluence guide. Omit document_id to see the catalog. Relative links resolve against the returned document URL.
Use the specialized live tools for reviews, company posts, changes, weekly digests and macro data. Search results may also suggest the relevant tool. Yardi and MRI Angus are supported where collected evidence exists.
Follow next_cursor or next_offset until it is null; keep the same filters. Record/document/digest offsets count text characters. Other offsets count results. A total describes the full matching set, not the current page. Collections can change between requests, so record your retrieval time and deduplicate IDs during long exports. A page that would exceed what AI clients accept (about 40,000 characters) returns fewer items with response_trimmed; continue from the returned cursor or offset and nothing is skipped.
Check get_coverage: it reports all matching sources, errors beyond the current page (summary.failing_sources and summary.errors_by_reason), last successful collection and the latest run. Pages a vendor has restricted to signed-in customers or removed remain visible with health: "unavailable": a known gap, rechecked weekly, not a collection fault. A scheduled job is not proof of a successful refresh.

An unavailable dependency produces an explicit error, not an empty evidence set. Retry with backoff; do not turn a failed lookup into a claim about a competitor. The public transport accepts at most 128 KiB per request and 10 calls per legacy batch, with a per-isolate request guard. Invalid types, unknown fields and out-of-range arguments are rejected. MCP reads do not start collectors or paid AI calls.

Suggested answer format: direct answer; supporting steps and sources; relevant visual evidence; dates and evidence status; what remains unknown. Proposed PRISM actions need the AI-generated attribution and separate checks against authorized internal code or documentation.

Troubleshooting

If connection fails, check whether your client supports remote Streamable HTTP, whether your organization allows custom connectors, and whether the endpoint responds. An HTTP 429 means retry after the indicated delay. The current rate limiter is per Worker isolate, not a guaranteed global quota.

If you cannot add an MCP, open the agent index or download Markdown from a report. Browsing and file-import capabilities depend on your tool.

If a source is stale, unavailable or incomplete, preserve that limitation in the answer. Check cited evidence before using generated wording in a customer conversation.