Internal CSM & AE Beta Test Guide
March 2026
|
Purpose of This Guide This guide helps CSMs and AEs run the first internal beta test of Compass MCP — CupixWorks' AI-powered site intelligence assistant. Use it to understand what Compass MCP does, how to set it up in Claude, what to test, and how to document feedback. |
1. What is Compass MCP ?
MCP stands for Model Context Protocol. Compass MCP connects CupixWorks project data to AI Agents (Claude, ChatGPT, Copilot and more) so users can ask natural language questions and get AI-generated insights — without logging into the CupixWorks web interface or navigating dashboards manually.
It covers two types of questions:
| Type | What You Can Ask |
|---|---|
| Capture-based | Questions about what's visible in site photos — safety issues, material conditions, site cleanliness, construction progress as captured on specific dates |
| Progress-based | Questions about SiteInsights analytics data — completion rates by trade or area, schedule delays, forecasted completion |
|
Note Compass MCP is currently accessed entirely through Claude (claude.ai or Cowork). There is no separate Compass MCP interface — everything happens in a Claude conversation after connecting the CupixWorks MCP connector. |
2. Available Tools
Compass MCP has four tools that Claude uses automatically depending on your question. You don't need to specify which tool to use — just ask naturally.
| Tool | What It Does |
|---|---|
| Site Image Explorer | Searches across site photos to find images matching your description — e.g. safety hazards, specific materials, or conditions at a particular date or location |
| Project Data Explorer | Queries structured project records — capture dates, levels, workareas, capturers, and progress counts |
| SiteInsights Analyzer | Analyzes construction progress data from SiteInsights Pro — overall completion, by-category breakdowns, schedule delays, and forecasts |
| Trade Visibility Analyzer | Summarizes trade detection results by workarea and trade — showing what has been visually confirmed on site |
3. What Compass MCP Can and Cannot DO
3.1 What it Can do
- Find site photos matching a natural language description
- Answer questions about capture history — dates, levels, who captured, how many
- Analyze SiteInsights progress — overall %, by trade, by area, by schedule
- Analyze work progress by work area and trade based on capture data (SiteInsights Lite)
- Identify which areas or trades are behind schedule
- Show direct links to specific site locations in SiteView
- Combine data from multiple sources to answer a single question
- Generate reports with analysis results
3.2 Current Limitations
| Limitation | What This Means |
|---|---|
| Read-only | Compass MCP cannot create annotations, update settings, or change any data in CupixWorks |
| One project per session | Each Claude session must be started with a specific project. Compass MCP does not switch projects automatically mid-conversation. |
| Image relevance may vary | Photo search sorts results by similarity. The first one or two links are usually the most relevant — later results may be less accurate. |
| SiteInsights license required | The SiteInsights Analyzer only works for projects where SiteInsights is set up and running. |
| Not real-time |
There are two reasons data may not reflect the latest activity. (1) Site photos become searchable only after post-processing and CQA review are complete — not at the earlier "Preview Ready" stage. (2) SiteInsights results sync to the database three times per day (every 8 hours) after SQA (SiteInsights Quality Assurance) review, so progress data may lag by up to 8 hours. |
4. Setup
4.1 What you need
- A Claude Team account (Cupix internal)
- Cupix Login Credentials
4.2 Connecting Compass MCP to Claude
- In Claude, go to Settings → Connectors.
- Find cupix-compass-mcp(PROD) and click Connect.
- Complete Warden Login when prompted.
- Once connected, click Configure next to the connect
- Under Tool Permissions, select Always Allow.
|
Tip Set Always Allow before starting a test session. Without it, Claude will ask for your permission before each tool call, which interrupts the flow of the conversation. |
4.3 Recommended Settings
| Setting | Recommended | Why |
|---|---|---|
| Model | Claude Sonnet 4.5 | Best balance of speed and accuracy for this use case |
| Extended Thinking | Off | Not needed here — adds wait time without improving results |
| Interface | Cowork (preferred) or Web Chat | Cowork gives a cleaner view by hiding behind-the-scenes tool activity |
4.4 Starting a Session
At the start of each new Claude conversation, paste the following line to connect to the test project. This skips the manual team and project selection steps.
|
Cupix Compass MCP. Team: cupix, Facility: cjj60d |
| Test Project | Details |
|---|---|
| Project Name | Madison Elementary School |
| Facility ID | cjj60d |
| Capture Range | Feb 2022 – Dec 2023 (8 capture dates, 37 captures) |
| Coverage | Early structural phase through post-finishing — wide variety of site conditions to explore |
|
Note If you skip the initialize prompt, select any team at the team step — Madison Elementary School (cjj60d) will always appear at the top of the facility list. |
5. What to Test
5.1 Test Goals
- Check that each tool returns accurate, relevant results for realistic questions
- Find cases where results are wrong, confusing, or unhelpful
- Assess whether responses would be genuinely useful to a project manager or site engineer
- Put yourself in the customer's shoes — think and ask like the end users who will actually use this product
Key Testing Mindset
You communicate with customers every day. Use that advantage. Before running a query, ask yourself: What would a Project Manager actually want to know on a Monday morning? What would a Superintendent ask when walking into a problem area? What does a Facility Manager need before a client meeting? The goal is not just to verify that the tools work — it's to find out whether the answers Compass MCP gives are genuinely useful to the people who will use it.
5.2 Sample Queries by Tools
Use these as starting points. Feel free to modify them or try your own variations — that's the point of the test.
Site Image Explorer - Finding Photos
| # | Sample Questions | What to Check |
|---|---|---|
| 1 | Show me locations that need cleanup. | Are the returned photos actually showing clutter or debris? Do the View Location links work? |
| 2 | Are there any safety hazards visible on site? | Does Claude find relevant hazards (blocked paths, unsecured materials, missing guardrails)? |
| 3 | Find images showing HVAC duct installation. | Are the photos relevant to HVAC work? Note if the first result is clearly off-topic. |
Project Data Explorer - Capture History & Records
| # | Sample Questions | What to Check |
|---|---|---|
| 1 | Show me all capture dates and which levels were captured on each date. | Does the result match the known capture dates for this project? Is it easy to read? |
| 2 | Which level was not captured on the most recent capture date? | Does Claude correctly identify the most recent date and find the missing level? |
| 3 | How many captures were taken per level overall? | Are the totals plausible? Does Claude explain its answer or just show raw numbers? |
SiteInsights Analyzer - Progress Analysis
| # | Sample Questions | What to Check |
|---|---|---|
| 1 | Show me the current SiteInsights progress status. | Does the overall % and category breakdown look consistent with what you'd see in the dashboard? |
| 2 | Which categories are most behind schedule? | Are the delay figures specific and plausible? Does Claude suggest what to do about it? |
| 3 | Analyze the progress on HVAC ducts. | Is the HVAC-specific breakdown accurate? Does Claude surface workarea-level detail? |
Trade Visibility Analyzer - Trade Detection
| # | Sample Questions | What to Check |
|---|---|---|
| 1 | Give me an overview of trade visibility across all work areas. | Is the breakdown by trade and work area clear and structured? |
| 2 | Which trades have the lowest detection rates? | Does Claude rank trades usefully and flag the ones needing attention? |
| 3 | Which work areas on Level 4 are performing best? | Does Claude filter correctly by level and rank work areas meaningfully? |
Multi-Tool Queries - Combining Data Sources
These queries require Claude to pull from more than one tool. They're the best test of how well Compass AI synthesizes information.
| # | Sample Questions | What to Check |
|---|---|---|
| 1 | For the areas with the lowest SiteInsights progress, find photos showing what they look like on site. | Does Claude use both SiteInsights data and photo search? Does the combined answer make sense? |
| 2 | Give me a full project status summary — capture history, progress, and any visible site issues. | Does Claude pull from all relevant tools? Is the summary useful as a weekly status report? |
6. Evaluating Results
6.1 What to Look for
| Question to Ask | What a Good Result Looks Like |
|---|---|
| Is the answer accurate? | Numbers and facts match what you'd see directly in CupixWorks |
| Is it easy to understand? | A project manager with no CupixWorks training could act on the response |
| Are photo results relevant? | The first image link matches the intent of the query |
| Is it actionable? | The response helps the user decide what to do next — not just raw data |
| Did it answer the question? | Claude addressed what was asked, rather than returning a generic or off-topic response |
6.2 Common Issues to Watch For
- Response has no data — usually means the session wasn't initialized. Start a new session with the initialize prompt.
- Photo results are clearly off-topic — especially for abstract or broad queries. Flag the query and the result.
- Numbers don't match the dashboard — note the specific discrepancy and which query produced it.
- Claude answers from general knowledge instead of project data — means no tool was called. Flag this.
-
Response is technically correct but hard to act on — flag for UX/prompt improvement.
6.3 Recording your questions & answers
- At the end of the test sessions, export all of your questions and answer to .md files by following this prompt
- Prompt [Export this conversation to .md files for download]
- Submit with your feedback through Salesforce
7. Selecting Pilot Customers
After internal testing wraps up, the goal is to identify 10 external pilot customers by the end of March 2026. (TBD) Use the criteria below when reviewing your accounts.
7.1 What to Look for
| Criteria | Ideal Profile |
|---|---|
| SiteInsights Active | At least one active project with SiteInsights Pro running and meaningful progress data (SiteInsights not required. Customers who are willing to test SI Lite can be included.) |
| Regular Captures | Monthly or more frequent site captures — enough image data for photo search to be useful |
| Project Scale | Multiple levels and workareas — small single-level projects won't showcase the value well |
| AI Agent Access | Customer has or is open to getting a Claude Pro or Team plan for the pilot or any other AI Agents. |
| Strong CSM Relationship | Responsive account, willing to give structured feedback over a few weeks |
| Clear Pain Point | An identifiable use case — e.g. weekly progress reporting, subcontractor tracking, safety walk-downs |
| Pilot Mindset | Comfortable with early-stage tools and understands that some results may be imperfect |
7.2 Accounts to Avoid
- Infrequent or irregular captures — photo search results will be sparse
- No path to Claude access — pilot cannot proceed without a Claude plan
-
Accounts in active churn risk or renewal negotiation — don't add beta complexity
7.3 How to Submit a Candidate
- To be determined
8. FAQ
| Question | Answer |
|---|---|
| Can I use this with a real customer project right now? | Not yet. This test is scoped to the Madison Elementary School demo project only. Do not connect Compass AI to live customer projects without confirmation from the product team. |
| Claude didn't use any tools — it just answered from general knowledge. What's wrong? | The session wasn't initialized with the test project. Start a new conversation and paste the initialize prompt at the top. |
| The SiteInsights Analyzer returned no data. | Confirm your session is initialized with facility ID cjj60d. If it is and there's still no data, flag it in the feedback channel. |
| Image links in the response aren't opening. | Make sure you're logged into CupixWorks in the same browser. SiteView links require an active CupixWorks session. |
| How do I reset and start a fresh session? | Start a new Claude conversation. Each conversation is independent — use the initialize prompt at the start of each one. |
| What Claude plan do external pilot customers need? | Claude Pro or Claude Team. This will be confirmed per account before the pilot starts. |
9. Questions and Feedback
| Topic | Where to Go |
|---|---|
| Setup or connection issues | Salesforce |
| Questions about test results or data accuracy | Salesforce |
| Submitting pilot customer candidates | TBD |
| General feedback, bugs, or unexpected behavior | Salesforce |
|
Internal Use Only Do not share this guide or Compass MCP access with customers or external contacts. All external communications about Compass MCP require approval from the product team before distribution. |