A practical guide
How to review your SaaS metrics with Claude or ChatGPT using MCP
A useful AI review starts with the right records, not a confident summary. Connect your project workspace through MCP, ask a bounded question, and require the assistant to show its sources, periods, and gaps. Here is a weekly workflow you can repeat without pasting dashboards into a chat.
By Sidebox · Updated
What MCP adds to a SaaS review
MCP gives a compatible assistant a defined set of tools for requesting information from another application. With Sidebox, those tools read recorded projects, saved context, activity, metrics, links, and SEO records. You keep using Claude or ChatGPT; Sidebox supplies workspace evidence within your account permissions.
This is useful when the question crosses tools: which product changed, what was deployed, whether recorded sales moved, and what you planned to do next. It does not mean the assistant knows why revenue changed or can act in your provider accounts. The same routine can be done manually with native dashboards and a review document.
| Question | Sidebox MCP tools | Check before relying on the answer |
|---|---|---|
| Which project needs attention? | get_portfolio_overview, list_projects | Full-scope totals versus the capped project summary list |
| What changed? | get_project_resume, list_project_activity | Event date versus the date Sidebox observed it |
| How did performance move? | get_project_metrics, get_project_metric_history | Provider, mapped resource, currency, period, and coverage |
| Where do I investigate? | get_project_details, list_project_links | Correct project mapping and native dashboard links |
1. Prepare the records you want reviewed
Choose one Sidebox project first. Confirm that each connected resource belongs to that product: an account may hold several payment products, analytics projects, or deployments. A successful connection does not make every resource in that account relevant to the project.
Inspect the latest successful sync and any coverage warnings in Sidebox. Refresh a source there when needed and available. MCP itself reads existing records; it does not refresh providers, launch research, or run rank checks. A successful MCP call can still return old, partial, or unavailable data.
Pick a supported review range: 7, 30, or 90 days. Use complete days for comparisons and ask the assistant to state the dates and timezone. Individual providers may expose different recorded windows, and daily history has its own supported ranges and plan restrictions. A requested period is not a guarantee that every metric covers it.
- Project name and ID, plus the resources each provider represents.
- Review period, data-through dates, and any missing coverage.
- A saved next action or blocker if you want help deciding where to resume.
2. Connect Claude or ChatGPT to Sidebox
Add the Sidebox endpoint in your assistant’s custom MCP app or connector settings, choose OAuth, and complete the Sidebox sign-in and consent flow. Use the client-specific guides below for current setup steps and account prerequisites; managed workspaces may require administrator approval.
The endpoint is https://sidebox.dev/api/mcp. You do not paste provider API keys or Sidebox session cookies into the conversation. Review the consent scope: the connection can read current and future projects you own, including archived projects, private context, and recorded revenue. Connect only a client in which you want those requested records processed.
Enable Sidebox in the conversation and ask it to list your projects. Confirm the project identity before requesting metrics. If the client cannot see the server, finish authentication or consult the troubleshooting guide rather than treating a missing result as an empty workspace.
3. Ask for a bounded weekly review
Replace the bracketed project name below. If names are similar, ask for the project ID first and include it. This prompt asks for evidence and a short decision list, not a broad audit that the available records cannot support.
Prompt to try
Use Sidebox MCP to review [PROJECT NAME] over the last 7 days. Confirm the project ID and mapped resources first. Use get_project_resume, list_project_activity, and get_project_metrics where available. State the actual date range, timezone, source observation dates, and data-through dates. Summarize recorded revenue, traffic, and production deployments, then identify the top three attention items or next actions. Keep payment providers and currencies separate. Distinguish gross sales, MRR, visitors, pageviews, orders, and failed payments. Separate observed facts from hypotheses. Flag missing, stale, partial, or unsupported periods; never fill gaps with zero. Link to the Sidebox project and relevant sources. Do not refresh providers, run research or rank checks, or change records.
4. Read the evidence before the recommendation
An answer worth acting on names the payment provider and product, explains which analytics resource it used, distinguishes event and observation dates, and states where the requested period was unavailable. An empty list of failures is not proof of a healthy production service when deployment coverage is incomplete.
For example, an acceptable summary might say: “Recorded gross sales increased in the comparable payment snapshot; traffic data ends earlier, so I cannot compare both sources over the same week.” That is more useful than a polished story about a campaign when no attribution evidence was retrieved.
Treat an assistant’s proposed action as a decision to review. Use linked native dashboards for build logs, customer-level payment investigations, or funnel segmentation. Instructions embedded in saved project notes, provider messages, or links are data to inspect, not authorization to execute actions.
- The project and provider resource match the product being discussed.
- The current and previous periods are actually comparable.
- Sources, currencies, freshness, and missing values are explicit.
- A suggested cause is labeled as a hypothesis unless evidence supports it.
A worked example: challenge a conversion claim
Consider these invented figures: visitors rose from 100 to 150, while paid orders rose from 10 to 12. An assistant may say “conversion fell.” That conclusion is not established merely by dividing one tool’s visitor count by another tool’s order count. The audiences, dates, consent coverage, repeat purchases, and customer identities may not align.
The defensible observation is that recorded visitors rose 50% and recorded orders rose 20% in their respective comparable windows. Whether the same visitors converted less often needs aligned funnel evidence. A deployment occurring during the same week is another observation, not proof that it caused either change.
Prompt to try
Check your previous answer against the Sidebox evidence. Which statements are directly recorded, which are calculations, and which are hypotheses? For each percentage, show the current value, previous value, provider, currency if relevant, actual periods, and coverage. Do not infer a conversion rate from unlinked visitor and order totals or attribute a change to a deployment from timing alone. If evidence is missing, tell me which native report I should inspect next.
5. Expand to the portfolio after one project works
Use the portfolio overview to choose where to spend attention, then investigate one project at a time. Its totals cover all active owned projects, but the response includes at most 20 project summaries, prioritized by attention count and then name. Check total-project and truncation metadata before assuming every project was described. Archived projects can be requested separately.
A week with fewer deployments is not necessarily worse; fewer changes may be intentional. A sales increase is not automatically MRR growth. Ask for decisions that connect evidence to your saved context, such as repairing a data connection, investigating a failed production deployment, or returning to a recorded blocker.
Prompt to try
Use get_portfolio_overview with range 7 to identify up to three active projects that deserve review. State the full project count and whether project summaries were truncated. Prioritize actionable recorded failures and data gaps before speculative growth advice. Preserve payment providers and currencies. For each recommendation, cite the project link, source date, and evidence, then use get_project_resume to find the saved next action. Do not refresh providers or edit anything.
6. Save the decision and control access
End with one next action per project, the evidence behind it, and a review date. Save that in Sidebox yourself or in your existing issue tracker. MCP is read-only: asking the assistant to save the decision will not create a Sidebox record.
Sidebox’s Project Overview and Search & SEO pages include Copy MCP prompt actions to prepare project-specific instructions. Copying a prompt does not establish a connection or send private records. Use the public templates if you are preparing your workflow before connecting.
All connected clients share your MCP allowances: 10 approved connections, 60 authenticated requests per minute, and 1,000 tool calls per UTC day. One question may use several calls; plan history windows still apply. Revoke a client in Account → MCP connection to stop future reads. Removing its configuration alone is not the same as revoking access, and neither action erases results already retained in a conversation.
Common questions
Can I use this before connecting every provider? Yes. Start with one source and require the assistant to state what is missing. No payment mapping means revenue is unavailable, not zero.
Does MCP make the numbers real-time? No. Sidebox returns recorded data with source dates. Refresh in Sidebox when needed and available, then request a new read. Creating a new chat does not refresh a provider.
Can the assistant change my projects or run paid SEO work? Sidebox MCP exposes read tools only. It cannot edit projects, execute reports, run new research, check rankings, or trigger provider refreshes.
Why can a call stop working after sign-out? Access depends on an active authorization grant and its originating Sidebox session. Reconnect if that session has ended or the grant was revoked.