Flag At-Risk Deals: An Agent That Reads Open Pipeline Across CRM and Warehouse

Your best reps already know which deals are slipping. The hard part is that the signal lives in two systems at once: the CRM that tracks the deal, and the warehouse that knows whether the customer is actually healthy. This tutorial builds one small, governed AI tool that reads both and flags the deals worth a call today.

In about fifteen minutes with Nexla MCP Studio, you will stand up a single MCP server, Pipeline Risk Sentinel, that any agent (Claude, ChatGPT, Gemini, or your own) can call to read open pipeline and surface at-risk deals. It is read-only by design: it surfaces risk, it never edits your CRM. We use Salesforce as the CRM here; HubSpot connects the same way.

The one idea that makes this work: purpose shapes the tools

Most MCP setups hand an agent a whole database and hope it picks the right table. MCP Studio works the other way around. You describe the job in plain language, and it builds only the few scoped tools that job needs. “Read open pipeline and flag at-risk deals” becomes five read-only tools, not direct access to hundreds of warehouse tables. Fewer tools means the agent chooses the right one more often, and there is far less it can get wrong.

What you will build

One governed server, two sources, five tools. The CRM supplies the deal; the warehouse supplies the truth about the account behind it.

What each source provides

Salesforce Open opportunities: stage, amount, owner, close date, last activity date, stage-entry timestamps, and engagement.

Snowflake Account health: product usage signals, invoice and payment status, support ticket volume, and historical win/loss outcomes.

Orientation: MCP Studio is split into a conversation pane on the left, where forms appear inline as you go, and a working area on the right with four tabs, Context, Tools, Explorer, and Build, plus a persistent Build MCP Server button.

1Describe the server

The opening prompt does most of the work. Name the server, name both sources, and state the outcome you want in plain language. The clearer the purpose and the systems, the better MCP Studio scopes the tools to match. Copy this prompt and adjust the wording to your data:

Notice what the prompt pins down: the scope (read-only, surfacing not editing), the two systems so the probe knows where to look, and the outcome you want, so Studio can shape the right tools. That single message drives every step that follows.

2Select and validate credentials

MCP Studio loads the connector instructions and detects credentials you already have. An inline form, Select Credentials for Pipeline Risk Sentinel, appears in the conversation:

Choose your CRM, Salesforce or HubSpot. Pick Salesforce, then select your Salesforce credential. Select your Snowflake credential. Click Submit Context. Studio validates each one and confirms Credential validated for both. If you pick HubSpot instead, you get a Connect OAuth step in place of the Salesforce picker.

3Discover the data

Studio opens a Credential Explorer for each source and runs an agentic probe against them. When it finishes you will see Probe complete: 5 candidate tools ready. You do not map columns by hand: the probe scans the connected systems and proposes tools matched to the purpose you described.

The Explorer tab lets you browse what it found. For Snowflake it opens a Database Explorer over your warehouse with a searchable table list and a schema view you can inspect or download, so you can confirm the probe is reading the right objects before you build.

4Review and scope the tools

Open the Tools tab. The five proposed tools are each mapped to a source and backed by an editable query macro you can adjust at runtime. Everything is read-only. Select or deselect tools, and edit the default query for any of them, here.

5Inspect a tool, where scope and safety get enforced

Open Inspect on any tool, for example list_open_pipeline, and you will see four parts. This is the governance layer, and it is worth understanding before you build.

Two things make this safe for production. Read-only is enforced at the connector, so even a confused agent cannot write to your CRM. And you decide which columns an agent is allowed to filter on, so the tool stays inside the lane you intend. Flip Requires approval on for anything you want a human to confirm before it runs.

6Build and register the server

Click Build MCP Server. The Build tab runs three steps in order, then registers and enables the server and connects it live to your session.

7Connect it to your agent

Open Integration config in the Build tab. There are two ways to connect, and which you use depends on the client.

For a desktop client, the Service Key block is the fastest path. Paste it into your client config and replace the two bracketed values with your server id and your Nexla API key:

Get your key from the profile menu at the bottom-left, Get Nexla API key. The Google and Azure OAuth blocks follow the same shape if your client authenticates that way instead.

8Test it

Connect the server and ask your agent the questions a RevOps lead actually asks. If the answers look right, the tools are wired correctly.

That is the whole loop. You described a job, MCP Studio scoped it into five governed tools across two systems, and now any MCP client can read your pipeline and flag the deals worth a call, without write access to anything. When the next quarter-end scramble starts, the agent is already watching.