Ad platforms
Shipped“Which campaigns are up week over week, and is the extra spend paying back?”
Shipped: Meta Ads and Google Ads, eighteen tools.
Custom Claude MCP connectors: ask your business systems a question, get real numbers back.
Any system with an API can become a question-and-answer tool inside Claude. We build the connector, the tool catalogue behind it, and the guardrails that keep it read-only and scoped to your accounts. The shipped example on this page is a connector for Meta Ads and Google Ads. The same pattern applies to your CRM, billing, support desk or warehouse.
Works overAd platformsCRMsBilling and financeSupport desksWarehouses and BIInternal APIs
How it works
How a custom MCP connector answers a question.
A question typed in Claude is routed to a connector deployed for your organisation. The connector calls the system through OAuth-scoped, read-only APIs and returns the numbers with the tools and date range cited. The diagram shows the shipped ads example.
You ask Claude
Desktop or Cowork
Custom MCP connector
Deployed per organisation, tokens encrypted at rest
Read-onlyOAuthMCP toolsComposes the right tool calls live, returns sourced numbers.
Meta Ads API
Business Manager, OAuth scoped
Google Ads API
Customer ID scoped per query
sourced answer back in Claude
01
Add a custom connector
Your team pastes a single URL into Claude's "Add custom connector" dialog, on Desktop or Cowork. Nothing to install on anyone's machine.
02
Authorise the system via OAuth
A standard OAuth handshake with the connected system, the same flow as logging into any third-party tool. In the shipped example that is Meta Business Manager and Google Ads. Tokens are encrypted at rest with per-tenant keys.
03
Read-only by default
Claude can read everything the system exposes but cannot write to it. Write actions, when a team wants them, ship separately behind an approval step. Analyst access is safe to share.
04
Isolated per organisation
No shared infrastructure across customers. Token refresh, deauth handling and rate-limiting are built in. Every team member with the connector installed gets the full tool catalogue.
The pattern
What a custom MCP connector is, and why it beats a dashboard.
A dashboard answers the questions someone predicted last quarter. A chatbot over your data guesses. A custom MCP connector gives Claude a catalogue of typed, read-only tools over the real API, so every answer is a live call with the numbers, the date range and the tools used cited back.
A dashboard, or a generic chatbot
- Answers only the questions it was built for. Anything new is a ticket to the data team.
- Numbers are as fresh as the last export or the last screenshot.
- A general chatbot with database access will guess when it cannot find the answer.
- Sharing access means sharing the whole tool, including the parts that can change things.
A custom MCP connector inside Claude
- Claude composes the right tool calls live, so free-form questions get real answers.
- Every number comes from the API at the moment you ask, in the account's own currency.
- Each answer cites the tools and date range used, so it can be reproduced and checked.
- Read-only by default and scoped by OAuth, so analyst access is safe to hand out.
Where it applies
Systems your team could be asking questions of.
If a team opens the same tool every week to pull the same numbers, a connector replaces that ritual. These are the systems we see most often, with the kind of question each one answers.
CRM and pipeline
“Which deals over fifty thousand have had no activity in two weeks, and who owns them?”
HubSpot, Salesforce, Pipedrive.
Billing and finance
“What is our net revenue retention this quarter, and which accounts drove the change?”
Stripe, Chargebee, Xero, QuickBooks.
Support desks
“Which product areas generated the most tickets this month, and how did first-response time move?”
Zendesk, Intercom, Freshdesk.
Warehouses and BI
“Show conversion by channel for the last four weeks against the same period last year.”
Snowflake, BigQuery, Databricks, Postgres.
Internal APIs
“How many orders are stuck in the fulfilment queue right now, grouped by warehouse?”
Anything with an authenticated HTTP API.
The shipped example
What the Meta Ads and Google Ads connector does.
Eighteen read-only tools over both platforms, packaged as slash commands for the recurring shapes and open to free-form questions for everything else. This is the level of depth a connector reaches once the questions are mapped properly.
01/weekly-report
Replace the Monday reporting ritual
Claude pulls last week's data across Meta and Google in about 30 seconds, compares it to the prior week, and writes you a slide-ready summary in your account currency.
- TL;DR with blended ROAS and week-over-week change
- Top five campaigns by spend, mixed Meta + Google
- Two to four prioritised recommendations grounded in actual numbers
02/creative-fatigue
Catch creative fatigue before it costs you
Scans your Meta ads for the high-frequency, declining-CTR pattern and surfaces the worst offenders. The bundled copy-agent then drafts three to five replacement variants per fatigued ad, matched to your brand voice.
- Fatigued ads ranked by wasted-spend estimate
- Original copy plus three to five fresh variants per ad
- Short "why this works" rationale for each variant
03Free-form question
Find wasted Google search-term spend
Just ask: "Show me search terms with cost over $10 and zero conversions in the last 14 days." Claude pulls the search-term report, filters for waste, and ranks the result so you can paste straight into Google Ads as negative keywords.
- Ranked list with cost, clicks, and conversion zero-state
- Audits routinely surface three-figure monthly waste in a single run
- Output is copy-paste ready for Google Ads negative-keyword bulk upload
04/anomaly-scan
Investigate anomalies without scanning charts
Daily-level statistical outlier detection across spend, ROAS, CTR, and frequency over the last thirty days, correlated with your account change history, so you don't just see "this date was weird," you see why.
- Severity-rated dates (low / medium / high) with the metric that moved
- Correlated changes within ±1 day of each anomaly
- A one-line suggested next step per finding
05Free-form question
Answer questions your dashboards can't
For the questions that don't fit a pre-built report, Claude composes the right tool calls live and returns the actual numbers in your currency, not a chatbot guess, not last-month's screenshot.
- Natural-language to live API call to numbers, in seconds
- Account-currency output, including non-USD multi-brand rollups
- Cites which tools and date range were used so the answer is reproducible
06Multi-account scope
Multi-account, no platform-hopping
Run multiple Meta ad accounts or Google customers, or manage several brands as an agency? Claude scopes each query to the right account on demand. No tab-switching, no CSV exports, no UI loading.
- Per-query account selection, no context-switching
- Same eighteen-tool catalogue across every brand or client
- Read-only by default, so analyst access is safe to share
A connector answer
What a connector answer looks like.
A representative /weekly-reportoutput from the ads connector. The structure and depth are what the connector produces; the numbers and names are illustrative, not a client's results.
/weekly-reportSend the Monday update.
returned in about 30 secondsSpend up 22% week-over-week ($18,400 → $22,400) but blended ROAS dropped from 3.1x to 2.6x, the extra spend isn't paying back.
- 1Pause Meta campaign "Spring Promo Q2"; $420 spent, 0 purchases, frequency 5.8 (likely fatigued).
- 2Reduce daily budget on Google "Brand-Defense" by 30%; search impression share is 98%, the extra spend is just showing the same audience.
Tools Claude called through the MCP server
Get Meta Account OverviewGet Google Account OverviewCompare PeriodsOr just ask
Free-form questions the ads connector answers.
Slash commands cover the recurring shapes. For everything else, Claude composes the right tool calls live and returns the actual numbers, with the date range and tools used cited.
"Which Meta audience had the lowest CAC last month?""How does our Tuesday spend compare to Monday across the last 4 weeks?""What's our best-performing Shopping product this quarter, by brand?""Show me search terms triggering on our brand from people who never bought.""Which creative format is winning on Meta Reels this month?""How much have we spent on prospecting vs retargeting across both platforms?"
How we build one
How a connector build runs, from the API to your team's Claude.
A proof of concept against your real system typically lands in two to four weeks. The shape is the same for every system: map the questions, build the tools, harden the guardrails, hand it over.
- 01
Map the questions
Week 1
What happens
We sit with the people who ask, list the recurring questions, and read the API to confirm each one can be answered live and cited.
What you get
Question list, API coverage map, tool catalogue design.
- 02
Build the tools
Weeks 2–3
What happens
The MCP server with typed, read-only tools over the API, OAuth with per-tenant encrypted tokens, rate limiting and deauth handling.
What you get
Working connector, first free-form answers with real numbers.
- 03
Commands and evals
Weeks 3–4
What happens
Recurring shapes become slash commands. An evaluation set of real questions checks every answer cites its tools and matches the platform's own numbers.
What you get
Slash commands, evaluation set and results, multi-account scoping.
- 04
Rollout and handover
Week 4 onward
What happens
Install instructions for the team, a runbook for adding a tool or a command, and the repository handed over.
What you get
Install guide, runbook, repository, thirty-day support window.
What you get
What is in place when we hand over.
The server, the tools, the commands and the evaluation set, all documented. These are the parts, using the shipped ads connector as the reference.
01
The MCP server
Deployed per organisation with encrypted tokens, token refresh, deauth handling and rate limiting built in.
02
A typed tool catalogue
Read-only tools over the system's API, each with a clear contract so Claude composes the right calls. Eighteen in the ads example.
03
Slash commands
The recurring shapes packaged as commands, such as a weekly report, a fatigue scan or an anomaly scan, each returning a structured answer.
04
An evaluation set
Real questions with expected answers, run before every change, so the connector is measured against the platform's own numbers.
05
Multi-account scoping
Per-query selection of the account, customer or brand, so agencies and multi-region teams use one connector.
06
Install guide, runbook and source
One URL for the team, a runbook for adding tools and commands, and the repository in your organisation.
Custom MCP connectors, answered.
A custom MCP connector is a small server that gives Claude a catalogue of typed tools over one of your business systems, so a person can ask a question inside Claude and get a live, cited answer from the real API. The shipped example on this page connects Meta Ads and Google Ads: eighteen read-only tools, slash commands for weekly reports, creative fatigue, wasted spend and anomalies, and free-form questions answered in the account's own currency. The same pattern applies to a CRM, a billing system, a support desk or a warehouse.
Anything with an authenticated API: ad platforms such as Meta, Google, TikTok, LinkedIn or Amazon Ads; CRMs such as HubSpot, Salesforce or Pipedrive; billing and finance tools such as Stripe, Chargebee, Xero or QuickBooks; support desks such as Zendesk, Intercom or Freshdesk; warehouses such as Snowflake, BigQuery, Databricks or Postgres; and internal HTTP APIs. Discovery starts by reading the API to confirm each recurring question can be answered live.
It is deployed per organisation with no shared infrastructure across customers. OAuth tokens are encrypted at rest with per-tenant keys. Access is read-only. Token refresh, deauth handling and rate limiting are built in. Nothing is installed on anyone's machine beyond Claude itself.
Not by default. The connector is strictly read-only, so Claude can analyse in detail but cannot write to the system. In the ads example that means no changes to campaigns, budgets or creatives. When a team wants write actions, they ship separately behind an approval step and an extended audit trail, the same way we build every internal tool.
Every answer cites the tools and the date range used, so it can be reproduced against the platform's own reporting. During the build we assemble an evaluation set of real questions with expected answers and run it before every change. If Claude cannot answer from the tools available, it says so rather than guessing.
A proof of concept against your real system typically lands in two to four weeks: one week to map the questions and the API, then the tools, then the commands and evaluation set. Rollout to the team is a single URL and an OAuth authorisation, about fifteen minutes per person. Systems with a well-documented API sit at the shorter end.
Yes. The connector scopes each query to a specific account, customer or brand on demand. In the ads example that means several Meta ad accounts or Google customers, which is how agencies and multi-region teams use one connector without exporting anything or switching platforms.
Yes. Numbers are returned in the account's actual currency rather than assumed USD, so a portfolio with INR, EUR and GBP accounts gets correctly denominated outputs per query.
The MCP server, the tool catalogue, the slash commands, the evaluation set, the install guide and the runbook, handed over with the repository. Your team can add a tool or a command without us. There is no per-query fee to Sarvaswa and no shared infrastructure to depend on.
Admin access to the system you want connected, a Claude Desktop or Cowork account for the people who will ask, and a list of the questions they ask most. If the list does not exist yet, the first week of discovery produces it.
Connectors are one piece of how we ship Claude in production. For the agent patterns behind them see AI agent development, or another internal tool with a person in the loop in the Slack data copilot.