The Improvado MCP server is a hosted Model Context Protocol endpoint that lets an AI agent such as Claude, ChatGPT or Cursor read, write and monitor your business systems through one governed connection. Its MCP index lists 187 entries, 186 of them named systems such as Salesforce, Google Ads, Gong and Jira, and the connector pages say credentials stay in Improvado's encrypted vault.
That is the short answer. The rest of this guide shows which systems are on the list, what the read, write and monitor scopes mean on each page, how authentication works, and which jobs teams can hand to an agent on day one. Every figure comes from the product pages themselves. Book a demo if you would rather see it run on your own stack.
What is the Improvado MCP server?
Start with the protocol. The Model Context Protocol site defines MCP as "an open-source standard for connecting AI applications to external systems." It compares MCP to a USB-C port: one standard plug between an AI application and the data sources, tools and workflows it needs.
An MCP server is the other end of that plug. It sits in front of a system and tells the agent what it can ask for and what it can do. Improvado runs that server for you. The connector pages describe it as hosted and pre-authenticated, so nobody installs Python, manages a local process or pastes keys into an agent's config file.
The design choice behind it is one gateway rather than one server per platform. Improvado's write-up of that architecture, one MCP gateway in front of everything, describes one OAuth grant per client and platform credentials that never leave the server side. For the protocol background and the wider market, see MCP servers for marketing analytics.
Which systems can an AI agent reach through it?
The count lives in the product, so here it is exactly as the files support it. The mcp content collection holds 187 entries, one page per system. One of them, named Example, is a prototyping page for testing an API connection, which leaves 186 named systems. The /mcp page headlines the larger number, "187+", because it counts every entry in the collection.
The list is not only marketing platforms. Alongside Google Ads, Facebook Ads and Google Analytics 4 it covers CRM and sales tools (Salesforce, HubSpot, Gong), meeting and project tools (Fireflies, Atlassian Jira, GitHub) and databases (ClickHouse, Google BigQuery). An MCP server is one route to these systems; for the other route, see our comparison of API integration platforms. The table shows ten of them with the example prompts each page publishes and the credential each page names.
| System | Example read prompt from its page | Example write prompt from its page | Credential the page names |
|---|---|---|---|
| Salesforce | Current pipeline value by stage, and which deals are at risk | Log a call and set a follow-up task | Salesforce OAuth |
| HubSpot | Email open and click rates for nurture campaigns | Enroll recent MQLs in a sales sequence | HubSpot OAuth |
| Google Ads | ROAS by campaign type this week | Create a Search campaign with a monthly budget | Google OAuth |
| Google Analytics 4 | Organic conversion rate by landing page | Create a remarketing audience | Google OAuth |
| Facebook Ads | ROAS by campaign objective across accounts | Which ad sets to pause based on CPM and purchase data | Meta API token, scoped |
| Gong | Deals where a competitor was mentioned in 30 days | Summarize calls and push a deal update to Salesforce | Gong API credentials |
| Fireflies | What was decided about pricing across sales calls | Draft a follow-up email from call action items | Fireflies API key |
| Atlassian Jira | High-priority issues stuck In Progress for 5+ days | Move unestimated issues to the backlog | Jira API token |
| GitHub | Open pull requests waiting for review over 3 days | Create an issue for a timeout error | Personal access token, scoped |
| ClickHouse | Daily active users by region from an events table | Create a materialized view of daily revenue | Database credentials, can be read-only |
Source for every row: the matching file in the site's MCP collection. The prompts are examples, not a limit.
What do read, write and monitor mean on a connector page?
All 187 pages are built the same way, with a Read section, a Write section and a Monitor section. That structure is the clearest statement of what the server is for.
- Read is a question in plain language, answered from the system's own data. The Salesforce page translates natural language into SOQL, Salesforce's query language, so the person asking never writes it.
- Write is an action in the system, such as logging an activity, changing a deal stage, pausing an ad set or opening an issue.
- Monitor is a standing watch. The pages describe alerts for things like stale deals, a conversion drop, a CTR drop or a slow query, raised before someone notices on a dashboard.
One caveat keeps this honest. The pages describe what each scope can do and give example prompts. What an agent is actually allowed to do in your account depends on the credential you connect, which is the next question.
How does an agent authenticate, and who sets the limits?
Each connector page names the credential type. In the table above you can see OAuth for Salesforce, HubSpot, Google Ads and Google Analytics 4, an API key for Fireflies, and tokens for Jira, GitHub and Meta. Improvado's docs page for the Jira connection also lists a Rovo MCP token, issued by Atlassian Rovo, as an alternative to an API token.
The limits follow the credential. Salesforce's page says permissions are controlled by the Salesforce user credentials you authenticate with and that field-level security from your org is respected. Gong's page says the same of its permission scopes: if a rep's calls are restricted to managers in Gong, the MCP integration enforces the same restriction. GitHub's page ties read and write rights to the token scope.
On storage, the Salesforce, Google Ads and HubSpot pages state that OAuth tokens sit in an encrypted vault certified to SOC 2 Type II, that the agent never touches the credentials directly, and that requests go through Improvado's proxy. Write access can also be gated. The ClickHouse page says write operations need explicit user confirmation, suggests starting with read-only credentials, and notes that executed statements show up in the ClickHouse query log with the MCP session identifier.
What can teams actually do with it?
The pages are written around concrete jobs. These six come straight from their Read, Write and Monitor prompts, and several combine two systems in one question.
- Pipeline risk review. Ask Salesforce for deals with no recent activity and Gong for deals where a competitor came up. The Gong page also lists a Monitor job for deal risk signals such as no next step set or a decision-maker who went silent.
- CRM hygiene after a call. The Salesforce page's write example turns one paragraph of notes into a logged call, an updated opportunity stage and a follow-up task, with no record navigation.
- Paid channel watch. Read ROAS by campaign type in Google Ads, then use the Monitor scope for anomalies that "eat your budget" before the weekly report. The Facebook Ads page does the same for CPM spikes and frequency fatigue.
- Organic traffic triage. The Google Search Console page asks which pages lost the most impressions against the prior period. Its write example submits a list of new URLs for indexing.
- Meeting memory. Search Fireflies transcripts for what was decided about pricing, then draft the follow-up email from the action items.
- Engineering and warehouse questions. Jira sprint health, GitHub pull requests waiting for review, ClickHouse and BigQuery queries from a plain-language request, so analysts and non-analysts hit the same data.
The pages also publish before-and-after times for the read prompts, for example "45 min to 1 min" on the Salesforce page's pipeline question. Those are Improvado's own estimates, not independent measurements, so test them on your data before you quote them.
For the full list of systems your own agent could reach next, browse the MCP index.
Which AI tools connect, and what does setup look like?
The /mcp page says you can connect any source to "Claude, ChatGPT, Cursor, or any MCP client." Improvado also runs dedicated pages for Claude, for OpenAI Codex and for custom agents you build yourself. The Claude page frames setup as a one-time admin step with nothing for the team to install.
Connector pages quote short setup times. Salesforce says under 2 minutes, and GitHub and Fireflies say under 5 minutes. The steps are the same shape each time: authenticate the system, choose what to include, and add the server to your AI client.
When is a hosted MCP server not enough?
Connecting an agent is the easy half. A single question that spans Salesforce, Google Ads and GA4 is only as good as how those systems name campaigns and define a conversion. If the same campaign has three names, the agent will report three campaigns, fluently. Improvado's guide to AI agent workflows on messy marketing data covers that failure mode in detail.
The other limit is scope. If you only need one system and your team is happy to run a local server, a single-platform MCP server may be enough. A hosted gateway earns its place when several people, several accounts and several systems are involved at once. If you are weighing it, talk to an expert about your stack.
Frequently asked questions
Can the Improvado MCP server write to my systems or only read?
Both. Every connector page has Read, Write and Monitor sections. What the agent may change depends on the credential you connect, such as the Salesforce user, the GitHub token scope or the Meta token scope. ClickHouse writes also require explicit user confirmation.
Which AI clients can use it?
Improvado's MCP page names Claude, ChatGPT and Cursor, and says any MCP client works. Because MCP is an open standard, a client that speaks the protocol can connect. Dedicated pages cover Claude, OpenAI Codex and custom agents.
Does the AI agent see my passwords or tokens?
No, per the connector pages. The Salesforce page says OAuth tokens are stored in an encrypted vault certified to SOC 2 Type II, calls are proxied through Improvado, and the agent never accesses credentials directly. Field-level security in your own org still applies.
How many systems does it cover?
The MCP collection has 187 entries: 186 named systems plus one prototyping page called Example. Improvado's /mcp page headlines the figure as 187+. Check the live index for the current list, because new systems are added as pages.
