Turn AI Assistants Into Product Intelligence Partners with MCP
You're probably using AI wrong.
Not wrong in the "you'll get fired" sense. Wrong in the "why are you manually copy-pasting context into ChatGPT like it's 2022" sense.
Product managers waste hours extracting information from Jira, Linear, GitHub, and Notion just to ask an AI assistant basic questions. "What features do we have around authentication?" "Which parts of our codebase have the most technical debt?" "What does Stripe have that we don't?"
You screenshot things. You paste tickets. You describe your product like you're explaining it to someone who's never seen it. The AI forgets everything the moment you close the tab.
There's a better way. It's called Model Context Protocol (MCP), and it lets AI assistants directly access structured data about your product. Not through web scraping or API wrappers. Direct server-to-AI connection.
MCP is Anthropic's open protocol for connecting AI assistants to external data sources. Think of it like USB-C but for AI context.
Before MCP, every AI tool needed custom integrations. Cursor had its own way of reading codebases. Copilot had another. Claude Desktop needed yet another approach. Building integrations meant writing bespoke code for each tool, each API, each data source.
MCP standardizes this. You build one MCP server that exposes your data. Any MCP-compatible AI assistant can connect to it. Claude Desktop, Cursor, custom agents — they all speak the same protocol.
Here's what makes it different from just "hitting an API":
Structured context, not raw data. MCP servers don't dump JSON at the AI. They expose tools (functions the AI can call), resources (data the AI can read), and prompts (pre-built workflows). The AI assistant decides what it needs and requests it.
Persistent sessions. The connection stays open. The AI can make multiple requests, build on previous answers, and maintain context across a conversation.
Type safety and validation. MCP uses JSON schemas. The AI knows what parameters each tool expects and what it'll get back. No more "parse this blob of markdown and hope."
This matters for product intelligence because your product data is complex. Features don't live in one place. They're scattered across code, documentation, tickets, and tribal knowledge. MCP lets you centralize that intelligence and make it queryable.
Why Product Intelligence Needs MCP
Product managers deal with three kinds of intelligence:
What you built. The features that actually exist in your codebase. Not the roadmap. Not what someone remembers from three quarters ago. What's actually shipped and working.
How healthy it is. Which parts of your product are technical debt disasters? Where's the code that breaks every time someone touches it? What features are orphaned with no active maintainer?
What you're missing. The competitive gap analysis. What do your competitors have that you don't? Where are you behind? Where are you ahead but nobody knows it?
Traditional product tools suck at all three.
Jira tells you what tickets closed. It doesn't tell you if the feature actually works or if it's a maintenance nightmare. GitHub shows commits. It doesn't explain what features those commits built or how they connect to user needs. Notion has meeting notes that nobody reads.
You end up with product managers who can't answer basic questions:
"Do we have SSO?" (Check five repos, three Slack threads, and ask engineering)
"Which features are breaking the most?" (Pull metrics from four tools and correlate manually)
"What payments features does Stripe have that we don't?" (Open 47 browser tabs)
MCP fixes this by letting AI assistants directly query structured product intelligence. Not searching Confluence. Not reading meeting notes. Actually querying a system that knows your codebase.
What Glue's MCP Server Exposes
Glue indexes your entire codebase — every file, function, API route, database schema, and feature. It uses AI agents to discover features automatically by analyzing code patterns, not by reading stale documentation.
With Glue's MCP server, you connect Claude, Cursor, or any MCP-compatible assistant to that intelligence. You can query:
Feature catalog. "List all authentication features." The AI gets back structured data: which repos, which files, which functions, when they were last changed, who maintains them.
Code health metrics. "Which features have the highest code churn?" The AI queries actual churn data — how often files change, how complex the code is, how many people touch it.
Competitive gaps. "What does our product have compared to Stripe's payment features?" The AI accesses gap analysis that maps your features against competitors.
Team knowledge. "Who knows the billing code?" The AI sees ownership data, contribution history, and knowledge concentration risks.
Here's what this looks like in practice.
Real Example: Feature Discovery
You're a PM. New quarter. Your CEO asks: "Do we have granular permissions at the org level or just workspace level?"
Traditional approach:
Ask in Slack. Wait for responses. Get three different answers.
Check Notion. Find a doc from 2022 that might be outdated.
Bug engineering. They check the code. Come back in two hours.
Still not sure if it's shipped or just planned.
With Glue + MCP:
You open Claude Desktop. Connected to Glue's MCP server. You ask:
"What permission features do we have? Break down by org-level vs workspace-level."
Claude calls get_features with filters for "permissions". Gets back:
OrganizationRoleManager in /backend/auth/org_roles.ts
WorkspacePermissions in /backend/auth/workspace_perms.ts
enforce_org_policy() function that checks org-level rules
Last changed: 3 weeks ago
Maintainer: Sarah's team
Claude responds:
"You have both. Org-level permissions are in OrganizationRoleManager and control who can create workspaces, invite members, and set billing. Workspace-level permissions are separate and control access within specific workspaces. The org-level system was added 3 months ago by Sarah's team."
Total time: 8 seconds.
The AI didn't search docs. It queried your actual codebase structure.
Real Example: Technical Debt Prioritization
You're planning the next quarter. Engineering keeps saying "we need to pay down technical debt." You need to know which debt actually matters.
Traditional approach:
Ask engineering for a list. Get 23 items with no priority.
Try to correlate with bug reports. Everything's on fire apparently.
Guess based on vibes and who complains loudest.
With Glue + MCP:
You ask Claude:
"What are the top 5 features with the worst code health? Include churn, complexity, and ownership risks."
Claude calls get_code_health sorted by a composite score. Gets back:
PaymentProcessing: High churn (changed 47 times in 3 months), high complexity (cyclomatic complexity 85), single owner (John, who's leaving).
NotificationService: Medium-high churn, orphaned (no clear owner), breaks other features.
ReportGenerator: High complexity, low test coverage, blocks new BI features.
Claude responds with a prioritized list and reasoning:
"PaymentProcessing is your biggest risk. It's changed constantly, extremely complex, and John (sole maintainer) is leaving next month. NotificationService is second because it breaks frequently and nobody owns it. ReportGenerator blocks your BI roadmap but is otherwise stable."
Now you can actually prioritize technical debt that matters for the business.
Real Example: Competitive Gap Analysis
Your competitor ships a new feature. Your CEO forwards the announcement. "Do we have this?"
Traditional approach:
Read their marketing copy. Guess what it actually does.
Ask around. Nobody's sure.
Check the codebase. Maybe? Different terminology.
Shrug and say "we'll look into it."
With Glue + MCP:
You ask Claude:
"Compare our authentication features to Auth0's feature list. What do they have that we don't?"
Claude calls get_competitive_gaps with Auth0's features. Gets back:
You have: Email/password, OAuth, SSO, MFA
They have: Passwordless, social login, enterprise connections, attack protection
Your gaps: No passwordless, no built-in social providers (you do OAuth manually), no attack protection dashboard
Claude responds:
"You're missing three things: passwordless auth (they have magic links and biometrics), pre-built social login (you support OAuth but users configure it themselves), and attack protection (brute force detection, bot detection). Your SSO and MFA are comparable."
You know exactly what the gap is. Not guessing from marketing copy.
How to Start Using MCP
First, install an MCP-compatible client. Claude Desktop is the easiest. Download it, sign in, and you're ready.
Second, connect an MCP server. If you're using Glue, this is a config file change: