Competitor Pulse

Real-time, AI-powered competitive intelligence: automatically capture and distill your competitors’ public signals into actionable digests delivered straight to Slack, email, and Notion using MCP

Comment

About the project

Inspiration

Keeping up with competitor news is a full-time job: blog updates, social posts, funding rounds, job listings, app-store reviews… yet most teams still rely on manual Google searches, Slack pings, and scattered spreadsheets. We knew there had to be a smarter way to surface only the signals that matter—on the channels teams already use. Enter Competitor Pulse: built for fast-moving startups that need AI-powered competitive intelligence in real time.

What we built

Competitor Pulse automatically:

  1. Crawls a watch-list of URLs (company sites, blogs, social feeds) via a Firecrawl MCP server
  2. Uses Claude (Anthropic API) to classify relevance and urgency
  3. Compiles structured digests via Gumloop MCP
  4. Delivers actionable notifications to Slack, Email, and Notion

A lightweight React frontend lets you manage your watch-list, preview digests, and tweak filters. Under the hood, a Node.js/Express API orchestrates MCP calls, stores metadata in Supabase, and handles webhooks for Slack and Notion.

What we learned

  • MCP orchestration: Chaining Firecrawl → Claude → Gumloop taught us how to design robust, retry-safe pipelines.
  • Prompt engineering at scale: Balancing precision (don’t miss anything important) with recall (don’t overload users) required iterative A/B testing of classification prompts.
  • Integrations: Building secure, low-latency webhooks into Slack and Notion APIs sharpened our understanding of enterprise-grade delivery.

Challenges we faced

  • Authentication scopes: Ensuring our service keys had just the right permissions across two MCP servers without over-privileging was tricky.
  • Rate limits: We hit Anthropic rate caps during heavy crawls—so we implemented exponential backoff and batched requests.
  • Deduplication: Many companies publish the same update across multiple channels; de-duplicating on semantic similarity took extra tuning with embedding-based hashing.

Built With

Share this project:

Updates