PT-Edge is an MCP server that gives AI assistants live, structured knowledge about the AI ecosystem. It indexes open-source projects, HuggingFace models and datasets, public APIs, and community discourse — then exposes 47 MCP tools, 3 resources, 3 resource templates, and 4 prompts for discovery, comparison, and trend analysis.
Built by Phase Transitions — a weekly newsletter on building with AI, from architecture decisions to production patterns.
- Daily ingests pull GitHub stats, package downloads, releases, HN posts, V2EX discussions, newsletter coverage, HuggingFace models/datasets, public API specs, and npm registry MCP servers
- Discovery indexes — 11K+ AI repos, 18K+ HuggingFace models, 42K+ datasets, 2,500+ public APIs, all with 256d semantic embeddings, hybrid search, name-match boosting, staleness signals, and pagination
- Materialized views compute derived metrics: momentum, hype ratio, tiers, lifecycle stage
- LLM-powered enrichment — Claude Haiku summarises releases and newsletter topics; OpenAI embeds everything for semantic search
- 47 MCP tools let you query this data naturally in conversation
- MCP resources & prompts — 3 static resources (methodology, categories, coverage), 3 parameterised resource templates (project, lab, category), and 4 compound query prompts (evaluate-technology, build-something, due-diligence, weekly-briefing)
- Community feedback system — corrections, article pitches, and lab event tracking
| Category | Tools |
|---|---|
| Discovery | about, whats_new, trending, lifecycle_map, hype_landscape |
| Deep Dives | project_pulse, lab_pulse, hype_check |
| Comparison | compare, movers, related, market_map |
| Project Discovery | radar, scout, deep_dive, sniff_projects, accept_candidate, topic, hn_pulse |
| AI Ecosystem Search | find_ai_tool, find_mcp_server, find_public_api, find_dataset, find_model (all support offset for pagination) |
| API Intelligence | get_api_spec, get_api_endpoints, get_dependencies, find_dependents |
| Community | submit_feedback, upvote_feedback, list_feedback, amend_feedback, propose_article, list_pitches, upvote_pitch, amend_pitch |
| Lab Intelligence | submit_lab_event, list_lab_events, lab_models |
| Methodology | explain |
| Power User | describe_schema, query, set_tier |
| Type | Items |
|---|---|
| Resources | methodology, categories, coverage |
| Resource Templates | project/{slug}, lab/{slug}, category/{category} |
| Prompts | evaluate-technology, build-something, due-diligence, weekly-briefing |
- Hype Ratio — stars / monthly downloads. High = GitHub tourism. Low = invisible infrastructure.
- Tiers — T1 Foundational (>10M downloads), T2 Major (>100K), T3 Notable (>10K), T4 Emerging
- Lifecycle — emerging → launching → growing → established → fading → dormant
- Momentum — star and download deltas over 7-day and 30-day windows
PT-Edge uses the MCP Streamable HTTP transport. Connect via:
https://mcp.phasetransitions.ai/mcp?token=YOUR_TOKEN
Works with Claude Desktop, Claude.ai (web connector), and any MCP-compatible client.
- Runtime: Python 3.11, FastAPI, FastMCP
- Database: PostgreSQL 16 with pgvector
- Embeddings: OpenAI text-embedding-3-large (256d Matryoshka for discovery indexes, 1536d for project/methodology)
- LLM: Claude Haiku 4.5 (release + newsletter summarisation)
- Hosting: Render (web service + cron + managed Postgres)
# Clone and set up
git clone https://github.com/grahamrowe82/pt-edge.git
cd pt-edge
cp .env.example .env # Add your API keys
# Start database
docker compose up -d
# Run migrations
python -m app.migrations.run
# Start server
uvicorn app.main:app --reload
# Run daily ingest
python scripts/ingest_all.pyMIT — see LICENSE.