How we curate

Lists tell you what exists. StackMap tells you what goes together — and why.

Typed relationships, not links

Every connection in the catalog is one of three kinds:

• complements — use them together; one covers what the other doesn't.
• alternative — different project, same job; we say when to pick which.
• built_with — depends on, builds on, or runs on.

And every connection carries a written "why" — the sentence a colleague would tell you over your shoulder. This is what a why looks like:

Nothing publishes without a human

1. 1 · Draft. A new repo is analyzed against the whole catalog — summary, topics, and candidate typed edges with confidence scores.
2. 2 · Review. A human reads every proposed edge and its why. Weak edges are cut, wrong ones rejected, the note sharpened. Repos that don't earn a place are refused.
3. 3 · Publish. Only approved entries reach the catalog. The curator's take on each repo page includes when not to use a tool — that's the part no awesome-list will tell you.

And we do say no

Curation you can't see is curation you can't trust. Recent refusals, with the reason:

• PartMode — Browser parametric CAD (OpenCascade WASM) with a genuinely good typed-agent permission model — but the core is mechanical design, no existing topic fits, and nothing in the catalog does CAD, so it lands with no honest edge. (scope)
• avoid-slop — A curated directory of anti-slop tools, not a tool — nothing can honestly pair with a link collection. (scope)
• skills (alfonsograziano) — Personal three-skill collection at 6 stars — the catalog vouches for what it maps; revisit if it grows. (maturity)
• index — Strong browser agent gone quiet — 14 months without a push, README pitching last year's models; the catalog vouches for what it maps. Revisit if it wakes. (maturity)
• Auto-Deep-Research — Ten months without a push and no license — two strikes; the catalog vouches for what it maps. Its parent AutoAgent framework may merit evaluation if it's still moving. (maturity)
• blueprint-mcp — A deployment recipe for Arcade's hosted MCP platform — sign-up, deploy and gateway required, no license. Vendor onboarding sample, not a self-contained tool. (scope)
• pdf-mcp — PDF MCP server at 52 stars — far below the bar, and the map's PDF tooling runs deep already. (maturity)
• jobseek — Career-page monitor at 163 stars — right neighborhood (would pair with ai-job-search), too early; revisit when it grows. (maturity)
• multi-agent-generator — Prompt-to-CrewAI/LangGraph code generator at 213 stars — thin surface today; revisit if it takes off. (maturity)
• bmad-module-creative-intelligence-suite — No license, 175 stars, and an add-on module for a method whose core isn't mapped — the catalog vouches for what it maps. (maturity)
• deepseek_ocr_app — Self-described 'quick vibe coded' demo wrapper around DeepSeek-OCR — an end-user app, and the model class is already mapped (unlimited-ocr, chandra). (scope)
• GeoLibre — Cloud-native GIS platform — a superb tool with no AI/agent surface; no honest topic or edge in a map of AI-stack building blocks. (scope)
• zero-employee-company-book — A book about AI-run companies, not a tool you can put in a stack. (scope)
• jobsync — An end-user job-tracker app that uses AI — no honest topic in a map of AI-stack building blocks. (scope)
• pake — Webpage-to-desktop-app packager — a superb tool with zero AI/agent relevance; wrong map. (scope)
• google-maps-agent-skills — Right scope (agent skills), 121 stars and weeks quiet — revisit when it grows. (maturity)
• seo-geo-claude-skills — A signpost repo pointing at its own replacement, 85 stars — map the canonical home if it earns it. (maturity)
• wigolo — Right scope (local-first web MCP), public beta at 85 stars — revisit when it grows. (maturity)
• awesome-harness-engineering — An awesome-list, not a tool — nothing can honestly pair with a link collection. (scope)
• FlashDB — Embedded MCU database, not an embedding database — no honest topic or edge in an AI/agent catalog. (scope)
• SwarnDB — Interesting vector+graph idea at 13 stars with no license — the catalog vouches for what it maps. (maturity)

The same gate applies to connections. Our discovery layer suggests semantically related pairs; a human promotes each one into a typed edge — or dismisses it. All 98 dismissals are public. The most recent:

• acontext ~ bemyagent — adjacent vocabulary, different jobs — neither swappable nor demonstrably used together (2026-08-24)
• acontext ~ squid — adjacent vocabulary, different jobs — neither swappable nor demonstrably used together (2026-08-24)
• adala ~ gigatoken — adjacent vocabulary, different jobs — neither swappable nor demonstrably used together (2026-08-24)
• adrian ~ plano — adjacent vocabulary, different jobs — neither swappable nor demonstrably used together (2026-08-24)
• adrian ~ agency-agents — extractor hub artifact — 'Adrian' as a concept node pulls in unrelated coding-agent repos; neither README mentions the other (2026-08-24)
• adrian ~ jcode — extractor hub artifact — 'Adrian' as a concept node pulls in unrelated coding-agent repos; neither README mentions the other (2026-08-24)
• adrian ~ maggy — extractor hub artifact — 'Adrian' as a concept node pulls in unrelated coding-agent repos; neither README mentions the other (2026-08-24)
• adrian ~ memanto — extractor hub artifact — 'Adrian' as a concept node pulls in unrelated coding-agent repos; neither README mentions the other (2026-08-24)

Kept honest, automatically

Stars, activity and imagery refresh twice a week straight from GitHub, so a card saying updated yesterday means yesterday. Know a repo that belongs on the map? Suggest it — every submission goes through the same human review.

Curious who's behind this? About StackMap & ShipWithAI →