AI-HUB
A structured knowledge base for AI developer tools — 102 registered tools generated from one verified registry.
- Status
- Active — verified Sep 2026
- Year
- 2026
- Role
- Registry design, validation and generation scripts, CI pipeline, and site structure.
01Problem
AI tooling changes weekly — pricing, models, platforms, and features drift constantly. A hand-maintained list of 100+ tools rots silently and becomes untrustworthy.
Most "awesome AI tools" pages are link dumps with no verification model, no status handling for discontinued tools, and no way to prove generated content matches its source.
02Solution
Treat content as data: edit ai-coding-tools/data/tools.yaml — the single source of truth — never the generated pages.
Validate before anything generates: schema, enums, duplicate detection, and reference checks must print RESULT: VALID with 0 errors.
Generate everything derived: 102 tool pages, indexes, comparison tables, and the master matrix all come from the registry via scripts.
03Implementation
Verify links in CI: internal link integrity must report 0 broken; external links re-checked weekly (report-only, because networks flake).
Mark uncertainty honestly: unverifiable facts are "Unknown"; archived, discontinued, or renamed tools stay in the registry with an explicit status instead of being silently deleted.
03ATechnical flow
data/tools.yaml
Single source of truth — 102 tools
validate-data.py
Schema, enums, duplicates, references
generate-tool-pages.py
40 detailed tool pages
generate-index.py
Indexes, comparisons, master matrix
check-links.py
Internal link integrity — 0 broken
check-external-links.py
Best-effort live URL check
CI — validate.yml
Regeneration parity + markdown lint
04Technology
- Python
- TypeScript
- YAML registry
- GitHub Actions
- markdownlint
- Static site generation
05Features
- 01102 registered tools across 6 categories, each with a Last verified date (current pass: September 2026).
- 0240 detailed tool pages generated from the registry, 10 comparison guides, and 1 master matrix.
- 03Compatibility classification (Recommended / Usable with limitations / Not ideal) for a specific 8 GB RAM machine — derived from documented requirements and hardware reasoning, not personal benchmarks.
- 04Beginner and reference guides: autocomplete vs chat vs agent, BYOK, MCP, RAG, autonomy scale (L0–L5), AI coding security, hardware and student guides.
- 05Explicit lifecycle status — e.g., Roo Code's May 2026 shutdown and gpt-engineer's April 2026 archive are documented, never silently removed.
- 06CI enforces 0 validation errors, regeneration parity, internal link integrity, and markdown linting on every push; external links re-checked weekly.
06Visual evidence
Working demo
Tool registry explorer
Search a small validated registry slice, just like the generated reference pages.
Local interaction · no fake API response
07Links
BEngineering decisions
- D01
Registry first, pages second
Generated files are never edited by hand. Contributors edit the registry and cite official sources — regeneration parity in CI fails the build if generated output drifts from the registry.
- D02
Unknown beats guessed
Anything that cannot be verified against an official source is marked Unknown. No benchmarks or statistics are invented; capability verdicts derive only from documented features.
- D03
Keep the dead tools
Discontinued, renamed, and acquired tools stay in the registry with explicit status flags — history is data, and deleting it would corrupt comparisons.
- D04
Transparency about AI involvement
The repository states plainly that research is AI-assisted, with human review against official sources before storage — disclosed as a process, not hidden.
CChallenges
- C01
Keeping 100+ pages honest
Manual sync is impossible at this scale — solved with generation parity checks so CI fails whenever pages no longer match the registry.
- C02
External link rot
Live URL checks are noisy because networks fail; they run weekly and report-only, while deterministic internal link checks gate every push.
- C03
Pricing and platform drift
Advertised price ≠ total cost and features change monthly — mitigated by per-entry verification dates instead of pretending content is permanently true.
DResult
- A working content pipeline: 102 tools, 40 tool pages, 6 categories, 10 comparison guides, and 1 master matrix — all generated from one registry.
- CI-gated quality: validation with 0 errors, regeneration parity, internal link integrity, and markdown lint on every push.
- MIT-licensed so others can fork, correct, and build on the system.
EWhat I learned
- “Content scales like software when you give it a schema, a validator, and a generator.”
- “A "Last verified" date is more honest than an evergreen claim.”
- “The most valuable engineering in a knowledge base is the machinery that proves it is still true.”