I build safety-first tooling for AI agents from Thailand — tools that measure instead of trust, ask before they spend, and verify or flag instead of guessing.
A mining series of quality/safety tools for AI coding agents — mostly cross-agent (Claude Code, Antigravity, Codex, Cursor and more); each tool's README states exactly how far it has been taken, and CoalTipple is Claude Code only. The State column tells the two kinds apart: a stable tool has a published stable Release, a beta tool has none yet.
New here? CoalMine is the longest-running tool and the natural starting point.
| Tool | What it does | State |
|---|---|---|
| CoalMine | 9 quality canaries — code rot, security, drift, supply-chain — wired to fire automatically at session end | stable |
| CoalTipple | Model/effort router — delegates cheap work down, escalates hard work up | stable |
| CoalBoard | Multi-lens consensus board for error-not-allowed decisions | stable |
| CoalHearth | Session warm-resume + handoff journal | stable |
| CoalFace | Fan-out discipline for multi-agent swarms — bounded cost, single-writer apply | stable |
| CoalWash | Memory washer/defragmenter with a mechanical fidelity gate | stable |
| CoalLedger | Docs-health canaries — "CoalMine for documentation" | beta |
| CoalGob | Recoverable-delete guard for AI coding agents. Classifier only: no hook yet, so nothing is blocked or rerouted today | beta |
📖 Docs: TheColliery documentation
License: every tool above is Apache-2.0.
House rules the whole series lives by: zero dependencies · offline-capable · fail-silent hooks that never break your session · every costly action consent-gated · honest claims only (a README must match what the code actually does).
Kolwen — a Thai-first LLM, being built the way this whole series works: verify or flag, never guess.
- Raised by many teachers, shadow of none — assembled from the world's open bloodlines, with every ingredient credited
- A birth certificate you can test: no frontier-model outputs in the training corpus, and we publish the independence evidence to prove it
- Free local weights first; details, benchmarks, and dates when they're real — this page will update as the kiln burns
(Details will be updated gradually — the kiln is still in operation.)
- Contributing test-generation fixes to move-test-gen
- Contributing fixes and hardening to Win11Debloat — 26 merged pull requests
- Was a team member at KatomnoiStudio — a Thai game project (org archived Sep 2026)
- Also on GitLab: gitlab.com/HetCreep
(links last verified 2026-10-23)
- SpriteDesignDatum — a geometry standard for 2D sprite art, and an honest map of which parts of it any authority actually publishes
- hetcreep-docs — my articles and notes
- Read the articles: Sprite Design Datum, the standard · Gacha Rate Design Datum
If these tools save you tokens or catch your bugs: Sponsor TheColliery — every bit funds more open tooling.
If you'd like to back my own work, the design-datum articles and my fixes to other projects such as Win11Debloat and move-test-gen: Sponsor HetCreep.
Tool questions: info@thecolliery.org · Kolwen and partnerships: info@kolwen.com
measured, not trusted · consent-gated · verify or flag