A Claude Code / Cowork plugin marketplace by Ink. Developer-grade automation, starting with memory hygiene.
| Plugin | Description |
|---|---|
| memory-keeper | Keep file-based agent memory lean and durable — compact the index, archive done/stale memories to cold storage, lint against a byte budget. Ships a skill, slash commands, a zero-dependency MCP server, and a stdlib CLI. |
| focus-notify | Focus-aware desktop notifications for Claude Code on macOS, Linux & Windows — notify only when your terminal is not the frontmost app. Three kinds (finished, approve, question), each with its own sound, plus click-to-focus. |
| kkskills-essentials | 13 generic, self-improving working skills — feedback discipline, clean-architecture & conventional-commits references, proactive task reminders, secret hygiene, concurrency, MCP authoring, timezone & local-LLM selection, plus a meta-skill that folds new learnings back in (versioned + changelogged). Skills-only. |
| kkskills-personal | Personal / project-specific skills split out from essentials — a user-profile calibration skill and a dated-filename reference convention. Fork and replace the specifics with your own. Skills-only. |
| loop-ledger | A guardrail for agent loops — a deterministic state machine (budget + stall-hash + exit predicate) that decides should-continue per iteration, plus an opt-in Stop hook that enforces it so the agent can neither quit early nor spin forever. Zero-dependency MCP server. |
| reflexion | Self-improving agent memory built on memory-keeper — recall matching lessons before a task, reflect a new lesson after it, browse the error/success library, and compress recurring lessons into higher-level patterns. |
| critique-gate | A quality brake for agent output — generate, critic scores against a rubric, rewrite, re-score, wired to loop-ledger's retry-count backstop. Four modes: score-retry, multi-critic, adversarial, judge-ensemble. Skills-only, no new server. |
| prompt-evolve | Evolve a prompt against a test set — run it, score every output, rewrite the prompt from its failures, repeat. Deterministic scorers (exact-match/regex/schema) plus a fully offline verification loop; stdlib CLI + skill, no model calls of its own. |
| agent-loops | Playbook for the 5 categories of agent loop design patterns (Quality, Memory, Planning, Exploration, System Optimization) — when to use each, its minimal shape, and the exact loop-ledger brake to wire. Plus a loop-selector router skill. Skills-only; pairs with loop-ledger for enforcement. |
| tool-metrics | Count Claude Code tool usage across every session — built-in tools, MCP tools grouped by server, skill invocations and subagent spawns. Merges session transcripts with a durable PostToolUse event log; /tool-stats prints a terminal table, /tool-stats-dashboard publishes an HTML dashboard. Stdlib-only, no dependencies. |
Together with loop-ledger and memory-keeper, these plugins complete the "self-improvement chain" —
see docs/proposals/SELF-IMPROVEMENT-CHAIN_SPEC.md for
how they wire together (no new plugin needed — it's a composition of the ones above).
Add the marketplace, then install a plugin:
/plugin marketplace add https://github.com/Ink01101011/kktest-dev
/plugin install memory-keeper@kktest-devOr from the CLI:
claude plugin marketplace add https://github.com/Ink01101011/kktest-dev
claude plugin install memory-keeper@kktest-devUpdate later with /plugin marketplace update kktest-dev.
git clone https://github.com/Ink01101011/kktest-dev.git
claude plugin marketplace add ./kktest-dev
claude plugin install memory-keeper@kktest-dev
claude plugin validate ./kktest-dev # validate marketplace
claude plugin validate ./kktest-dev/plugins/memory-keeper # validate the pluginPer-plugin guides live under docs/, one folder per plugin.
memory-keeper
focus-notify
kkskills-essentials
kkskills-personal
loop-ledger
MIT — see LICENSE.