Plugin marketplace for Claude Code and compatible
agents (Kimi Code via shared SKILL.md / hooks format).
Add this marketplace once:
/plugin marketplace add madlexa/skills
Then install any plugin from it:
/plugin install niblet@madlexa-skills
/plugin install speculator@madlexa-skills
To update later:
/plugin marketplace update madlexa-skills
The diligent crumb-keeper for AI coding sessions. After every turn, Niblet
auto-writes findings to the project knowledge base via a single secure
helper. At session end, a sub-agent extracts reusable workflow patterns and
lands them as proposals you review before promoting (via the action-aware
niblet-promote helper — never a raw mv).
Five checkpoint layers:
- FAST (every turn) — agent writes findings to
.claude/kb/and memory feedback. Auto-write, local, reversible. - DEEP (SessionEnd) — sub-agent extracts workflow patterns; skills,
agents, commands, CLAUDE.md edits, and global writes are staged as
proposals in
.niblet/proposals/. Promote vianiblet-promote. - DISTILL (when KB > 20 files or 200 KB) — sub-agent consolidates overlapping KB entries to keep the knowledge base dense and non-redundant.
- AUDIT (every 5 sessions) — sub-agent scans the artifact index for stale paths and contradictions between KB entries.
niblet-status— dashboard showing KB counts, pending proposals, promoted artifacts, and queue depths.
Opt-in guarded auto-apply — set NIBLET_GUARDED_APPLY=1 and run
niblet-promote --guarded-sweep to auto-promote risk=low + confidence=high
MERGE_KB_ENTRY / UPDATE_KB_ENTRY proposals. A timestamped backup is
written before each overwrite. All other action types remain manual.
Sanitized capture — observe.sh logs only tool name + safe path + exit
code. tool_input and tool_response content (where secrets and untrusted
text live) is never stored. Closes the persistent prompt-injection vector.
Validated writes — every ACTION the agent applies goes through
bin/niblet-apply, which enforces slug rules and canonical path
containment. Sub-agent-supplied filenames cannot escape their allowed dir.
Cross-session DEEP queue — SessionEnd writes a queue entry that any
subsequent session drains, even with a new session id. No orphan markers.
Honest KB surfacing — a SessionStart hook emits a compact KB index
reminder so the agent knows which .claude/kb/ topics exist.
A graph-oriented knowledge base for AI coding sessions. Speculator stores your architecture as a knowledge graph of entities and edges right in your repo, so Claude queries the graph instead of re-reading dozens of files before writing a single line.
TypeScript CLI + MCP server — a speculator CLI (add, get, list, search,
update, stats, export, …) plus an MCP server exposing the same graph operations
as tools the agent can call directly.
Hook-driven context injection — a SessionStart hook surfaces knowledge-base stats, and a UserPromptSubmit hook searches the graph for the current prompt and injects the most relevant entities and edges — zero manual lookup.
Curation agents — ships speculator-ai-engineer,
speculator-typescript-cli-developer, and speculator-mcp-protocol-engineer
for building and maintaining the graph.
Requires Node.js (≥ 20) for the CLI and MCP server. The hooks are POSIX shell.
See plugins/speculator/README.md.
skills/
├── .claude-plugin/
│ └── marketplace.json
└── plugins/
└── <plugin-name>/
├── .claude-plugin/plugin.json
├── README.md
├── skills/
├── agents/
├── hooks/
└── ...
Plugins in this marketplace ship as POSIX shell scripts. They run natively
on macOS and Linux. Windows users need WSL2 (recommended) or
Git Bash — cmd.exe and PowerShell are not supported. See each
plugin's README for its specific dependencies (typically bash, jq,
python3, and git).
To work on a plugin without publishing:
claude --plugin-dir ./plugins/niblet