smart-mcps
A Claude Code plugin that installs 8 skills and session hooks for agentmemory, codegraph, NotebookLM, and Perplexity. All tools run as CLI commands — no MCP servers, no context-window pollution.
Installation
Clone the repo into your project (or anywhere you want the skills to live) so you can edit them:
git clone https://github.com/Giovani-Merlin/smart_mcps
Then install the plugin in Claude Code:
/plugin marketplace add https://github.com/Giovani-Merlin/smart_mcps
/plugin install smart-mcps
Skills are registered automatically on install. Restart Claude Code after installation.
Skills are meant to be edited. After installing, configure each skill you use — see Setup below. The notebooklm-chat skill in particular ships with a placeholder notebook map that you must fill in.
Skills
| Skill | Trigger | What it does |
|---|
codegraph | /codegraph | Symbol lookup, call graph tracing, impact analysis via codegraph CLI |
notebooklm-chat | /notebooklm-chat | Chat with a notebook by topic — query only, requires configuring your notebook map in the skill file |
notebooklm-complete | /notebooklm-complete | Full NotebookLM management: query, create, add sources, audio/video artifacts |
perplexity | /perplexity | Web-grounded search, research, and reasoning via smart-mcps-perplexity CLI |
handoff | /handoff | Resume most recent agentmemory session ("where were we") |
recall | /recall | Search past agentmemory observations and lessons |
remember | /remember | Save an insight or decision to agentmemory long-term storage |
recap | /recap | Summarize recent sessions for the current project |
Setup (per tool)
Codegraph
Requires an index built from your project's source. Run once per project:
pip install codegraph # or: uv tool install codegraph
codegraph init
codegraph index
The Setup hook runs codegraph index automatically on each session start if codegraph is installed and an index exists.
Perplexity
Requires a PERPLEXITY_API_KEY. The MCP server is spawned by the VS Code extension host, so the key must be in that process's environment — not just a terminal session.
Recommended: Install the mkhl.direnv VS Code extension and create a per-project .envrc:
cp .envrc.example .envrc # fill in your key
direnv allow
Shell profile fallback (all projects share the same key):
export PERPLEXITY_API_KEY="pplx-..." # in ~/.bashrc or ~/.profile
NotebookLM
Step 1 — one-time browser auth:
uv tool install notebooklm-mcp-cli
nlm login
Re-run nlm login when cookies expire. Auth state is stored in ~/.notebooklm-mcp-cli/ and persists across projects.
Step 2 — seed aliases from your notebook titles:
bash scripts/seed-nlm-aliases.sh
This registers each notebook as an nlm alias (e.g. ltx-2-3-engineering-...). Aliases persist globally across projects. Re-run whenever you create new notebooks. Inspect with nlm alias list.
Step 3 — configure the notebooklm-chat skill (required):
Open skills/notebooklm-chat/SKILL.md and fill in the notebook map:
| Topic | Notebook alias or ID |
| ----- | -------------------- |
| LTX 2.3 / video generation / cinematography | `ltx-2-3-engineering-and-ic-lora-implementation-guide` |
| ControlNet / pose / depth | `floed-and-comfyui-controlnet-aux-development-summaries` |
Map by semantic topic (not just notebook title) so the skill can match natural-language questions. Use the alias slugs from step 2 or raw UUIDs — both work.
Use /notebooklm-complete for everything beyond querying: adding sources, creating audio overviews, managing notes.
AgentMemory
Requires the agentmemory daemon running locally:
systemctl --user start agentmemory
# or
~/.agentmemory/start.sh
The recall, remember, handoff, and recap skills use the smart-mcps-agentmemory CLI which talks to this daemon.
Hooks
The plugin installs session hooks that wire agentmemory capture automatically:
| Hook | What it does |
|---|
sessionStart | Loads session context from agentmemory |
userPromptSubmitted | Records user intent |
preToolUse / postToolUse | Captures tool calls and results |
postToolUseFailure | Records failures for learning |
preCompact | Saves context before compaction |
sessionEnd / agentStop | Closes session and flushes observations |
subagentStart / subagentStop | Tracks sub-agent lifecycle |
notification | Handles async notifications |
Editing skills
