InertialAI Claude Code plugin

Time-series embedding and analysis tools for Claude Code, powered by the InertialAI API.

Install

/plugin marketplace add InertialAI/claude-plugin
/plugin install inertialai@inertialai-plugins

Then configure your API key (see below). Requires uv on PATH — the MCP server runs as a uv script with inline dependencies, nothing else to install.

API key

The server resolves the key in this order: INERTIAL_API_KEY env var → OS keychain → none. Pick whichever fits your platform:

• OS keychain (recommended on macOS / Linux desktop / Windows). In Claude Code, run /inertialai:setup — it walks you through running a one-time terminal command that reads the key via getpass, validates it, and stores it in your OS keychain. The key never enters the chat.
• Environment variable (works everywhere including headless Linux / WSL). Add export INERTIAL_API_KEY='your-key' to ~/.zshrc or ~/.bashrc, restart your shell.

Never paste your API key into a Claude Code prompt — it would land in transcripts and request logs. Use one of the two paths above.

Layout

.
├── .claude-plugin/marketplace.json          # marketplace catalog
└── plugins/inertialai/
    ├── .claude-plugin/plugin.json           # plugin manifest
    ├── .mcp.json                            # declares the MCP server
    ├── scripts/setup-key.py                 # one-time interactive key installer
    ├── server/                              # MCP server
    │   ├── server.py                        # entry point (PEP 723 script): builds FastMCP, registers tools
    │   ├── store.py                         # SQLite-backed embedding store + vector helpers
    │   ├── auth.py                          # API key resolution + setup-error payload
    │   └── tools/                           # one file per MCP tool, each a Tool subclass
    │       ├── _base.py                     # Tool base class
    │       ├── __init__.py                  # ALL_TOOLS registry
    │       ├── create_embedding.py
    │       ├── list_models.py
    │       ├── compare.py
    │       ├── find_similar.py
    │       ├── classify.py
    │       ├── list_embeddings.py
    │       ├── delete_embedding.py
    │       └── check_setup.py
    └── skills/                              # playbooks Claude auto-invokes
        ├── TEMPLATE.md                      # copy this into skills/<name>/SKILL.md
        ├── analyze-imu-session/SKILL.md
        ├── find-anomalies/SKILL.md
        ├── compare-sessions/SKILL.md
        └── setup/SKILL.md

Adding a new MCP tool

1. Create plugins/inertialai/server/tools/<your_tool>.py defining a subclass of Tool (tools/_base.py). Set name = "<your_tool>" and implement run (sync or async) with typed parameters and a docstring — FastMCP derives the JSON schema and description from it.
2. Import the class in tools/__init__.py and append it to ALL_TOOLS.
3. Restart the MCP server (or run /reload-plugins).

No edits to server.py required.

Adding a new skill

Copy plugins/inertialai/skills/TEMPLATE.md into a new directory:

mkdir plugins/inertialai/skills/<your-skill>/
cp plugins/inertialai/skills/TEMPLATE.md plugins/inertialai/skills/<your-skill>/SKILL.md

Fill in the frontmatter and sections, then reload.

Tools

All tools operate on stored handles rather than raw vectors, so 512-dim floats never enter the model's context.

Tool	Purpose
create_embedding	Call the API, store the vector locally, return a handle. Pass label to add to a classifier corpus.
list_models	List available embedding models.
compare	Cosine similarity between two handles.
find_similar	Top-k most similar handles to a query.
classify	Nearest-neighbor classify against the labeled corpus.
list_embeddings / delete_embedding	Manage the local store.
check_setup	Diagnose API key state and surface setup instructions.

Embeddings persist in ${CLAUDE_PLUGIN_DATA}/embeddings.db.

Adding new models

The API takes model as a parameter, so new embedding models work without a plugin update. New endpoint families (forecasting, classification-as-a-service, etc.) get new tools added to server/server.py.

Roadmap

Currently Claude Code only. The MCP server runs as a local subprocess (uv run --script), which Cowork's hosted runtime doesn't support. Cowork compatibility is pending a hosted HTTP MCP gateway at mcp.inertialai.com; once that ships, the same plugin will work in both products.