# Claude Code plugin

The LocalData MCP repository is also a Claude Code plugin. Its manifest,
`.claude-plugin/plugin.json`, registers the `localdata` MCP server with the
command `uvx localdata-mcp` and ships 18 skills and 11 agents that drive the
tools documented in the [tools reference](tools-reference.md).

Skills and agents call the MCP server; they add no analytical capability of
their own. Everything they do is reachable by calling the tools directly.

## Installing

Clone the repository and point Claude Code at it with `--plugin-dir`:

```bash
git clone https://github.com/ChrisGVE/localdata-mcp.git
claude --plugin-dir /path/to/localdata-mcp
```

Confirm what was picked up:

```bash
claude --plugin-dir /path/to/localdata-mcp plugin details localdata-mcp
```

That prints the component inventory — it should report 18 skills and 11 agents.
`claude --plugin-dir /path/to/localdata-mcp mcp list` shows the MCP server as
`plugin:localdata-mcp:localdata`; it launches through `uvx`, so the first start
downloads the package and its dependencies.

`claude plugin install localdata-mcp` does not work yet. That command installs
from a marketplace, and this repository does not ship a
`.claude-plugin/marketplace.json`, so `claude plugin marketplace add` on it
fails with "Marketplace file not found". `--plugin-dir` is the supported route
for now.

The MCP server on its own — without the skills and agents — needs no plugin
machinery. Add it to any MCP client as described in
[getting started](getting-started.md).

## Using a skill

A skill is invoked by name as a slash command in a Claude Code session:

```
/forecast sales_db revenue
```

The arguments follow each skill's `argument-hint`; `forecast` declares
`"<database-name> <column-name>"`. Claude also selects a skill on its own when
the request matches its `description`, so "forecast next quarter's revenue from
`sales_db`" reaches the same place without the slash command.

Agents are not invoked by slash command. Claude delegates to one when a request
matches the agent's `description`, or you can ask for it directly — "use the
forecaster agent on this series".

Skills and agents call the MCP tools on your behalf, so the `localdata` server
has to be connected first. Connect your data source with `connect_database`
before invoking one; a skill has no way to guess which file or database you
mean.

## Skills

Each skill is a `SKILL.md` inside a group directory: `skills/<group>/<skill-name>/SKILL.md`.
The manifest lists the five group directories under its `skills` key, and Claude
Code scans one level below each of them — which is why the group directories are
named individually rather than pointing at `skills/` and relying on a recursive
walk. Each skill declares an `allowed-tools` list restricting it to the tools it
needs, and an `argument-hint` describing what to pass.

### `exploration/`

| Skill | Purpose |
| --- | --- |
| `explore-data` | Connect to a source, profile schema and quality, recommend which analyses fit |
| `data-quality` | Assess completeness, consistency, validity, and uniqueness before analysis |
| `find-reference-data` | Identify and prepare external reference data — benchmarks, demographics, economic indicators, geographic context |

### `statistical/`

| Skill | Purpose |
| --- | --- |
| `hypothesis-test` | Check assumptions, select the test, report the result in plain language |
| `ab-test` | Analyze an experiment and return a ship / iterate / no-ship recommendation |
| `analyze-correlations` | Find strong relationships between variables and suggest regression models |
| `sampling-estimation` | Design a sampling strategy and estimate parameters with confidence intervals |

### `modeling/`

| Skill | Purpose |
| --- | --- |
| `regression` | Fit and evaluate models predicting a target variable |
| `cluster-analysis` | Discover groupings, with cluster-count evaluation |
| `anomaly-detection` | Flag outliers with isolation forest or local outlier factor |
| `dimensionality-reduction` | Compress high-dimensional data with PCA, t-SNE, or UMAP |
| `forecast` | Analyze a time series and forecast forward with confidence intervals |
| `geospatial` | Distances, geographic clusters, accessibility |
| `optimization` | Resource allocation, scheduling, and process optimization under constraints |

### `graph-data/`

| Skill | Purpose |
| --- | --- |
| `graph-data-explore` | Walk a DOT, GML, GraphML, or Mermaid graph — nodes, edges, statistics, paths, export |

### `workflow/`

| Skill | Purpose |
| --- | --- |
| `data-pipeline` | End-to-end run: connect, profile, analyze, report |
| `research-pipeline` | Hypotheses, power analysis, assumption checks, reproducible reporting |
| `process-control` | Control charts, process stability and capability, out-of-control conditions |

## Agents

Agents live in `agents/`, one Markdown file per agent. All run on Sonnet; the
turn budget in the table is the agent's `maxTurns` and bounds how long an
analysis may run before it must report back.

| Agent | Turns | Scope |
| --- | --- | --- |
| `data-explorer` | 15 | Profiles an unfamiliar dataset: schema, quality, patterns, summary report |
| `data-scientist` | 30 | Composes multi-step pipelines across domains when the right approach is unclear |
| `statistical-analyst` | 20 | Hypothesis tests, ANOVA, effect sizes, sampling design, bootstrap estimation, non-parametric tests |
| `ml-analyst` | 25 | Clustering, anomaly detection, dimensionality reduction, regression modeling |
| `forecaster` | 20 | Decomposition, stationarity testing, ARIMA/ETS choice, forecasts with uncertainty bounds |
| `bi-analyst` | 20 | A/B testing, cohort analysis, CLV, attribution, funnels; translates statistics into business recommendations |
| `graph-data-analyst` | 15 | Centrality, community detection, path finding, visualization export |
| `geospatial-analyst` | 20 | Coordinate systems, spatial relationships, distances, clustering, interpolation, accessibility |
| `operations-analyst` | 20 | Statistical process control, optimization, capacity planning, efficiency analysis |
| `research-analyst` | 25 | Methodological rigor: assumption documentation, power analysis, reproducible reporting |
| `data-researcher` | 20 | Finds, downloads, and prepares public reference datasets to enrich user data |

## Adding a skill or agent

Place a new skill in the domain directory it belongs to — `skills/exploration/`,
`skills/statistical/`, `skills/modeling/`, `skills/graph-data/`, or
`skills/workflow/` — as `<skill-name>/SKILL.md`. The directory name is the skill
name and must match the `name` field in the file's frontmatter. Agents are a
single `agents/<agent-name>.md`, with the same name-matching rule.

A new group directory has to be added to the `skills` array in
`.claude-plugin/plugin.json` as well; a skill in a directory that is not listed
there is not discovered, and nothing reports an error. Check with
`claude --plugin-dir . plugin details localdata-mcp` — the skill count in the
component inventory is the answer.