Skip to main content
Glama
centia-io

Centia MCP Server

Official
by centia-io
README.md
# Centia MCP Server

A Model Context Protocol (MCP) server that exposes the Centia API as MCP tools generated from the OpenAPI spec. It communicates over STDIO and can be used with any MCP-compatible client (Claude Desktop, MCP Inspector, etc.).

## Prerequisites
- Node.js 18+ (20+ recommended)
- npm 9+

## Install
```bash
# From the project root
npm install
```

## Authentication

The MCP server uses the same login as the [`gc2-cli`](https://www.npmjs.com/package/@mapcentia/gc2-cli):

```bash
npm i -g @mapcentia/gc2-cli
gc2 connect https://api.centia.io   # only if not the default
gc2 login                            # browser, device, or password
```

The MCP server reads `~/.config/configstore/gc2-env.json` (managed by `gc2-cli`) and refreshes the access token automatically when it expires.

For headless / CI use, set `API_TOKEN` (and optionally `API_BASE_URL`) — env vars take precedence over the stored session:

- `API_BASE_URL` (optional) — Base URL for the Centia API. Default: `https://api.centia.io`
- `API_TOKEN` (optional) — Static personal access token. Bypasses the stored `gc2-cli` session.

The API surface is defined in `centia-api.json` (already included in the repo). The server reads it at runtime to generate tools.

## Run locally (development)
Hot‑reload development run with `tsx`:
```bash
npm run dev
```
This starts the MCP server on STDIO. It is meant to be launched by an MCP client (see below), but you can also smoke‑test it with the MCP Inspector.

## Run with npx
You can run the server directly using `npx`.

If running from the source locally:
```bash
npm run build
npx .
```

If the package is installed or you want to use the published version:
```bash
npx @centia/mcp-server
```

## Build and run (production)
```bash
npm run build
npm start
```
This compiles TypeScript to `dist/` and starts `node dist/index.js`.

If you prefer a static token over the stored `gc2-cli` session (e.g. CI):
```bash
API_TOKEN=your_token_here npm start
# or
API_BASE_URL=https://api.centia.io API_TOKEN=your_token_here npm run dev
```

## Using with MCP Inspector (recommended for local testing)
MCP Inspector lets you connect to the server and try tools interactively.

1. Start the Inspector UI:
   ```bash
   npx @modelcontextprotocol/inspector -- npx @centia/mcp-server
   ```
2. If needed, add environment variables in the Inspector connection dialog (e.g., `API_TOKEN`).

## Using with Claude Desktop
Add the server to your Claude Desktop MCP config (e.g., `claude_desktop_config.json`).

### Using npx (easiest)
```json
{
  "mcpServers": {
    "centia": {
      "command": "npx",
      "args": ["-y", "@centia/mcp-server"]
    }
  }
}
```

Run `gc2 login` once and the server will pick up the session. To force a static token instead, add `"env": { "API_TOKEN": "YOUR_CENTIA_TOKEN" }`.

### Using local source
```json
{
  "mcpServers": {
    "centia": {
      "command": "npm",
      "args": ["run", "start"],
      "cwd": "/absolute/path/to/your/mcp-server"
    }
  }
}
```
- For development, replace `"args": ["run", "start"]` with `"args": ["run", "dev"]`.
- Ensure `cwd` points to this project directory.

## NPM scripts
- `npm run dev` — Run TypeScript directly with `tsx` (hot‑reload style dev loop).
- `npm run build` — Compile TypeScript to CommonJS/ESM in `dist/` via `tsc`.
- `npm start` — Run the built server (`node dist/index.js`).

## Using AGENTS.md and Skills with AI Coding Agents
This repository includes an `AGENTS.md` file with global hard rules and a `skills/` folder with task-specific guidance for AI coding agents (like Claude Code, Junie, etc.) when working with Centia BaaS.

`AGENTS.md` contains the core rules (tool priority, security baselines, destructive-ops policy), while each skill in `skills/*/SKILL.md` covers a specific workflow area.

To ensure your AI agent follows these rules while developing your application:

1. Copy `AGENTS.md` and the `skills/` folder from this repository to the root of **your own application's** repository.
2. When starting a session with your AI agent, it will automatically find and follow the guidelines.

**AGENTS.md** covers:
- **Prime Directive**: Preferring MCP tools and official SDKs.
- **Tool Priority**: Order of interaction (MCP tools > SDK > HTTP).
- **Security & Auth**: How to handle tokens and OAuth flows correctly.
- **Destructive Operations Policy**: Requiring explicit confirmation before destructive changes.

**Skills** (`skills/*/SKILL.md`) cover task-specific workflows:
- `centia-runtime-sdk` — Building JS/TS runtime code with `@centia-io/sdk`
- `centia-provisioning` — Schema, tables, columns, constraints, indexes, migrations
- `centia-json-rpc` — JSON-RPC method lifecycle and TypeScript interfaces
- `centia-auth-model` — Auth across browser, backend, CLI, and provisioning
- `centia-types-formats` — PostgreSQL types, casts, type hints, output formats
- `centia-file-import` — File import workflow
- `centia-ows` — WMS/WFS/WFS-T and MapCache proxy endpoints and their access model
- `centia-ogc` — OGC API Features and Maps (GeoJSON items, map images)
- `centia-openapi-docs` — OpenAPI and docs fallback policy
- `centia-delivery` — Delivery checklist and quality gate

## Troubleshooting
- Tools missing or inputs look odd: ensure `centia-api.json` exists and is valid. The server generates tools from this file at startup.
- 401/403 errors: run `gc2 login` (or set a valid `API_TOKEN` in the environment for headless use).
- JSON Schema validation errors: schemas are auto‑normalized/sanitized for MCP, but if you updated `centia-api.json`, re-run and check logs for details.
- ESM/CommonJS issues: this project uses ESM (`"type": "module"`). Use Node.js 18+ and run scripts via npm as shown above.

## License
ISC

TDQS

C2.8/5.0

Scored across 56 tools

Disambiguation5/5

Every tool name uniquely pairs a verb (delete, get, patch, post) with a resource (client, column, etc.), leaving no ambiguity. Even with 56 tools, each combination is distinct and clearly indicates the resource and action.

Naming Consistency5/5

All tool names follow a strict verb_noun pattern using PascalCase and HTTP verbs (delete, get, patch, post). This consistency makes the tool set predictable and easy to navigate.

Tool Count2/5

With 56 tools, the count far exceeds the recommended range (3-15). While the domain may require many operations, this volume is likely to overwhelm agents and suggests an overly broad scope.

Completeness4/5

The tool set covers CRUD for most major resources (tables, schemas, columns, etc.) and includes file handling, SQL, GraphQL, and OAuth. Minor gaps exist (e.g., no delete for events or devices), but the core database operations are well-covered.

Maintenance

ActivityNo data
ResponsivenessNo issues