Openemis Pro

title: openemis-mcp-pro — Read + Write MCP Server for OpenEMIS School Management description: openemis-mcp-pro is the read and write MCP server that bridges AI assistants to the OpenEMIS school management information system — 678 resources, 3361 endpoints, 40 playbooks. keywords:

• OpenEMIS
• school management system
• education management
• student attendance
• student risks
• MCP server
• write tools

openemis-mcp-pro — Read + Write AI Bridge for OpenEMIS School Management

A natural-language bridge between MCP-aware agents (Claude, Codex, Cursor, etc.) and any OpenEMIS school — with full read + write access.

OpenEMIS is a free, open-source school management information system developed by UNESCO and KORDIT. It runs the day-to-day administration of every kind of educational institution — kindergartens, primary schools, secondary schools, secondary vocational institutions, technical colleges, and universities — managing students, staff, attendance, assessment, infrastructure, meals, scholarships, examinations, training, and ministry-level reporting. This MCP-pro server adds full read + write access plus per-user authentication on top of any OpenEMIS school.

Translations: Русский · Español · हिन्दी · العربية

Built on top of the published OpenEMIS Core API (reference docs at api.openemis.org/core) and verified end-to-end against the public demo at [demo.openemis.org/core](https://demo.openemis.org/core) with real credentials, real data, real round-trips.

Ask in English:

*"How many current students are at Avory Primary?"*

The agent plans the calls, this MCP delivers the data, and you get the answer:

*"Avory Primary School (code P1002) has 553 currently enrolled students."*

You never write a line of code. You never see JSON. You just ask.

Status: v1.0.0 — full CRUD for non-workflow resources. Read queries work against every OpenEMIS v5 resource. Write tools (create/update/delete) are live for all resources that do not flow through the CakePHP Workflow plugin. Workflow-controlled resources (attendance, staff leave) are blocked at the tool level and redirect to the appropriate playbook.

What this is

openemis-mcp-pro is the read + write MCP server that connects AI assistants to the OpenEMIS school management system. It exposes 678 resources (students, attendance, risks, staff, exams, infrastructure) across 40 curated playbooks — 26 read and 14 write/auth. The pro distribution adds direct write tools (openemis_create, openemis_update, openemis_delete), HTTP server mode for ChatGPT Custom GPT, and per-user authentication on top of the free read-only distribution.

Why this exists

The OpenEMIS Core REST API is large — the v5 surface alone exposes 3,361 endpoints across 678 resources (Core 5.13.0). No AI agent can hold that in context, and raw Swagger-style introspection floods a conversation with noise that has nothing to do with the user's actual question.

This MCP solves that in two ways:

1. Domain-scoped discovery. Instead of dumping the whole API into the agent's context, the openemis_discover(topic) tool narrows to the ~20–30 endpoints relevant to what the user is actually asking about ("attendance", "students", "assessment") — powered by a small curated knowledge pack of Domain-*.md notes.
2. A single, composable getter. One openemis_get tool covers list + singleton + filtered search across every resource. The agent supplies resource + optional id + optional params (_fields, _conditions, orderby, page, limit) and the rest of the OpenEMIS CakePHP-style query DSL maps straight through.

The net effect: agents answer natural-language questions in 2–4 tool calls, not 30.

Tools

A representative natural-language question like *"how many teachers at Avory Primary, how many vacant positions?"* resolves to three openemis_get calls — chained by the agent, narrowed by _conditions, delivered back as a single English answer. A write request like *"enrol a new student"* uses openemis_get_playbook to load the step-by-step guide, then openemis_create for each write step.

Core compatibility

Tested against OpenEMIS Core 5.13.0 (master, June 2026). Earlier 5.7 – 5.12 deployments are also supported — the API surface is backwards-compatible.

Capability flag — POCOR-9660 multi-id GET

openemis_get accepts params.ids = "1,2,3" for batch lookups. Core 5.10+ carries POCOR-9660 (?id=1,2,3 and _conditions= :IN(...) support in CrudApiController), so the handler collapses the batch into a single round-trip by default. Pointing at an older Core (5.7 – 5.9) without the native operator? Force the legacy parallel fan-out:

OPENEMIS_CORE_IN_OPERATOR=off

For composite-PK or view resources — where ids does not apply — use _conditions= :IN(1,2,3) instead; it filters any field by a value list and works regardless of this flag. Filtering on a field that does not exist on a resource now returns HTTP 400 (Core 5.10+, POCOR-9697), so use exact field names.

Verified against demo.openemis.org

Every claim in this README was proven against the public demo instance before being written:

• POST /api/v5/login with { username, password, api_key } → JWT cached, 331 chars
• GET /api/v5/institutions?limit=200&_fields=id,name,code → 24 institutions incl. "Avory Primary School" (id=6, code P1002)
• GET /api/v5/institution-students?institution_id=6&student_status_id=1&limit=1 → pagination reports last_page: 553 → 553 currently enrolled students
• GET /api/v5/academic-periods → 7 pages of real academic-year data
• GET /api/v5/absence-types → EXCUSED, UNEXCUSED, LATE, etc.

The sample scripts/smoke-login.mjs shipped with this repo performs the login test step-by-step so you can confirm reachability against your own instance before wiring it into Claude Code.

Compatible agents

openemis-mcp speaks the Model Context Protocol over stdio — any MCP-compatible client works:

Stdio mode (local machine) — connects as a subprocess:

HTTP server mode (OPENEMIS_TRANSPORT=http, install once on Oracle/VPS) — connects by URL:

Install

Requires Node 22+ (for built-in fetch and AbortController) and Python 3.10+ (for the manifest builder and playbook generator scripts in mcp-openemis-gen/). The MCP server itself is Node-only; Python is only needed if you rebuild the manifest from source.

From GitHub

git clone https://github.com/tixuz/openemis-mcp.git

cd openemis-mcp

npm install npm run build

cp .env.example .env $EDITOR .env

Configure

.env.example documents every variable. At minimum you need the three credentials your OpenEMIS admin issues:

OPENEMIS_BASE_URL=https://demo.openemis.org/core   # or your own instance

OPENEMIS_USERNAME=admin OPENEMIS_PASSWORD=your_password OPENEMIS_API_KEY=your_api_key

Optional

OPENEMIS_TIMEOUT_MS=30000 OPENEMIS_VAULT_PATH=/absolute/path/to/domain-notes OPENEMIS_MANIFEST_PATH=/absolute/path/to/manifest.jsonl

The server logs in lazily on the first authenticated tool call, POSTing to /api/v5/login, parsing the JWT out of data.token, and caching it in memory. On a 401 it re-logs in and retries once.

OPENEMIS_VAULT_PATH points at the folder containing the curated Domain-*.md notes used by openemis_discover. If missing, discovery degrades gracefully to keyword matching against the manifest alone.

OPENEMIS_MANIFEST_PATH points at the JSONL output of the companion builder in ../mcp-openemis-gen/. If absent, the discovery tools return a friendly "manifest not built yet" hint — they don't crash.

Smoke-test reachability

set -a && source .env && set +a

node scripts/smoke-login.mjs

Expected:

[Test] Loading config...

[OK] Config loaded: baseUrl=https://demo.openemis.org/core [Test] Creating client... [OK] Client created [Test] Attempting login... [OpenEMIS] Login successful; cached JWT (331 chars) [OK] Login successful

Register with Claude Code

claude mcp add openemis \

--env OPENEMIS_BASE_URL="https://your-openemis/core" \ --env OPENEMIS_USERNAME="…" \ --env OPENEMIS_PASSWORD="…" \ --env OPENEMIS_API_KEY="…" \ --env OPENEMIS_VAULT_PATH="/absolute/path/to/vault" \ -- node "$(pwd)/dist/server.js"

Verify

claude mcp list | grep openemis

Expected: openemis: node /…/dist/server.js - ✓ Connected

Any new Claude Code session in this project will see all nine tools automatically.

Server mode (Oracle Always Free / any VPS)

Set OPENEMIS_TRANSPORT=http to run as a persistent HTTP server instead of a local subprocess. Install once on your server; every MCP-compatible client (Claude Code, Cursor, Cline, Windsurf) connects by URL.

On your server:

git clone https://github.com/tixuz/openemis-mcp-pro.git

cd openemis-mcp-pro npm install && npm run build cp .env.example .env $EDITOR .env