Skip to main content
Glama

ed-tech-system-mcp

Domain-Driven MCP (Model Context Protocol) server for ed-tech workflows. The server exposes validated MCP tools backed by LangGraph agents, web search, and YouTube video discovery — all organized with Clean Architecture and DDD. Document embedding and retrieval live in the backend (ed-tech-system-backend embedding service); the MCP server no longer runs RAG.

What this project does

External MCP clients call MCP tools that validate input with Pydantic, delegate to application workflows and LangGraph agents, and reach external systems through domain ports implemented in the infrastructure layer.

MCP tools

Tool

Module

Purpose

health_check

custom_tools

Liveness probe

search_youtube

custom_tools

Educational YouTube search

build_lesson_enrichment_query

custom_tools

Expand lesson metadata into 4–5 search terms for document/video lookup

research_article

custom_tools_agent_workflows

Research article generation workflow

content_generation

custom_tools_agent_workflows

Lesson/quiz/project content generation

author_lesson_pipeline

custom_tools_authoring

Graph leaf → generate → validate → save (draft/publish)

search_graph_nodes

custom_tools_authoring

Curriculum graph leaf search

validate_lesson / validate_quiz / validate_project / validate_test_boilerplate

custom_tools_authoring

Content validation

save_to_backend

custom_tools_authoring

Persist authored lesson tree via backend RPCs

generate_mock_test_structure / validate_mock_test

custom_tools_authoring

Mock assessment scaffolding

socratic_tutor

custom_tools_socratic

Socratic tutor turn

collect_project_review_context / project_review

custom_tools_project_review

Project mentor review

Handlers use MCP tool caching (when enabled), latency logging, privileged auth where required, and domain error mapping at the protocol boundary.

Integrations

Capability

Integration

Curriculum graph + authoring RPCs

Supabase RPCs via authoring backend client (anon + manager JWT)

Web search

Tavily

Video discovery

YouTube Data API v3

Agent orchestration

LangChain / LangGraph

Caching

Redis when CACHE_ENABLED=true

Document embedding / RAG

Backend ed-tech-system-backend embedding service + mcp-find-documents edge function

Related MCP server: langgraph-clean-arch

Architecture

Clean Architecture under src/mcp_server/. Dependency rule: Domain has no framework I/O; Infrastructure implements Domain ports; Interface and Application depend inward only.

entrypoint → interface → application → domain ← infrastructure

Layer

Path

Responsibility

domain

domain/

Entities, ports, validators, curriculum enums — no MCP/LangGraph/Supabase

application

application/

LangGraph agents/, runners, LLM routing, authoring services

interface

interface/

MCP tools (custom_tools*.py), validation, error mapping

infrastructure

infrastructure/

Supabase, search, YouTube, Groq, Redis, cache adapters

entrypoint

main.py, wiring.py, settings.py, …

Bootstrap and composition root only

Changelog folders use the same names plus tests, performance, code-health, refactor — see ARCHITECTURE.md § Changelog layer names.

Read next: ARCHITECTURE.md (layer rules, tree, anti-patterns) · AGENTIC_ARCHITECTURE.md (graphs / tools) · OBSERVABILITY.md (execution traces).

Quick start

Prerequisites

  • Python 3.12 (see requires-python in pyproject.toml)

  • uv — environment and dependency manager

  • Doppler CLI (recommended for secrets) or a local gitignored .env

Install

uv python install 3.12
uv sync --all-groups

Configure secrets

Secrets never enter git. Use Doppler (team) or a local .env (solo dev).

doppler login
./scripts/doppler/setup-local.sh
./scripts/doppler/bootstrap-from-env-example.sh   # first time only — uploads placeholders
# Fill real values in the Doppler dashboard → ed-harness-system

Required variables: APP_ENV, SUPABASE_URL, SUPABASE_SERVICE_ROLE_KEY, YOUTUBE_API_KEY.

Optional: GROQ_API_KEY (only when an LLM path is invoked — lazy-init at first use), TAVILY_API_KEY, LOG_LEVEL (applied at bootstrap via configure_logging()). Staging/production also require CACHE_ENABLED=true and REDIS_URL (local/CI keep the default off).

See ENVIRONMENT_SETUP.md for the full secrets workflow.

Run the MCP server

# With Doppler
doppler run -- uv run mcp-server

# With local .env (APP_ENV=development)
uv run mcp-server

Inspect traces in tests

Traces are captured programmatically by workflow_trace.py and workflow_llm_trace.py and asserted in pytest. See OBSERVABILITY.md for trace field details and debugging patterns.

Development

Day-to-day commands

uv sync --frozen              # after pulling lockfile changes
uv run mcp-server             # start server
uv run ruff check src/        # lint
uv run ruff format --check src/
uv run mypy src/              # type check
uv run pytest                 # tests (143 cases as of 2026-07-21)

Engineering backlog

Audit findings are triaged into backlog/BACKLOG.md (RICE-ranked, traceable to changelog audits). As of 2026-07-21: 23 done, 6 deferred (adapter HTTP implementation, profiling, trace IDs).

Add dependencies

uv add some-package           # runtime
uv add --group dev some-tool  # dev only

Do not use pip install in this repo — it bypasses the lockfile.

Quality gates (CI parity)

uv sync --frozen --all-groups
uv run ruff check src/
uv run mypy src/
npm run lint:architecture   # layer imports + boundary patterns (also runs on git push)
uv run pytest

Git hooks: Husky pre-commit runs public-repo safety checks (sensitive files, tracked leaks, secret scanners); pre-push re-checks tracked safety, scans pushed commits for secret content, then runs architecture lint — neither blocks the other tier.

Run quality-gate commands from the repository root (ed-tech-system-mcp/), not ui/. The same scripts are also available inside ui/ via npm run hooks:test and npm run lint:architecture.

Project layout

.
├── src/mcp_server/
│   ├── domain/              # Entities, ports, validators, enums
│   ├── application/         # Agents, runners, LLM routing, authoring services
│   │   └── agents/          # LangGraph packages (content_generation, socratic, …)
│   ├── interface/           # MCP tools + validation
│   ├── infrastructure/      # Adapters (search, video, LLM, cache, clients)
│   ├── wiring.py            # Composition root
│   ├── settings.py
│   └── main.py              # mcp-server entry
├── tests/                   # pytest + architecture lint
├── changelog/               # Agent memory: {DATE}/{LAYER}/ (local)
├── scripts/                 # Doppler, hooks, Render, dev helpers
├── docs/assets/             # README screenshots
├── ARCHITECTURE.md          # Layer boundaries (canonical)
├── AGENTIC_ARCHITECTURE.md  # Agent graphs and tool orchestration
├── OBSERVABILITY.md         # Workflow UI, trace replay
└── ENVIRONMENT_SETUP.md     # uv, secrets, CI, MCP client config

Documentation index

See the documentation matrix at the end of this file for canonical docs and changelog artifacts.

MCP client integration

Register the server in your MCP host using the project interpreter:

{
  "mcpServers": {
    "ed-tech-system": {
      "command": "doppler",
      "args": ["run", "--", "uv", "--directory", "/absolute/path/to/ed-tech-system-mcp", "run", "mcp-server"]
    }
  }
}

Alternative patterns (local .env, uv launcher) are in ENVIRONMENT_SETUP.md § MCP client integration.

Documentation matrix

Read the minimum doc set for your task. Do not load everything.

Canonical docs (repo root)

Document

Read when

README.md

First visit — overview, quick start, MCP tools

ARCHITECTURE.md

Any code change — layers, ports/adapters, deps per layer, file layout, anti-patterns

AGENTIC_ARCHITECTURE.md

LangGraph/LangChain agents, LLM wiring, tool taxonomy, DB/web/video flows

OBSERVABILITY.md

Execution traces and debugging

ENVIRONMENT_SETUP.md

uv, lockfile, deps, env vars, ruff/mypy/pytest, CI, MCP client setup

Conflict resolution: ARCHITECTURE.md wins on layer boundaries; AGENTIC_ARCHITECTURE.md wins on orchestration semantics.

Engineering backlog

Document

Read when

backlog/BACKLOG.md

RICE-ranked tasks from audits; status tracking

backlog/RICE.md

Scoring rubric and priority formula for backlog items

Changelog memory (changelog/{DATE}/{LAYER}/)

Local engineering memory (often gitignored). {LAYER} must match an architecture layer or audit folder:

domain · application · interface · infrastructure · entrypoint · tests · performance · code-health · refactor

File pattern

Purpose

INVESTIGATION{N}.md

Scope and gaps before coding

IMPLEMENTATION{N}.md

Execution checklist and status

CODE_REVIEW{N}.md

Pre-merge review findings

TEST{N}.md

Behavior catalog before writing tests

HOMOLOGATION.md

Coverage verdict after tests pass

PERFORMANCE_AUDIT{N}.md

Performance bottleneck findings

CODE_HEALTH_AUDIT{N}.md

Maintainability / dead-code findings

REFACTOR{N}.md

Merged refactor actions from audits

LOOP_BREAK{N}.md

Recursive-loop parameters and iteration log

Pairing: IMPLEMENTATION{N} ↔ INVESTIGATION{N}; CODE_REVIEW{N} ↔ same {N}. Full protocol: .cursor/rules/changelog-agent-memory.mdc.

Quick routing

Code in a layer?        → ARCHITECTURE.md (+ AGENTIC_ARCHITECTURE.md if agents/tools/LLM)
Traces / debugging?     → OBSERVABILITY.md
Environment / CI?       → ENVIRONMENT_SETUP.md
Secrets / Doppler?      → ENVIRONMENT_SETUP.md § Secrets & safety
Tests / merge gate?     → pytest + quality gates in ENVIRONMENT_SETUP.md
Audits / cleanup?       → backlog/BACKLOG.md

License

MIT

Available Tools

4 tools
find_documentsB

Retrieve educational documents enriched with complementary videos.

Inputs: query (required), video_limit, document_limit. Returns: Validated output for document + video discovery tool calls.

health_checkA

Verify the MCP server is running.

Inputs: no named parameters. Returns: result

run_workflowC

Execute the document + video discovery LangGraph workflow.

Inputs: query (required), video_limit, document_limit. Returns: Validated output for LangGraph workflow execution.

search_youtubeC

Search for educational YouTube videos matching a query.

Inputs: query (required), language, max_results, safe_search. Returns: Validated output for video search tool calls.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3/5.0

    Scored across 4 tools

    Disambiguation4/5

    Tools are distinct: health_check is server status, search_youtube targets only videos, find_documents focuses on documents with videos, and run_workflow executes a combined workflow. Minimal overlap.

    Naming Consistency4/5

    All tools use snake_case with a verb_noun pattern, though health_check is noun_verb. The pattern is consistent and readable.

    Tool Count2/5

    With 4 tools, the server feels too sparse for an 'ed-tech system.' The core functionality (search, documents, workflow) is minimal, leaving many expected operations uncovered.

    Completeness2/5

    The server lacks CRUD operations for resources, user management, and other typical ed-tech features. The tool set is focused on discovery only, which is incomplete for the broad server name.

    Related MCP Connectors

    Related MCP Servers