Knowledge Graph for Education Global MCP

 

Knowledge Graph for Education Global MCP is a curriculum-agnostic FastMCP server for exploring curriculum knowledge graphs from countries, states, and educational organizations. It provides one interface for discovering frameworks, searching academic standards, navigating curriculum hierarchies, retrieving provenance-aware resources, and assembling deterministic evidence for cross-framework comparison.

Claude Desktop acts as the reasoning and generation layer. The server remains read-only and deterministic: it loads accepted graph packages, validates and indexes them, returns source-grounded evidence, and does not call a server-side LLM.

Architecture

Each curriculum is retained as an independent, immutable, and versioned graph package. A shared catalog and common MCP services make the packages searchable through one server without flattening them into a single source graph.

This preserves:

• exact framework, snapshot, and package identity;
• source terminology and local grade labels;
• provenance, rights, and validation evidence;
• tree and multi-parent DAG topology; and
• unresolved relationship statuses.

The current implementation focuses on Academic Standards graphs and is designed to remain configuration-driven rather than curriculum-specific.

Current MCP surface

The server exposes:

• 19 tools
• 1 fixed resource
• 14 resource templates
• 9 prompts

Tools

• list_frameworks
• get_framework
• search_standards
• get_standard
• get_standard_context
• search_learning_components
• get_learning_component
• get_learning_component_context
• get_learning_components_for_standard
• get_framework_statistics
• get_capabilities
• compare_framework_evidence
• get_learning_progression
• get_standard_progressions
• search_learning_progressions
• traverse_learning_progressions
• get_learning_progression_paths
• read_evidence
• get_workflow_instructions

Prompts

• student_study_support
• teacher_guide_draft
• student_handbook_section
• multigrade_lesson_plan
• learning_progression_teaching_sequence
• learning_progression_support_plan
• learning_progression_curriculum_review
• administrator_alignment_review
• cross_framework_comparison

compare_framework_evidence exposes text matching through flat scalar fields matchMode (tokens or exact_phrase) and matchOperator (all or any). Do not send a nested match object to this tool. The adapter constructs the existing typed internal text-match policy before calling the ordinary comparison service.

Lexical topic discovery

Text search uses exact normalized description tokens or contiguous normalized phrases. It does not stem words or expand synonyms, so fraction and fractions are different queries. For concept discovery, Claude should search the user's original wording first and may then issue at most three separate conservative alternatives, such as a clear singular/plural or retrieved local-terminology variant. Every alternative must preserve the same framework, snapshot, grade, subject, statement-type, grouping, match, and limit settings. Cross-framework alternatives must be applied symmetrically. Each call retains its own bounds, scores, has_more value, and cursor; a zero-match result does not prove curriculum absence. Once a relevant branch is found, bounded hierarchy context is preferred over unlimited synonym generation.

Stored progression queries retrieve buildsTowards and relatesTo evidence in one exact framework/snapshot. The three new prompts support teaching sequences, support planning and curriculum review using standards, relationships, provenance and supporting Learning Components. Grade arrays are entered as JSON arrays at the prompt boundary. Stored edges are IDinsight-generated judgments, without publisher endorsement; confidence is not a calibrated learner-success probability. Missing edges do not prove absence of a pedagogical relationship. See the progression guide and maintainer update process.

Prerequisites

Install:

• uv
• Git
• Python 3.13 through uv
• Claude Desktop for local MCP use

None of this is needed to use a hosted deployment; see Connect to a hosted server.

The project requires Python >=3.13,<3.14.

Install the required Python version:

uv python install 3.13

Node.js and npm are needed only when building an optional .mcpb distribution.

Repository setup

Clone the repository and enter its root:

git clone <repository-url>
cd KGForEdGlobalMCP

Create or update the locked backend environment:

uv --directory backend sync --locked --no-dev

The application resolves the repository-level config/ and data/ directories from its installed source layout. Optional KGFEGMCP_* environment variables can override those paths, but are not required for ordinary repository-local commands.

To install development tooling as well:

uv --directory backend sync --locked --extra dev

Verify the local server

Run the real server through a separate locked STDIO subprocess:

uv --directory backend run --locked --no-dev kgfegmcp-stdio-smoke

The smoke command:

1. starts python -m kgfegmcp.mcpb_server;
2. completes an MCP handshake;
3. verifies the exact 13-tool, 1-resource, 12-template, and 7-prompt inventory; and
4. confirms that the subprocess exits cleanly.

A successful run returns a JSON result with "status": "passed".

The server entry point can also be started directly:

uv --directory backend run --locked --no-dev \
  python -m kgfegmcp.mcpb_server

That command is not an interactive shell. It waits for an MCP client and reserves stdout for protocol traffic.

Connect Claude Desktop

The documented local integration uses Claude Desktop's MCP configuration file. On 2026-10-08 a user connected Desktop on macOS to server 0.4.0 this way and through the .mcpb bundle, with no failures observed; see Use with Claude Desktop and claude.ai.

~/Library/Application Support/Claude/claude_desktop_config.json

Find the absolute paths required by Claude Desktop:

command -v uv
pwd

Add the following entry under the existing top-level mcpServers object. Replace /absolute/path/to/uv and /absolute/path/to/repository with the values from your machine. Preserve unrelated Claude Desktop settings already present in the file.

{
  "mcpServers": {
    "curriculum-knowledge-graph": {
      "command": "/absolute/path/to/uv",
      "args": [
        "--directory",
        "/absolute/path/to/repository/backend",
        "run",
        "--locked",
        "--no-dev",
        "python",
        "-m",
        "kgfegmcp.mcpb_server"
      ],
      "env": {
        "KGFEGMCP_CONFIG_ROOT": "/absolute/path/to/repository/config",
        "KGFEGMCP_DATA_ROOT": "/absolute/path/to/repository/data",
        "KGFEGMCP_ENV": "local",
        "KGFEGMCP_GRAPH_PACKAGES_ROOT": "/absolute/path/to/repository/data/graph_packages",
        "KGFEGMCP_INVALID_PACKAGE_POLICY": "fail",
        "KGFEGMCP_LOG_LEVEL": "INFO",
        "KGFEGMCP_PROFILE_ROOT": "/absolute/path/to/repository/config/profiles",
        "KGFEGMCP_PROMPT_ROOT": "/absolute/path/to/repository/config/prompts",
        "PATHS_PROJECT_DIR": "/absolute/path/to/repository"
      }
    }
  }
}

When the file already contains other keys, replace only the empty mcpServers object or add the curriculum-knowledge-graph member to the existing object.

Validate the JSON:

jq empty "$HOME/Library/Application Support/Claude/claude_desktop_config.json"

Fully quit and reopen Claude Desktop:

osascript -e 'quit app "Claude"'

After restart, enable curriculum-knowledge-graph under Connectors in a new conversation.

A simple first request is:

Use the curriculum-knowledge-graph connector to list all available frameworks.

Connect to a hosted server

The same server can run as a hosted Streamable HTTP service. Clients then connect to a URL instead of starting a local process, and nothing is installed on the user's machine. The project's public deployment is at:

https://kg-for-ed-global-mcp.up.railway.app/mcp

In Claude, add a custom connector in the connector settings and enter that URL, leaving authentication empty. A user connected a claude.ai custom connector to this deployment, running 0.4.0, on 2026-10-08 with no failures observed. In Claude Code:

claude mcp add --transport http curriculum-knowledge-graph https://kg-for-ed-global-mcp.up.railway.app/mcp

The hosted service is built from the root Dockerfile, which bakes config/ and data/graph_packages/ into the image and starts python -m kgfegmcp.http_server. The MCP surface is identical to the local STDIO server. Build and verify it locally with:

docker build -t kgfegmcp-http:local .
docker run --rm --read-only -e PORT=8000 -p 8000:8000 kgfegmcp-http:local

uv --directory backend run --locked --no-dev kgfegmcp-http-smoke \
  --url http://localhost:8000/mcp

See docs/operations/deployment.md for the hosting workflow, rollback, and the access and rights posture of a public endpoint.

Optional MCPB package

The repository can build a deterministic MCP Bundle containing the locked Python metadata, source package, profiles, prompt configurations, and accepted graph packages.

Install the official MCPB CLI:

npm install -g @anthropic-ai/mcpb

Build the bundle:

uv --directory backend run --locked --no-dev kgfegmcp-build-mcpb

The default output is:

dist/kgfegmcp-1.0.0.mcpb

Retain the exact staging directory for review:

rm -rf ./dist/kgfegmcp-stage

uv --directory backend run --locked --no-dev kgfegmcp-build-mcpb \
  --stage-output ./dist/kgfegmcp-stage

Smoke-test the staged runtime:

uv --directory backend run --locked --no-dev kgfegmcp-stdio-smoke \
  --bundle-root ./dist/kgfegmcp-stage

The bundle does not contain a virtual environment or vendored dependencies. Its runtime uses locked uv project metadata and starts the server with:

python -m kgfegmcp.mcpb_server

Claude Desktop custom-extension installation behavior may vary by client build. Manual registration through claude_desktop_config.json is the documented local connection method. A user installed and used the 0.4.0 bundle in Desktop on macOS on 2026-10-08 with no failures observed; other Desktop builds and operating systems are untested.

Useful CLI commands

Show command help:

uv --directory backend run --locked --no-dev kgfegmcp-stdio-smoke --help
uv --directory backend run --locked --no-dev kgfegmcp-http-smoke --help
uv --directory backend run --locked --no-dev kgfegmcp-build-mcpb --help
uv --directory backend run --locked --no-dev kgfegmcp-build-manifest --help
uv --directory backend run --locked --no-dev kgfegmcp-validate-packages --help
uv --directory backend run --locked --no-dev kgfegmcp-verify-evidence-archive --help

A clone holds only the runtime graph packages. Rebuilding or accepting a package also needs the archived evidence, which is git-ignored; see Add or update a framework.

Troubleshooting

Connector does not appear

Confirm that:

• command is the absolute result of command -v uv;
• the repository and backend paths are absolute;
• the JSON passes jq empty;
• Claude Desktop was fully quit and reopened; and
• the repository STDIO smoke still passes.

Inspect Claude's logs on macOS:

find "$HOME/Library/Logs/Claude" \
  -maxdepth 1 \
  -type f \
  -iname '*mcp*' \
  -print

No module named 'mcp.types'

Always launch the server as a module:

python -m kgfegmcp.mcpb_server

Do not execute src/kgfegmcp/mcpb_server.py by filesystem path. Direct file execution can cause the internal kgfegmcp.mcp package to shadow the external MCP SDK package.

Protocol output errors

STDIO stdout is reserved for MCP protocol messages. Application logging must remain on stderr, and production code should not use uncontrolled print() calls.

Project documentation

• AGENTS.md supplies project agent instructions; docs/architecture.md describes the maintained architecture.
• backend/README.md contains backend-specific operational notes.
• packaging/mcpb/README.md describes the MCP Bundle packaging workflow.

License

See LICENSE.