Skip to content

Latest commit

 

History

History

README.md

Rubric

An environment for working with office documents (Word, PowerPoint, PDF) with LLM-based grading.

Build and upload Docker image

Follow the end of main README to log in

cd examples/rubric
docker buildx build --build-context taiga=../../ --push --platform linux/amd64 -t us-east1-docker.pkg.dev/gcp-taiga/shared/rubric:0.0.1 -f Dockerfile .

Running this example

Prerequisites:

First, build the environment:

cd examples/rubric
docker build --build-context taiga=../../ --platform linux/amd64 -t rubric -f Dockerfile .

Start the mcp_client (if not already running):

cd mcp_client

# Install all dependencies (backend + frontend)
make setup

# Build the frontend for production
make build-frontend

# Start the combined server with your API key
ANTHROPIC_API_KEY=your_key make run

Finally, open a new terminal and run:

Note: -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY is needed to run rubric grading

docker_id=$(docker run -d -i -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY rubric uv --offline --directory /mcp_server run rubric mcp) &&
open -a "Google Chrome" "http://localhost:5000/app?container_id=$docker_id&max_tokens=64000&problem_id=rubric-test"

This will open a browser where you can watch Claude solve the rubric-test problem.

Available problems

  • rubric-test - Very simple question, expect short answer
  • office-test - Create and manipulate office documents (DOCX, PPTX files)

Hot-reload for development

If you're modifying the code, you can mount the source directory to avoid rebuilding:

docker_id=$(docker run -d -i -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY -v ./src:/mcp_server/src rubric uv --offline --directory /mcp_server run rubric mcp) &&
open -a "Google Chrome" "http://localhost:5000/app?container_id=$docker_id&max_tokens=64000&problem_id=clauppy-test"

Environment details

This environment includes tools for:

  • Creating and editing Word documents (DOCX)
  • Creating and editing PowerPoint presentations (PPTX)
  • Working with PDF files
  • LLM-based grading for evaluating document quality

Spreadsheet Copilot MCP

This environment contains an alternative MCP server that exposes the Spreadsheet Copilot tools. These tools have the same inputs and outputs as https://github.com/anthropics/spreadsheet-agent/blob/main/shared/src/tools.ts, but are implemented in Python and operate on local excel files. To use this MCP server, build and push as usual, but run with run spreadsheets mcp rather than run rubric mcp. Note that you should not give access to the bash tool in the taiga UI, otherwise Claude can cheat and use the underlying Python libraries rather than the tools to do its work.