Skip to content

Test MCP implementation with different LLMs and improve accordingly #26

@esgn

Description

@esgn

Context

This is a task we will repeat at different stages

  • Some models will not have access to MCP resources
  • Some small models will struggle with complex behavior
  • Some environments implement only part of the MCP standard or with specifics
  • MCP tools description must be tested with different models and different environments

Therefore we must test and improve accordingly.

Actions

  • Prepare work in a dedicated repository : ignfab/geocontext-test
  • Allow the creation of a langgraph agent for test purpose
  • Create the first tests relying on the agent
  • Integrate ignfab/geocontext-test in geocontext (npm run test:e2e)
  • Archive ignfab/geocontext-test
  • Add missing dependency for google models ( npm install --save-dev @langchain/google-genai )
  • Define a list of model to test (see bellow, in progress)
  • Run npm run test:e2e for each model and produce a report

Models

  • anthropic:claude-haiku-4-5
  • anthropic:claude-sonnet-4-6
  • google-genai:gemini-2.5-flash (WARNING : google_genai with an underscore in python)
  • google-genai:gemini-3.1-flash-lite-preview
  • ollama:qwen3.5:9b : too small to chain tool calls
  • ... TO BE COMPLETED at least with some mistral model ...

Linked issues

Metadata

Metadata

Assignees

Labels

testingUnit or integration testing related issue

Type

No fields configured for Task.

Projects

No projects

Relationships

None yet

Development

No branches or pull requests

Issue actions