Tl;dr
Use DeepWiki and Anthropic’s Sonnet 4.5 and Opus 4.1MCP is new
Coral agents connect to their coral servers via MCP (Model Context Protocol). MCP is a protocol for exposing tools and other context-related primitives to LLMs via convenient transports like HTTP. MCP was announced in November 2024, as a consequence it is outside the training window of many LLMs (expect a 1-2 year lag in general). As a consequence, stating that the coral server provides an MCP interface for agents means much more to a human than to an LLM. Coral itself is also too new to be in the training window of many LLMs, so stating that an agent is “Coralized” or “built with Coral” also will not be understood by many LLMs. Unfortunately, many LLMs will hallucinate that they understand MCP and Coral, and will produce code that is syntactically correct but semantically incorrect.Table of models and their awareness of MCP and Coral
Model | MCP Awareness | Coral Awareness | Notes |
---|---|---|---|
Sonnet 4.5 | Yes | No | |
Opus 4.1 | Yes | No | Seems less familiar than Sonnet 4.5 |
Gemini 2.5 Pro | No | No | Training cut-off after MCP and Coral release |
GPT-4 | No | No | Training cut-off before MCP and Coral release |
Llama 3 | No | No | Training cut-off before MCP and Coral release |
Mistral 7B | No | No | Training cut-off before MCP and Coral release |
DeepWiki
DeepWiki provides an LLM-generated documentation of the coral server’s code. It is generally pretty accurate. It can be asked questions directly, or be used as a source of truth for other agents.Prompt engineering
When using LLMs for coding tasks, prompt engineering is crucial. Here are some tips:- Encourage the model to ask questions to you about Coral and MCP if it is unsure.
- Work from high quality up to date examples.
- For agentic coding tools without access to DeepWiki, consider giving them access to the coral server source