The heart of Graffiticode is a collection of task-specific languages. Each language is a specialized generative toolchain that converts natural language requests into formal language programs that are compiled to computing artifacts. Each language has its own capabilities, constraints, internal representation, and rendering behavior. Graffiticode combines the best of LLM creativity and programming language verifiability.
- Client agents connect to graffiticode mcp server and load the graffiticode skills
- Graffiticode MCP server provides the following tools:
- list_languages: dynamic list of languages updated in real-time
- get_language_info: precise description of the capabilities and scope of each language
- create_item: creates an item based on an english intended outcome and returns an item ID
- update_item: updates an item based on an english description of intended outcome
- render_item: renders an item form view
- get_item: returns the internal JSON representation of an item
- get_spec: generates a specification of an item detailed enough to recreate it anew
- Micro-agent uses language specific training data to route and translate the intent into a program that can be compiled
- Router reads the capabilities and scope of each language and identifies a suitable language to complete the intended task. If no such language exists, the router rejects the request
- Code Generator, using RAG and other prompt optimization techniques, translates the request into code in the chosen dialect
- Compiler first checks the code for valid semantics providing the code generator with information on errors found, then compiles the code into a data model used to instantiate the task artifact. The compiler may call out to external web services for side effects (read / write data) related to the task
- Task Store contains items indexed by a hash of the code
- Compiler validates the code and transforms it into a task data model. Errors are fed back to the code generator. Data is stored in the task
- Artifact is instantiated using the compiled data model