AI Assistant Integration

TLang ships a Model Context Protocol server that gives AI coding assistants — Claude, Cursor, GitHub Copilot, and others — full access to TLang's compiler, runtime, and embedded documentation. The MCP server turns any MCP-compatible assistant into a TLang-aware pair programmer.

Starting the MCP Server

      tlang mcp-server

The server communicates over stdin/stdout using JSON-RPC 2.0, following the MCP 2024-11-05 specification. Connect it to your AI assistant using the assistant's MCP configuration. Example configuration for Claude Desktop (claude_desktop_config.json):

      {
        "mcpServers": {
          "tlang": {
            "command": "tlang",
            "args": ["mcp-server"]
          }
        }
      }

Available Tools

The MCP server exposes a set of tools that mirror the TLang CLI. The assistant can call these tools directly during a conversation:

tlang_init — initialise a new TLang project with a manifest.yml
tlang_compile — compile a .tlang file to bytecode without running it
tlang_run — run a pre-compiled .tlang bytecode file
tlang_both — compile and run in a single step (the typical workflow)
tlang_package — package a project into a .tbag archive for publishing
tlang_clean — remove compiled bytecode caches
tlang_index — index the current project for symbol search
tlang_topic — retrieve embedded documentation on a specific TLang topic

When a project context is active, the server also exposes dynamic tools derived from the project's exposed functions, allowing the assistant to call specific generators directly.

Resources

The MCP server exposes embedded documentation as resources the assistant can read to understand TLang before generating code:

tlang://guide — the full TLang onboarding guide
tlang://project — project structure and manifest reference
tlang://lang — language syntax reference
tlang://templates — template block reference (lang, doc, style, data, cmd, raw)
tlang://helpers — helper function patterns
tlang://builtins — standard library module reference
tlang://patterns — idiomatic TLang patterns
tlang://mistakes — common mistakes and how to avoid them

These resources are served directly from the binary — no internet connection required.

System Prompt

The MCP server provides a ready-made system prompt for AI assistants. When connected, the assistant automatically receives context about:

TLang syntax and semantics
Template types and their calling conventions
The set + template binding pattern
The standard library API
Common pitfalls (StringBuilder.build vs toString, etc.)
How to write idiomatic TLang code

This eliminates the need to paste TLang documentation into every conversation. The assistant starts with full context.

AI-Native Workflow

The intended workflow with TLang and an AI assistant:

Describe your model to the assistant (entities, fields, relationships)
The assistant writes the TLang templates using set declarations and template blocks
The assistant calls tlang_both to compile and run the templates
TLang generates all target files — the assistant reviews the output
You add more entities or refine the templates — the assistant regenerates

The AI does the creative work (understanding your domain and writing the template) once. TLang handles the mechanical work (generating the same pattern across all entities) at zero additional AI cost. A project with 200 entities costs the same in AI tokens as a project with 2 — because the AI writes one template, not 200 files.

Project Context

Use the set_project_context tool to point the MCP server at a TLang project:

      # In the assistant conversation:
      # "Set the project context to my-app/"

Once set, the server indexes the project and exposes its exposed functions as dynamic tools. The assistant can then call your project's generators directly without writing any CLI commands.