The missing half of AI-assisted CAP development: putting the compiler in the loop
The author argues @cap-js/mcp-server@cap-js/mcp-serverSAP's MCP server for CAP development, run as cds-mcp: search_model searches the project's compiled CDS model, search_docs searches the CAP documentation indexed as embeddings on the developer's own machine.sap official, cap solves retrieval — letting an agent search CAP docs and the compiled model before writing — but nothing in that loop checks what the agent actually wrote, and built capforgecapforgeAn MCP server that scaffolds CAP and CDS for a coding agent and then holds it to cds compile — the agent does not stop until the generated code builds clean.independent, cap to cover the other half: a local MCP server, no LLM call of its own, whose generatorGeneratorsGetting from nothing to a project that runs, and from a project that runs to one that has the piece you just decided you need. Two different moments: the first is a one-time decision about layout, the second recurs for the life of the project.SAP and community, narrow by design scaffolds a full CDS vertical slice, including Fiori annotationsFiori annotationsDescribing a service so a Fiori client can render it without hand-written UI code — titles, value helps, hierarchies and visibility expressed as model annotations rather than screen logic.SAP and community, unassessed and mock seed dataSeed dataGetting realistic data into a service so it can be developed and demonstrated — initial data, per-tenant content, and believable volumes for testing.SAP and community, narrow by design, then loops cds compile, `cds lint`LintingCatching in the editor what would otherwise be caught at deploy time, or not at all — problems in source code and models and in the project's configuration, found by rules rather than by review.SAP and community, narrow by design and a throwaway SQLite deploy of that data until it is clean, with matching checks for UI5 freestyle views. A cap_project_context tool compiles the real project to CSN so the agent reuses existing names instead of inventing them — the agent access to your codeAgent access to your codeLetting a coding agent read the project or the system you actually have, reach its model, its objects and its checks, and act on it. In file-based environments that works out of the box; complex systems need dedicated access capabilities. Whatever the environment, the agent also needs project- and framework-specific context, so it understands the thing in front of it rather than guessing from what it remembers of its training data. The transport is incidental: an MCP server, a language server or a CLI answer the same question, and what differs between them is how much of the real thing the agent can see. Distinct from MCP servers, which is a protocol adapter on your own running service, and from agent guidance, which is knowledge handed over as files rather than read from the thing itself.SAP and community, unassessed half retrieval alone does not cover.
Where it sits
Who is behind it
Properties
About this work
onedevatlas is open for browsing. Find tools for SAP development, compare their documented capabilities, ask your AI assistant, and join a flight during TechEd week.
Details
Graph
Source
History
- Entered the atlasDiscovered via SAP Community · cap-tag.
For your agent
- Ref
content:the-missing-half-of-ai-assisted-cap-development-putting-the-compiler-in-the-loop- JSON
- /content/the-missing-half-of-ai-assisted-cap-development-putting-the-compiler-in-the-loop.json
- MCP
- Connecting an agent