CAPforge

MCP server that makes your AI agent (Claude Code, Cursor, …) get SAP CAP/CDS right: scaffolding that follows real conventions, plus real validation against cds compile. The agent doesn't stop until the code compiles clean.

Website · npm · Leelo en español

It calls no LLM itself: your agent brings the model (BYOK). Zero inference cost, nothing leaves your machine.

Why

General-purpose models hallucinate CDS — annotation terms that don't exist, OData v2 syntax mixed into v4, invented fields — because there's barely any public CAP code in training data. CAPforge gives your agent the two things it's missing:

1. Curated, sourced knowledge — modern CAP/UI5 best practices distilled from official SAP docs (capire, OData vocabularies, reference samples), every rule with its source URL.
2. A real feedback loop — tools that run the actual compiler and linter and feed errors back until everything is green.

Tools

| Tool | What it does | |---|---| | cap_project_context | Codebase-aware: compiles your real project to CSN and returns the existing model (services, entities, fields, associations) so the agent reuses instead of inventing | | cap_scaffold_entity | Generates all 5 layers of an entity following conventions: CDS model with aspects (cuid, managed), projection, annotations.cds (HeaderInfo, LineItem, Facets, FieldGroups), mock CSV, i18n | | cap_write_files | Writes files into the project | | cap_validate | Runs cds compile and returns errors (the loop) | | cap_lint | Runs cds lint (best practices / anti-patterns) | | cap_deploy_check | Layer 3: deploys to a throwaway sqlite DB and loads the mock CSVs — catches what compile can't (CSV columns that don't exist, bad data, duplicate keys) | | cap_analyze_legacy_ui5 | Migration mode: inventories a legacy freestyle UI5 app (read-only) — datasources, sections, columns, bound fields, value helps, actions, i18n | | ui5_scaffold_section | Generates a freestyle ObjectPage section (SmartTable with p13nData, or SimpleForm) + handler stubs + i18n | | ui5_validate_view | The UI5 loop: well-formed XML + every declared event handler exists in the controller | | ui5_lint | Layer 3: runs SAP's official ui5lint — deprecated APIs, UI5 2.x readiness, CSP issues | | resource conventions://cap | Canonical knowledge core + your team's convention profile + migration map |

Normal flow: context → scaffold → write → validate → correct → until green. Migration flow: analyze_legacy → context (target) → propose map (REUSED / NEW / DISCARDED) → scaffold → validate.

Requirements

• Node.js 18+
• @sap/cds-dk available (globally or in the project) for cap_validate to work.

Install & try

npm install
node test-loop.js   # full-loop demo against the bundled sandbox

VS Code / SAP Business Application Studio

Install the "CAPforge" extension from the VS Code Marketplace (also on Open VSX): it registers the MCP server automatically for Copilot agent mode, no mcp.json editing.

Point the capforge.conventionsPath setting to your team's private conventions. BAS note: BAS installs extensions from Open VSX, but its AI assistant does not consume MCP servers natively yet — the extension is ready for the day it does.

Claude Code / Cursor

Register in Claude Code (project .mcp.json or global config):

{
  "mcpServers": {
    "capforge": {
      "command": "npx",
      "args": ["-y", "capforge"]
    }
  }
}

Load your team's private conventions (optional):

"env": { "CAPFORGE_CONVENTIONS": "PATH/to/team-conventions.md" }

Make your agent actually use it

Registering an MCP server does not force the agent to call it: agents only invoke tools when they decide the task needs them, and for CAP they often just write CDS from memory. Two ways to fix that:

1. Workspace instructions (recommended). Add this to your project's agent instructions file — .github/copilot-instructions.md (Copilot), CLAUDE.md (Claude Code) or .cursorrules (Cursor):

For every SAP CAP or UI5 task: call cap_project_context before generating and reuse existing names; scaffold with cap_scaffold_entity / ui5_scaffold_section; validate with cap_validate, cap_lint and cap_deploy_check (UI5: ui5_validate_view, ui5_lint) and correct until everything is green. Never hand-write CDS without a clean cap_validate run.

2. The cap_workflow prompt. CAPforge ships an MCP prompt with the same workflow: pick it from your client's prompt list (in VS Code: type / in chat) at the start of a CAP session.

You can also just ask explicitly ("scaffold this with the capforge tools and validate until clean") — but the instructions file is what makes it automatic for the whole team.

Knowledge

• knowledge/ — canonical core: modern CAP/UI5 best practices (aspects, compositions, drafts, authorization, Fiori annotations, i18n, tooling, migration map), anchored to official SAP docs with cited sources.
• conventions/ — house profile: one team's style, overridable via CAPFORGE_CONVENTIONS. Wins over the core on style conflicts.

The conventions://cap resource serves both, combined, to the agent.

Community

The roadmap is decided by the people who build CAP every day: roadmap discussion. Contributions welcome — see CONTRIBUTING.md.

License

MIT.