Work P3 Open source
jda-mcp
Lets an AI agent query a warehouse management system without being able to break it.
- MCP tools
- 27
- MOCA transports
- 2
- writes by default
- 0
Signal path
- AI agent Cursor, Claude Desktop, VS Code
- jda-mcp 27 tools over stdio
- Read-only guard blocks hidden writes
- MOCA HTTPS or TCP
- WMS Blue Yonder / JDA
The problem
Blue Yonder’s warehouse management system (formerly JDA, and RedPrairie before that) is driven through MOCA, its command language and server. The people who know a warehouse’s data best are often not the people who can write MOCA or SQL against it.
An AI assistant could close that gap. But nobody sensible hands a language model an open connection to a production WMS.
What I built
jda-mcp is a Model Context Protocol server. It gives any MCP-capable client, such as Cursor, Claude Desktop or VS Code, a set of 27 tools for connecting to a MOCA server, running queries, and exploring the schema and the command repository.
It speaks both MOCA transports: XML over HTTP or HTTPS, and the legacy TCP wire protocol.
Safe by default
The interesting engineering is in what the server refuses to do.
- Read-only unless told otherwise. A guard inspects every statement before it is sent. It blocks mutating verbs, and also the ways a write can hide: multi-statement SQL, DML smuggled inside a CTE (
with ... delete),select ... into,execand PL/SQL blocks, and statements tucked behind comments or string literals. - Writes need a human. Write tools exist but are disabled by default. They need per-operation approval through MCP elicitation, or an explicit
--allow-writeflag at startup. - Writes never retry. Read queries log back in and retry once after a dropped socket or an expired session. Writes do not, so nothing can run twice.
- Credentials are encrypted at rest, using Windows DPAPI where it is available and AES otherwise.
Built for a model to use
Results come back as structured content with a declared output schema, so a client does not have to re-parse text. Large results can be paged, and a positional row format uses far fewer tokens than rows of objects, so more data fits in each response.
The server also ships its own reference material and guided prompts for connecting and for exploring a table.