PowerDesigner MCP
This server lets AI agents drive PowerDesigner end-to-end through MCP: design, create, inspect, validate, convert, and generate SQL for database models (and falls back to an in-memory mock when PowerDesigner is absent).
Model management: create/open/save/save-as/close/list/inspect PDM, CDM, and LDM models, plus packages and domains.
Schema inspection: list/search tables, columns, keys, indexes, references, relationships; get full object details; take summary/detail snapshots; inspect whole schema; compare two models.
Schema modifications: create/rename/update/delete tables, columns, primary keys (single & composite), references with automatic FK-column migration, and indexes (normal/unique/composite).
High-level automation: create an entire database schema from one JSON spec, apply atomic structured schema patches, or materialize a full design spec into a new/existing model.
Validation & rules: run structural model checks (
validate_model) and customizable design-rule checks such as naming patterns, required columns, PK/FK/index naming, allowed data types, and non-empty comments.DDL & conversion: generate native PowerDesigner SQL DDL, and convert CDM → LDM → PDM.
Safety & transactions: every mutating tool supports
dry_run; batch tools are atomic with rollback; explicit begin/commit/rollback transactions, model backups, and restore from backup.Operational/resource support: server info, open-model listing, MCP resources and prompts for database design workflow and schema review.
Enables generation of native MySQL DDL scripts from PowerDesigner models and conversion of conceptual/logical models to MySQL physical data models.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@PowerDesigner MCP Generate MySQL DDL for all tables in the PDM"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
PowerDesigner MCP
Let AI agents operate PowerDesigner like a professional database modeling engineer.
English | 中文
An MCP (Model Context Protocol) server that drives Sybase PowerDesigner 16.5
through its official COM Automation API — never by touching .pdm binaries —
covering the full database course-design loop:
requirements → data dictionary → CDM → E-R → LDM → PDM
→ constraints → indexes → DDL (SQL) → model checks → saveVerified end-to-end on a real PowerDesigner 16.5 installation: model creation, tables, columns, primary keys (single & composite), foreign keys with automatic FK-column migration, indexes, native MySQL DDL generation, validation, and save-as — 42 unit tests + live acceptance tests, all passing.
Why this exists
AI coding assistants can reason about database design, but they cannot touch PowerDesigner. This MCP server is the missing execution layer:
The AI does the design reasoning. The server does reliable execution.
55+ granular tools (not one giant tool), plus high-level batch operations.
Safety first: every mutating tool supports
dry_run; batch tools are atomic with automatic rollback; file-level transactions restore the exact pre-transaction state.Graceful degradation: without PowerDesigner installed, the server falls back to an in-memory mock backend with identical tool semantics — so the whole pipeline is testable anywhere.
Related MCP server: solidworks-mcp
Architecture
MCP Client (Cursor / Claude Desktop / Claude Code)
│ MCP stdio (JSON-RPC)
▼
FastMCP Server ──── Tools (55+) / Resources / Prompts
▼
Services (schema orchestration · validation engine · inspection · transactions)
▼
PowerDesignerAdapter (interface)
├── ComAdapter ── single-threaded STA COM dispatcher ── PowerDesigner 16.5 COM
└── MockAdapter ── in-memory model store (tests / PD-less development)Adapter isolation means version differences (16.x / 17.x) stay inside the COM
layer; every assumption is verified against the vendor's own constants file,
.NET interop metadata, and live experiments — see
docs/com-api-notes.md.
Quick start
Install — one command, no path editing
git clone https://github.com/NovolineGast/powerdesigner-mcp.git
cd powerdesigner-mcp
.\install.ps1That is the whole install. There is no path to type and no client JSON to hand-edit:
The package is installed as a tool —
uv tool install --editable .putspowerdesigner-mcpin an isolated environment with a launcher onPATH(without uv it falls back to an editable install in.venv).The client config is written for you — the installer then runs
powerdesigner-mcp install, resolves the launch command for this machine and merges it into every MCP client it detects:%USERPROFILE%\.workbuddy\mcp.json,%APPDATA%\Claude\claude_desktop_config.json,%USERPROFILE%\.cursor\mcp.json, andclaude mcp addwhen that CLI is present. A replaced config is backed up next to the original asmcp.json.bak-<timestamp>.The result is verified — a real PowerDesigner COM probe and a stdio smoke test run on the exact command that was just registered, so the entry is proven to launch rather than merely written.
Already installed? Re-register at any time (after moving the project, changing clients, etc.):
powerdesigner-mcp install # detected clients
powerdesigner-mcp install --client workbuddy
powerdesigner-mcp install --dry-run # preview, write nothing
powerdesigner-mcp install --print-only # print the entry JSONWhy this needed absolute paths before: the project used to be installed as dependencies only into a
.venv, so every client entry had to carry the venv's interpreter path and aPYTHONPATHpointing atsrc. Installing the package itself removes both, andinstallcomputes whatever remains by itself — including the uv tool launcher when the host does not inheritPATH.
Even shorter: no clone needed
# after the first PyPI release
uvx powerdesigner-mcp install
# straight from the repository (uv must be able to reach GitHub - see note)
uvx --from git+https://github.com/NovolineGast/powerdesigner-mcp powerdesigner-mcp installuvx runs in a throw-away environment, which would produce a config pointing
into a cache that gets pruned — so install detects that, re-installs the
package as a proper uv tool (reusing the recorded source: PyPI name + version,
git URL + commit, or archive URL) and registers that launcher instead.
Some networks (TLS-inspecting proxies) block uv's git transport while the
gitCLI still works; uv then fails without an error message. If that happens, install from a checkout instead:git clone … ; uv tool install --editable . ; powerdesigner-mcp install.
Manual / other platforms
uv tool install --editable . # recommended: launcher on PATH
powerdesigner-mcp install # writes the client configs
powerdesigner-mcp probe # COM capability check
powerdesigner-mcp serve # start MCP server (stdio)
# without uv:
py -3 -m venv .venv
.venv\Scripts\python.exe -m pip install -e .
.venv\Scripts\python.exe -m pd_mcp installMCP client configuration
You normally never write this by hand — powerdesigner-mcp install does it.
Here is what it produces, for reference:
{
"mcpServers": {
"powerdesigner": {
"command": "C:\\Users\\<you>\\.local\\bin\\powerdesigner-mcp.exe",
"args": ["serve"],
"env": { "PDMCP_ATTACH_MODE": "auto", "PDMCP_DEFAULT_DBMS": "MySQL 5.0" }
}
}
}Same structure as above, under the mcpServers key.
claude mcp add powerdesigner -- "$env:USERPROFILE\.local\bin\powerdesigner-mcp.exe" serveSame structure as the Claude Desktop example. Then open Connector management →
Custom connectors (top-right) → Trust the powerdesigner server; new MCP
servers do not activate automatically.
If your client cannot pass env at all, point command at
scripts\powerdesigner-mcp.cmd (portable wrapper, leave args empty) — the
installer also accepts an explicit launcher:
powerdesigner-mcp install --command <path>.
Tool catalog (55+)
Group | Tools |
Model management |
|
Read / inspect |
|
Mutations (all support |
|
High-level automation |
|
Validation loop |
|
DDL & conversion |
|
Transactions |
|
Resources: powerdesigner://models, powerdesigner://model/{id},
.../tables, .../table/{code}, .../relationships, .../indexes
Prompts: database_design_workflow (full course-design loop), schema_review
The course-design workflow
1. get_server_info # verify connection
2. design_from_spec(kind=CDM) # AI designs entities/relations → materialized
3. convert_cdm_to_ldm # resolve M:N
4. convert_cdm_to_pdm(dbms="MySQL 5.0") # physical model
5. create_reference / create_index # FKs & indexes (dry_run first)
6. validate_model # fix every error
7. check_database_design(rules) # your course naming/comment rules
8. fix → re-validate # closed loop
9. generate_ddl # native SQL
10. save_model # save .pdmSafety model
Mechanism | Behavior |
| Returns the exact execution plan, touches nothing |
| Atomic: any failure rolls the model back |
| Saves + backs up the model file |
| Restores the exact pre-transaction file and reopens |
No file backup + destructive op | Explicit error — never fails silently |
Configuration
Env var | Default | Description |
|
|
|
|
|
|
| auto-detect | Path to |
|
| Show PowerDesigner window when we launch it |
| PD default | e.g. |
|
| Per COM-call timeout (seconds) |
Full list in src/pd_mcp/config.py. A JSON config file
(pdmcp.json) is also supported.
Development & testing
.venv\Scripts\python.exe -m pytest tests -m "not live" # unit/smoke tests (mock backend)
$env:PDMCP_LIVE = "1"
.venv\Scripts\python.exe -m pytest tests -m live # live acceptance (real PowerDesigner)
.venv\Scripts\python.exe -m pd_mcp probe # standalone COM capability probeThe mock backend mirrors PowerDesigner semantics (PK via Primary flags,
reference FK migration, index column binding) so the full pipeline — including
schema orchestration, validation, transactions and DDL — is tested without a
PowerDesigner license.
CI runs the suite on windows-latest (Python 3.10 and 3.13) and separately
builds the sdist/wheel, validates the metadata with twine check --strict,
asserts the wheel carries the server and its entry point, and installs it before
running the CLI. Releases are cut by pushing a v* tag — see
docs/RELEASING.md
for the one-time PyPI Trusted Publisher setup.
Verified environment & honest limitations
Verified: PowerDesigner 16.5.0.3982 on Windows 10/11, 64-bit Python 3.13. COM facts verified against the vendor's
VBScriptConstants.vbs,Interop.*.dllmetadata, official C# sample and live probes — see docs/com-api-notes.md.CheckModel()returnsNoneon 16.5 (results go to PD's Result List window); machine-readable findings come fromvalidate_model/check_database_design.PowerDesigner has no SaveAs: saving an unsaved model to a new path is implemented via ShellNew-template file binding + content copy (returns a
copiedsummary). Diagram symbols are re-attached and auto-laid-out as part of the copy, and the saved file is verified to hold the model rather than the ShellNew stub (PD 16.5 sometimes writes to an extension-less sibling).CDM/LDM are first-class: entities/attributes, identifiers (primaries), and conceptual relationships with per-end cardinalities and a dependent side. Indexes are a physical concept, so index tools reject a CDM/LDM explicitly.
CDM inheritance is not exposed yet; the adapter interface makes adding it straightforward.
Generating for a DBMS different from the model's DBMS returns a clear error — use
create_model(dbms=...)/ChangeDBMSfirst.
License
PowerDesigner MCP(中文文档)
让 AI Agent 像专业数据库建模工程师一样操作 PowerDesigner。
English | 中文
一个 MCP(Model Context Protocol)服务器,通过 PowerDesigner 的官方 COM
Automation API(绝不直接修改 .pdm 二进制)驱动 Sybase PowerDesigner 16.5,
覆盖数据库课程设计完整闭环:
需求分析 → 数据字典 → CDM → E-R 模型 → LDM → PDM
→ 完整性约束 → 索引 → DDL(SQL) → 模型检查 → 保存模型已在真实 PowerDesigner 16.5 环境端到端验收:建模、表、列、主键(单列/联合)、 外键(自动迁移 FK 列)、索引、原生 MySQL DDL 生成、模型校验、另存为—— 42 项单元测试 + 真机验收测试全部通过。
中文目录
为什么需要它
AI 助手擅长数据库设计推理,却无法"触碰" PowerDesigner。本 MCP Server 是缺失的 执行层:
AI 负责设计推理,服务器负责可靠执行;
55+ 细粒度工具(拒绝单一巨型工具)+ 高层批量编排;
安全第一:所有修改工具支持
dry_run;批量操作原子执行、失败自动回滚; 文件级事务可精确还原事务前状态;优雅降级:未安装 PowerDesigner时自动切换语义一致的内存 Mock 后端, 整条流水线在任何机器上均可测试。
架构
MCP 客户端(Cursor / Claude Desktop / Claude Code)
│ MCP stdio(JSON-RPC)
▼
FastMCP Server ──── Tools(55+)/ Resources / Prompts
▼
Services(schema 编排 · 校验规则引擎 · 检查 · 事务)
▼
PowerDesignerAdapter(抽象接口)
├── ComAdapter ── 单线程 STA COM 调度 ── PowerDesigner 16.5 COM
└── MockAdapter ── 内存模型(测试 / 无 PD 环境)Adapter 隔离使版本差异(16.x/17.x)被限制在 COM 层内;所有 API 假设均经厂商 常量文件、.NET Interop 元数据与真机实验验证——见 docs/com-api-notes.md。
快速开始
git clone https://github.com/NovolineGast/powerdesigner-mcp.git
cd powerdesigner-mcp
.\install.ps1 # 一条命令:装包 + 注册客户端 + 真机 COM 探测 + 冒烟测试不需要手写任何路径,安装脚本自己完成三件事:
把包装成工具(
uv tool install --editable .):独立环境 + PATH 上的 启动器;没有 uv 时退回项目.venv可编辑安装。自动写客户端配置:随后执行
powerdesigner-mcp install,自行解析本机 启动命令,并合并进它探测到的所有客户端配置(WorkBuddy / Claude Desktop / Cursor,以及存在claudeCLI 时的claude mcp add);被替换的配置会备份为mcp.json.bak-<时间戳>。验证配置真能启动:真机 COM 探测 + 用刚刚注册的那条命令跑 stdio 冒烟 测试,而不是只把 JSON 写进去。
已装过?随时重新注册(搬目录、换客户端之后):
powerdesigner-mcp install # 自动探测客户端
powerdesigner-mcp install --client workbuddy
powerdesigner-mcp install --dry-run # 只预览,不写文件
powerdesigner-mcp install --print-only # 打印配置 JSON以前为什么要手写路径:项目过去只把依赖装进
.venv,包本身没装, 于是每个客户端条目都必须带 venv 解释器路径 加 指向src的PYTHONPATH。 现在把包装成工具后两者都不需要,剩下的路径由install自己算——包括宿主不 继承PATH时改走 uv 工具启动器。
手动安装 / 其他平台:
uv tool install --editable .
powerdesigner-mcp install
powerdesigner-mcp probe # COM 能力探测
powerdesigner-mcp serve # 启动 MCP Server(stdio)
# 不用 uv 时:
py -3 -m venv .venv
.venv\Scripts\python.exe -m pip install -e .
.venv\Scripts\python.exe -m pd_mcp install客户端配置样例(正常无需手写,由 powerdesigner-mcp install 生成)见上文英文部分。
连 clone 都不需要的一条命令:
# PyPI 首次发布之后
uvx powerdesigner-mcp install
# 直接来自仓库(要求 uv 能访问 GitHub,见下方提示)
uvx --from git+https://github.com/NovolineGast/powerdesigner-mcp powerdesigner-mcp installuvx 跑在一次性环境里,写进配置的路径会被缓存清理掉;install 会识别这一点,
按记录的来源(PyPI 名称+版本 / git 地址+提交 / 归档地址)把包重装成常驻的
uv tool,再注册那个启动器。
少数网络(会做 TLS 拦截的代理)会让 uv 的 git 传输失败而
git命令行仍可用, 此时 uv 不打印任何错误。遇到这种情况请改用本地检出:git clone … ; uv tool install --editable . ; powerdesigner-mcp install。
发布流程(打标签自动发 PyPI)见 docs/RELEASING.md。
工具目录
分组 | 工具 |
模型管理 |
|
读取/检查 |
|
修改类(全部支持 |
|
高层自动化 |
|
校验闭环 |
|
DDL 与转换 |
|
事务 |
|
Resources:powerdesigner://models、powerdesigner://model/{id} 及 tables /
table / relationships / indexes 子资源 Prompts:database_design_workflow(课程设计全流程)、schema_review
课程设计工作流
1. get_server_info # 确认连接
2. design_from_spec(kind=CDM) # AI 设计实体/联系 → 一次性物化
3. convert_cdm_to_ldm # 概念 → 逻辑(拆解 M:N)
4. convert_cdm_to_pdm(dbms="MySQL 5.0") # 逻辑 → 物理
5. create_reference / create_index # 补外键与索引(先 dry_run)
6. validate_model # 修复全部 error
7. check_database_design(rules) # 课程规范(命名/注释/PK/必填列)
8. 修复 → 复检 # 闭环
9. generate_ddl # 生成 SQL
10. save_model # 保存 .pdm安全模型
机制 | 行为 |
| 只返回执行计划,不触碰模型 |
| 原子执行:任一步失败自动回滚 |
| 先保存并备份模型文件 |
| 精确还原事务前文件并重新打开 |
无文件备份 + 破坏性操作 | 明确报错——绝不静默失败 |
配置
环境变量 | 默认 | 说明 |
|
|
|
|
|
|
| 自动检测 |
|
|
| 自启动时是否显示 PD 窗口 |
| PD 默认 | 新建 PDM 的 DBMS,如 |
|
| 单次 COM 调用超时(秒) |
完整清单见 src/pd_mcp/config.py;也支持 JSON 配置文件(pdmcp.json)。
开发与测试
.venv\Scripts\python.exe -m pytest tests -m "not live" # 42 项单测/冒烟(Mock 后端)
$env:PDMCP_LIVE = "1"
.venv\Scripts\python.exe -m pytest tests -m live # 真机验收
.venv\Scripts\python.exe -m pd_mcp probe # 独立 COM 能力探测Mock 后端复刻 PowerDesigner 语义(列 Primary 主键、引用 FK 迁移、索引列绑定), 因此 schema 编排、校验、事务与 DDL 全流程在无 PD 许可证的机器上同样可测。
验证环境与诚实声明
验证环境:Windows 10/11 + PowerDesigner 16.5.0.3982 + 64 位 Python 3.13。 COM 事实经厂商
VBScriptConstants.vbs、Interop.*.dll元数据、官方 C# 样例 与真机实验交叉验证——见 docs/com-api-notes.md。16.5 的
CheckModel()返回None(结果进 PD Result List 窗口);机器可读的 检查结果由validate_model/check_database_design提供。PowerDesigner 没有 SaveAs:未保存模型另存通过 ShellNew 模板文件绑定 + 内容复制实现(返回
copied统计)。复制过程会重新挂载图符号并自动布局, 且会校验请求路径确实拿到模型内容而非模板壳(PD 16.5 有时会把内容写到 去掉扩展名的同级文件)。CDM/LDM 为一等公民:实体/属性、标识符(主标识)、概念联系(两端基数 + 依赖侧)。索引属物理概念,索引类工具对 CDM/LDM 显式报错。
CDM 继承(Inheritance)暂未暴露;Adapter 接口使补充实现非常直接。
生成与模型 DBMS 不符的 SQL 会返回明确错误——请先
create_model(dbms=...)或 ChangeDBMS。
许可证
Available Tools
60 toolsapply_schema_patchA
Apply an ordered list of structured patch operations to a model. operations: [{op: create_table|alter_table|rename_table|drop_table|add_column|alter_column|drop_column|create_primary_key|drop_primary_key|create_reference|alter_reference|drop_reference|create_index|alter_index|drop_index|create_domain, ...}]. Execution stops at the first failure and (atomic, default) rolls back. dry_run=true returns the plan.
Inputs: model_id (required), operations (required), atomic, dry_run.
begin_transactionA
Start a transaction before a batch of modifications. For file-backed models the current file is saved and backed up so rollback can restore the exact pre-transaction state. Pass the returned txn_id to commit_transaction / rollback_transaction.
Inputs: model_id (required), auto_save.
check_database_designA
Check a model against YOUR design rules. rules: [{id, description, type, params?, severity?}]. Machine-checkable types: TABLE_HAS_PK, TABLE_CODE_PATTERN, TABLE_NAME_STYLE, COLUMN_HAS_COMMENT, COLUMN_CODE_PATTERN, COLUMN_MANDATORY, REQUIRED_COLUMNS, PK_NAME_PATTERN, FK_NAME_PATTERN, INDEX_NAME_PATTERN, DATA_TYPE_ALLOWED, NO_DUPLICATE_INDEX, NO_EMPTY_COMMENT_TABLE, REGEX. Rules with unknown types are reported in 'skipped' (never silently ignored). Returns passed/errors/warnings/violations.
Inputs: rules (required), model_id (required).
close_modelB
Close a model in PowerDesigner. save=true saves to its file before closing.
Inputs: model_id (required), save.
commit_transactionA
Commit a transaction: keeps all changes and discards the backup.
Inputs: txn_id (required).
compare_modelA
Structurally compare two open models: added/removed/modified tables, added/removed/modified columns per table, references and domains.
Inputs: model_id_a (required), model_id_b (required).
convert_cdm_to_ldmC
Convert a CDM model to an LDM. Uses PowerDesigner's native conversion when available, otherwise a structured mapping. Returns the new model info.
Inputs: model_id (required).
convert_cdm_to_pdmA
Convert a CDM model to a PDM (optionally pass dbms, e.g. 'MySQL 5.0'). Native conversion first, structured mapping as fallback.
Inputs: model_id (required), dbms.
convert_ldm_to_pdmC
Convert an LDM model to a PDM (optionally pass dbms). Native conversion first, structured mapping as fallback.
Inputs: model_id (required), dbms.
create_columnA
Add a column to a table. Supported properties: name, code, data_type (e.g. 'VARCHAR', 'INT', 'DATETIME'), length, precision, mandatory, default_value, comment, description, domain (ref or code), primary (include in primary key). Supports dry_run.
Inputs: name (required), model_id (required), table_ref (required), code, +9 more.
create_database_schemaA
Create a whole database schema in one call from a JSON spec: {tables:[{name, code, comment, columns:[{name, code, data_type, length, precision, mandatory, default_value, comment, primary}], indexes:[...]}], relationships:[{parent_table, child_table, parent_columns, child_columns, name, cardinality}], domains:[...], indexes:[...]}. Primary keys are derived from columns with primary:true (or primary_key_columns). dry_run=true returns the execution plan without touching the model; atomic=true (default) rolls the model back if any step fails.
Inputs: spec (required), model_id (required), atomic, dry_run.
create_indexB
Create an index on a table: normal, unique or composite. Example: columns=['user_id','create_time'], name='idx_order_user_time'. Supports dry_run.
Inputs: columns (required), model_id (required), table_ref (required), name, +3 more.
create_modelA
Create a new empty model. kind: 'PDM' | 'CDM' | 'LDM'. For PDM you may pass dbms (e.g. 'MySQL 5.0'); PowerDesigner picks its default when omitted. Returns the new model info including its model_id.
Inputs: kind (required), name (required), code, dbms.
create_primary_keyB
Create (or replace) the primary key of a table from the given column codes. Supports single-column and composite keys, e.g. columns=['order_id','product_id']. Existing PK membership is cleared. Supports dry_run.
Inputs: columns (required), model_id (required), table_ref (required), code, +2 more.
create_referenceA
Create a foreign-key reference between two PDM tables (1:N from parent to child). If parent_columns/child_columns are omitted, the parent's primary key is used and FK columns are auto-created in the child (update_key=true, e.g. user 1--N order creates order.user_id). cardinality like '0,n' or '1,1'. Supports dry_run.
Inputs: model_id (required), child_table (required), parent_table (required), code, +7 more.
create_tableA
Create a table (PDM) or entity (CDM/LDM). Optionally pass an inline 'columns' array (same shape as create_column's column spec) to create columns in the same call. Supports dry_run.
Inputs: name (required), model_id (required), code, columns, +2 more.
delete_columnA
Delete a column from a table. Destructive: prefer dry_run first.
Inputs: model_id (required), table_ref (required), column_ref (required), dry_run.
delete_indexA
Delete an index. Destructive: prefer dry_run first.
Inputs: model_id (required), index_ref (required), table_ref (required), dry_run.
delete_referenceA
Delete a foreign-key reference (the FK columns in the child are kept). Destructive: prefer dry_run first.
Inputs: model_id (required), reference_ref (required), dry_run.
delete_tableA
Delete a table (and PowerDesigner cascades its columns/keys/indexes and attached references). Destructive: prefer dry_run first.
Inputs: model_id (required), table_ref (required), dry_run.
design_from_specA
Materialise a full design spec into a NEW model (or extend an existing one via model_id). spec: {model:{kind: 'PDM'|'CDM'|'LDM', name, code, dbms}, tables/entities:[...], relationships:[...], domains:[...], indexes:[...]}. The AI supplies the design; this tool executes it. Supports dry_run.
Inputs: spec (required), atomic, dry_run, model_id.
generate_ddlA
Generate SQL DDL for a PDM using PowerDesigner's native database generation for the model's DBMS. output_path: target .sql file (created if the directory exists). Returns the SQL text plus file info. The model's own DBMS is used; requesting a different dbms returns a clear error.
Inputs: model_id (required), output_path (required), dbms.
get_columnC
Get one column's full detail.
Inputs: model_id (required), table_ref (required), column_ref (required).
get_domainC
Get one domain's detail.
Inputs: model_id (required), domain_ref (required).
get_indexC
Get one index's full detail.
Inputs: model_id (required), index_ref (required), table_ref (required).
get_keyB
Get one key by ref (from list_keys).
Inputs: key_ref (required), model_id (required), table_ref (required).
get_model_infoA
Get detailed info about one open model: kind, name, code, file, DBMS, and counts of tables/entities, columns, keys, references, indexes, domains.
Inputs: model_id (required).
get_packageC
Get details of one package by ref, code or name.
Inputs: model_id (required), package_ref (required).
get_referenceC
Get one reference's full detail.
Inputs: model_id (required), reference_ref (required).
get_relationshipB
Get one CDM/LDM relationship by ref.
Inputs: model_id (required), relationship_ref (required).
get_server_infoA
Check the PowerDesigner connection and report server capabilities. Call this first to verify PowerDesigner is reachable (or that the mock backend is active). Returns version, attach mode, open model count.
Inputs: no named parameters.
get_tableB
Get full detail of one table/entity by ref or code: columns (type, length, mandatory, default, comment, primary), keys, primary key, indexes and incoming/outgoing references.
Inputs: model_id (required), table_ref (required).
inspect_schemaA
Inspect the whole schema at once. mode='summary'|'detail', optional single table, paginated. Returns tables + foreign keys view; use this instead of many list_* calls to save context.
Inputs: model_id (required), mode, page, table, +1 more.
list_columnsC
List columns of a table with full attributes. Optional text search.
Inputs: model_id (required), table_ref (required), query.
list_domainsB
List domains (reusable data type definitions) of a model.
Inputs: model_id (required).
list_indexesA
List indexes of one table or of the whole model. Shows unique flag and covered columns (supports simple, unique and composite indexes).
Inputs: model_id (required), table_ref.
list_keysB
List keys of a table. The primary key has primary=true and its columns are included (supports single-column and composite keys).
Inputs: model_id (required), table_ref (required).
list_model_backupsA
List available model backup files (created by transactions and save operations) in the backup directory.
Inputs: no named parameters.
list_open_modelsA
List all models currently open in PowerDesigner with kind (PDM/CDM/LDM), name, code, file path, DBMS and object counts.
Inputs: no named parameters.
list_packagesA
List packages (sub-models/namespaces) of a model.
Inputs: model_id (required).
list_referencesA
List foreign-key references (PDM) of a model: parent/child tables, join column pairs, cardinality and mandatory flag.
Inputs: model_id (required).
list_relationshipsA
List relationships of a CDM/LDM model (entity-to-entity relations). For PDM models use list_references instead.
Inputs: model_id (required).
list_tablesA
List tables (PDM) or entities (CDM/LDM) of a model. Supports text search over code/name and optional package filter. Returns brief rows - use get_table for full detail.
Inputs: model_id (required), page, query, page_size, +1 more.
model_snapshotB
Convert the model into compact JSON the AI can reason over. mode='summary': one line per table + references. mode='detail': full columns/keys/indexes/references. Optionally restrict to given table codes.
Inputs: model_id (required), mode.
open_modelA
Open a PowerDesigner model file (.pdm / .cdm / .ldm) and return its info. Set read_only=true to prevent modifications.
Inputs: path (required), read_only.
remove_primary_keyA
Remove the primary key from a table. Destructive: prefer dry_run first.
Inputs: model_id (required), table_ref (required), dry_run.
rename_columnB
Rename a column (Name and/or Code). Supports dry_run.
Inputs: model_id (required), table_ref (required), column_ref (required), dry_run, +2 more.
rename_tableB
Rename a table (Name and/or Code). Supports dry_run.
Inputs: model_id (required), table_ref (required), dry_run, new_code, +1 more.
rollback_modelA
Restore a model from a backup file (latest backup, or pass backup_file from list_model_backups). Closes the model, restores the file and reopens it.
Inputs: model_id (required), backup_file.
rollback_transactionA
Roll a transaction back: file-backed models are restored from the pre-transaction backup and reopened (the model_id may change - run list_open_models). Without a file backup, create/update operations are undone via the operation journal; deletions without backup cannot be undone and are reported explicitly.
Inputs: txn_id (required).
save_modelA
Save a model to its current file. Fails with a clear error when the model was never saved (use save_model_as first).
Inputs: model_id (required).
save_model_asA
Save a model to a new file path (.pdm/.cdm/.ldm extension is appended automatically). The target file must not exist yet.
Inputs: path (required), model_id (required).
search_columnsA
Search columns across the whole model (or one table) by substring in code/name. Returns table_code + column detail rows.
Inputs: query (required), model_id (required), page, page_size, +1 more.
search_tablesB
Search tables by substring in code or name (case-insensitive).
Inputs: query (required), model_id (required), page, page_size.
set_primary_keyA
Alias of create_primary_key: replaces the table's primary key with the given columns (composite supported).
Inputs: columns (required), model_id (required), table_ref (required), name, +1 more.
update_columnA
Update column properties: name, code, data_type, length, precision, mandatory, default_value, comment, description, domain, primary. Only supplied values are changed. Supports dry_run.
Inputs: model_id (required), table_ref (required), column_ref (required), code, +10 more.
update_indexC
Update an index: name, unique flag, comment, or replace its columns. Supports dry_run.
Inputs: model_id (required), index_ref (required), table_ref (required), name, +4 more.
update_referenceC
Update a reference: name, code, comment, mandatory, parent_role, child_role, cardinality ('0,n', '1,1', ...). Supports dry_run.
Inputs: model_id (required), reference_ref (required), code, name, +6 more.
update_tableC
Update table properties: name, code, comment. Supports dry_run.
Inputs: model_id (required), table_ref (required), code, name, +2 more.
validate_modelA
Run built-in structural checks on a model: missing primary keys, duplicate table/column codes, missing data types, missing comments, FK type mismatches, dangling references, duplicate/redundant indexes. Returns passed/errors/warnings with object locations - fix issues and re-run to close the loop.
Inputs: model_id (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 60 tools
Tools are mostly grouped by resource type with clear list/get/create/update/delete distinctions, but there is notable overlap: list_tables vs search_tables, list_columns vs search_columns, model_snapshot vs inspect_schema, and validate_model vs check_database_design. The aliased create_primary_key/set_primary_key pair and multiple batch-modification tools (create_database_schema, apply_schema_patch, design_from_spec) add further selection ambiguity, though descriptions do mitigate confusion.
The overwhelming majority of tools follow a clear verb_noun snake_case pattern: list_tables, get_column, create_reference, update_index, delete_reference. Minor deviations exist: model_snapshot, inspect_schema, design_from_spec, and begin/commit/rollback_transaction use noun-first or non-standard forms, but they are still readable and the pattern is highly consistent overall.
At 60 tools, the surface is extremely large and exceeds the 50+ threshold for extreme mismatch, even accounting for PowerDesigner's broad domain. While each tool is individually meaningful, the sheer count creates a heavy integration burden and overlaps among search/snapshot/batch tools suggest it could be consolidated.
The toolset covers nearly the full lifecycle for models, tables, columns, keys, indexes, and references, and adds transaction management, validation, DDL generation, and model conversion. Minor gaps exist: CDM/LDM relationships only have list/get with no dedicated create/update/delete, domains lack direct CRUD, and there is no delete_model tool, though these can be worked around via apply_schema_patch or design_from_spec.
Maintenance
Related MCP Connectors
Exactly 50 data transformation and live web verification tools for AI agents.
PostgreSQL, MySQL, and SQL Server in one session. 26 read-only MCP tools for AI agents.
The agent-native cloud: database, functions, AI, storage, computers. 53 tools. OAuth or API key.
Protocol-native energy infrastructure orchestration for AI data centers. Provides 46 MCP tools across 8 grid protocols (IEC-61850, DNP3, Modbus, OCPP, OpenADR, IEEE 2030.5, IEC 60870-5-104, ICCP) with 5 core API primitives: connect, dispatch, settle, comply, and intel. Enables AI agents to programmatically interact with substations, grid interfaces, and energy assets for real-time workload-grid coordination.
Related MCP Servers
- AlicenseAqualityBmaintenanceEnables local Power BI project (.pbip) automation, including model (TMDL) and report (PBIR) layer manipulation, without cloud dependencies. Provides 56 tools for building, editing, and validating reports and measures through natural language prompts.13MIT
- AlicenseNot gradedqualityAmaintenanceEnables AI agents to control a running SOLIDWORKS session through its COM API, with tools for sketching, feature creation, assemblies, and visual feedback via screenshots.22Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to build, measure, verify, and correct parametric SolidWorks parts and assemblies locally via the COM API, supporting operations like sketching, extruding, modeling, dimensioning, material assignment, export, and assembly mating.MIT
- FlicenseNot gradedqualityAmaintenanceEnables AI agents to draft, inspect, and validate CAD drawings through a large typed tool catalog, driving either a live desktop CAD seat over COM or a headless ezdxf engine on any platform. It covers ISO GD&T and dimension-tolerance checks plus P&ID, mechanical, and architectural sheet generation, with selectable tool packs and a low-token discovery mode.-