代理工具
Astah Pro MCP Main Branch Compresses Nearly 400 UML Operations into 83 Programmable Tools
The experimental main branch of Astah Pro MCP adds programmatic tool calling, replacing nearly 400 fine-grained operations in direct mode with 83 tools to reduce the burden of loading large tool catalogs for agents. The changes do not yet appear in Astah’s official stable release information for version 0.2.9, and model data, logs, and editing permissions are still exposed to external agents.
The independently developed Astah Pro MCP plugin is testing a new programmatic tool-calling mode on its main branch, allowing agents such as Claude Code and Codex CLI to directly read and write UML models in Astah Professional. The legacy direct mode remains available on port `18888` and exposes 398 tools, while the new mode uses local port `8888` and exposes only 83 higher-level interfaces to the model. For agents that limit the number of tools or place complete schemas in the context, a smaller tool catalog can reduce token overhead and the risk of selecting the wrong operation from several similar options.
The plugin can create or query class diagrams, sequence diagrams, activity diagrams, state machines, use-case diagrams, requirement diagrams, and mind maps. It can also convert between models and source code. Communication uses Streamable HTTP but listens only on localhost; Astah displays a confirmation dialog whenever a new agent session connects. The README shows three agents taking approximately 13, 19, and 41 minutes to complete the same set of diagramming tasks. However, these are only single demonstrations, with no success rates, token counts, or amount of manual correction reported, so they should not be treated as a model ranking.
The update also exposes a deployment versioning issue: Astah’s official plugin page still lists version 0.2.9 as the latest release and describes approximately 300 editable tools and 100 read-only tools, while the GitHub main branch requires Astah Pro 12, uses different ports, and provides 83 or 398 interfaces depending on the mode. Teams tracking the main branch directly cannot reuse the official stable-release configuration or assume tool-schema compatibility. They should pin a commit or JAR, preserve the MCP tool manifest, and run regression tests.
In addition, local transport does not mean the data remains local: project content and logs are still sent to the connected model service, and the agent can modify design files. The maintainer recommends creating a backup with Git or a separate copy first. For confidential architectures, teams should also restrict vendor data retention, review connection confirmations, and wait for an official release and more comprehensive consistency testing.