The design that matters is the split: one call pulls stories and bugs into a local file, and everything after that — search, statistics, trends, vector index, word frequency — reads from disk. That keeps API pressure low and makes repeated analysis cheap. Two cautions from the project itself: the fake-data generator will overwrite your real local file if you do not give it a path, and the LLM summarisation is where your ticket text leaves the machine.
An MCP server for TAPD, the Tencent project- and defect-tracking platform. It fetches a workspace's stories and bugs through the Open API and keeps them locally, then layers analysis on top: exact-field search, priority filtering, distribution statistics, time-trend charts, semantic search over a vector index, and word-frequency analysis. The heavy summarisation steps call an external LLM you configure separately.
- Stories and bugs fetched and saved locally with counts returned — the recommended first call — or returned directly as JSON when the volume is small — `get_tapd_data`, `get_tapd_stories`, `get_tapd_bugs`
- Exact or fuzzy search over any field, scoped to stories, bugs or both, with case sensitivity as an option — `precise_search_tapd_data`
- Priority filtering with high/medium/low presets, and a full distribution report covering counts, priority, status and creator — `search_tapd_by_priority`, `get_tapd_data_statistics`
- Time-trend analysis over a chosen date field and range, as a count or line chart — `analyze_time_trends`
- A vector index built over your own data, its status inspected, and natural-language search against it — `vectorize_data`, `get_vector_info`, `search_data`
- Keyword frequency analysis that feeds better search terms back to you — `analyze_word_frequency`
- Synthetic TAPD data for testing and demos, with a warning that it overwrites your local file if you do not name an output path — `generate_fake_tapd_data`
- An LLM-written project overview across a date range, with a token budget — `generate_tapd_overview`
- HTML stripped from `description` fields with a preview mode that shows the compression ratio before you commit — `preview_tapd_description_cleaning`, `preprocess_tapd_description` (documented as still in development, with document and image export blocked upstream)
TAPD Open API credentials in an `api.txt` in the project root: `API_USER`, `API_PASSWORD` and `WORKSPACE_ID`, with the account needing read permission on the project. That file is gitignored so the secret does not travel with the repo. The LLM features are optional and separately keyed: `DS_KEY` for DeepSeek or `SF_KEY` for SiliconFlow, set as environment variables and requiring an editor restart afterwards; without them the summarisation tools return an error and everything else keeps working. The README carries an explicit warning not to configure an external LLM at all if your TAPD data is confidential.