[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7CRQ4QJ0CD2A69EK8DTYVWQ":3,"skill-same:01M7CRQ4QJ0CD2A69EK8DTYVWQ":50},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":27,"commitSha":28,"skillDirectory":29,"forksCount":30,"repositoryHtmlUrl":31,"repositoryHomepageUrl":32,"ownerHtmlUrl":33,"ownerBlogUrl":6,"ownerDescription":34,"updatedAt":35,"install":36,"resources":49},"01M7CRQ4QJ0CD2A69EK8DTYVWQ","data-mcp-tools",null,"Use when tools from the jupyter, duckdb, dbt, mlflow, huggingface or postgres MCP servers appear in your tool list, or before reaching for one — what each is good for, what it can damage, and the CLI fallback when it is not connected","makifbaysal\u002Ftasktrooper","makifbaysal","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F25031940?v=4","01M7CRG3ZCSYZSSEVDMW3DEYC8","Mehmet Akif Baysal","01M7CRG401FJ8YAXH5ER730XYV",122,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fmakifbaysal\u002Ftasktrooper\u002Fblob\u002Fmain\u002Fcatalog\u002Fagents\u002Fdata-scientist\u002Fskills\u002Fdata-mcp-tools\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: data-mcp-tools\ncategory: tools\ndescription: Use when tools from the jupyter, duckdb, dbt, mlflow, huggingface or postgres MCP servers appear in your tool list, or before reaching for one — what each is good for, what it can damage, and the CLI fallback when it is not connected\ntech_stack: MLOps\nsource: original; informed by the datalayer jupyter-mcp-server, motherduckdb mcp-server-motherduck, dbt-labs dbt-mcp, MLflow MCP server, Hugging Face MCP and modelcontextprotocol server-postgres documentation (cited, not reproduced)\n---\n# Data MCP Tools\n\n## Overview\n\nTaskTrooper can connect optional MCP servers that make data work faster: a live Jupyter kernel, DuckDB, dbt, MLflow, Hugging Face search and Postgres. They are disabled by default; a person enables them in Settings → MCP servers. They are conveniences on top of the workspace, never a replacement for it: the repository's code and its test suite remain the deliverable and the evidence.\n\n**Core principle:** a server is available only if its tools are in your tool list right now. If they are not, use the CLI equivalent and carry on — never stop a task, and never ask a person to install or enable a server mid-task.\n\n## 1. Detecting availability\n\nLook at your tool list. Tools from a server carry its id in the name — `mcp_duckdb_…` in TaskTrooper's own runtime, `mcp__duckdb__…` in a CLI runtime. Tool names differ between server versions; read the descriptions in your tool list rather than assuming a name from this skill. A server whose tools are absent is not connected for this run, whatever the repository's docs say.\n\n## 2. The servers\n\n| Server id | What it gives you | Use it for | Fallback without it |\n|---|---|---|---|\n| `jupyter` | read, insert, edit and execute cells in a running local JupyterLab | looking at data interactively, checking a notebook runs cell by cell | `uv run jupyter nbconvert --execute`, papermill, or a script |\n| `duckdb` | SQL over Parquet\u002FCSV\u002FDuckDB files, in-memory by default | EDA, profiling (`SUMMARIZE`), reconciling a result two ways | `duckdb` CLI or `uv run python -c \"import duckdb; ...\"` |\n| `dbt` | lineage, compile, list, show, test, run, build for the project | finding a model's parents and children, compiling SQL, running tests on the dev target | the `dbt` CLI with the same commands |\n| `mlflow` | experiments, runs and registered models from the tracking server (ML tool set) | finding the current champion and its metrics to compare your candidate against | a short script with `mlflow.MlflowClient` |\n| `huggingface` | model search on the Hub | finding a pretrained model and its exact revision | the Hub website docs or `huggingface_hub` in a script |\n| `postgres` | read-only SQL against a configured Postgres | inspecting a source table's shape and row counts | `psql` with the repository's dev connection |\n\n## 3. jupyter — the kernel runs whatever you send\n\n- Executing a cell is exactly as powerful as run_terminal: it can delete files, call the network and write to databases. Apply the same care, and never send credentials into a cell.\n- The kernel keeps state between calls. A result that depends on cells you ran earlier, out of order, is not reproducible — re-run from a fresh kernel before trusting it.\n- Exploration in the kernel is fine; anything the deliverable needs moves into the package with tests (rule no-notebook-only-logic, notebook-to-module). A notebook edited through the server is still committed the repository's way (stripped outputs, jupytext pair or marimo file).\n- The server talks to a JupyterLab a person started on this machine; it may have nothing to do with your task workspace. Check the notebook path and the kernel's working directory before reading results as evidence about your branch.\n\n## 4. duckdb — fast SQL for EDA and result QA\n\n- Point it at files in your workspace or `\u002Ftmp\u002Ftt-\u003Ctask key>\u002F`: `SELECT * FROM read_parquet('data\u002Fsample\u002F*.parquet')`, `SUMMARIZE ...`, join-count checks (analytics-result-qa).\n- Results come back capped (about 1,024 rows by default): aggregate and profile in SQL instead of pulling raw rows to read by eye.\n- The default database is in-memory — tables you create disappear with the session. Anything a pipeline needs is written by code in the diff, not by a tool call.\n- A DuckDB file or a MotherDuck (`md:`) connection may be configured instead: treat writes there like writes to a shared database — do not.\n\n## 5. dbt — prefer read-only and dev-target commands\n\n- Lineage, `list`, `compile`, `parse` and `show` read the project and are safe; use them to find what a change affects before editing (sql-analytics-and-dbt).\n- `run` and `build` (and `test`, which queries the warehouse) execute against the warehouse with the credentials in `profiles.yml`. Use them only against a development target, never production, and only for the models your task touches (`--select fct_orders+`). A person may have disabled them for the server — then use the CLI on the dev target, or say they could not run.\n- The tool result is evidence only for the target it ran on; name the target in your closing message.\n\n## 6. mlflow — read to compare, never delete\n\n- Use it to read experiments, runs and registered models: the champion's version and alias, its logged metrics, the data fingerprint and git SHA it was trained on. Compare your candidate against those numbers on the same data (model-training-and-evaluation).\n- Never delete or rename runs, experiments, models or traces, and never move a registry alias: promotion is deploy-time work for a person (rule deploy-runbook-fields).\n- Your own test and exercise runs log to a local store in `\u002Ftmp\u002Ftt-\u003Ctask key>\u002F`, not to the server behind this tool (experiment-tracking-and-reproducibility).\n\n## 7. huggingface and postgres\n\n- Hugging Face: search, then pin what you choose by commit revision in code (deep-learning-pytorch); check the model's license and whether it needs `trust_remote_code` before proposing it.\n- Postgres: queries run read-only. Use them to learn a source's columns, types and row counts; never paste production rows into fixtures, commits or comments.\n\n## 8. Evidence\n\nA tool call is not a test. Whatever you learned through a server is either encoded in code with tests in the diff, or reported in your closing message as what you looked at (server, target, query). The hand-off gate runs the repository's checks (data-build-check); MCP results never replace them.\n\n## Common Mistakes\n\n- Waiting for, or asking for, a server that is not connected instead of using the CLI.\n- Building the deliverable in the Jupyter kernel and never moving it into the package.\n- Running `dbt build` through the server against the production target.\n- Pulling thousands of rows through DuckDB to inspect by eye.\n- Treating the shared MLflow server as a scratchpad for test runs.\n\n## Red Flags\n\n- A conclusion in your closing message that only exists in kernel state.\n- A dbt or SQL tool call whose target you do not know.\n- A delete, rename or alias change on a shared MLflow server.\n- Production data copied from a tool result into a fixture.\n","# Data MCP Tools\n\n## Overview\n\nTaskTrooper can connect optional MCP servers that make data work faster: a live Jupyter kernel, DuckDB, dbt, MLflow, Hugging Face search and Postgres. They are disabled by default; a person enables them in Settings → MCP servers. They are conveniences on top of the workspace, never a replacement for it: the repository's code and its test suite remain the deliverable and the evidence.\n\n**Core principle:** a server is available only if its tools are in your tool list right now. If they are not, use the CLI equivalent and carry on — never stop a task, and never ask a person to install or enable a server mid-task.\n\n## 1. Detecting availability\n\nLook at your tool list. Tools from a server carry its id in the name — `mcp_duckdb_…` in TaskTrooper's own runtime, `mcp__duckdb__…` in a CLI runtime. Tool names differ between server versions; read the descriptions in your tool list rather than assuming a name from this skill. A server whose tools are absent is not connected for this run, whatever the repository's docs say.\n\n## 2. The servers\n\n| Server id | What it gives you | Use it for | Fallback without it |\n|---|---|---|---|\n| `jupyter` | read, insert, edit and execute cells in a running local JupyterLab | looking at data interactively, checking a notebook runs cell by cell | `uv run jupyter nbconvert --execute`, papermill, or a script |\n| `duckdb` | SQL over Parquet\u002FCSV\u002FDuckDB files, in-memory by default | EDA, profiling (`SUMMARIZE`), reconciling a result two ways | `duckdb` CLI or `uv run python -c \"import duckdb; ...\"` |\n| `dbt` | lineage, compile, list, show, test, run, build for the project | finding a model's parents and children, compiling SQL, running tests on the dev target | the `dbt` CLI with the same commands |\n| `mlflow` | experiments, runs and registered models from the tracking server (ML tool set) | finding the current champion and its metrics to compare your candidate against | a short script with `mlflow.MlflowClient` |\n| `huggingface` | model search on the Hub | finding a pretrained model and its exact revision | the Hub website docs or `huggingface_hub` in a script |\n| `postgres` | read-only SQL against a configured Postgres | inspecting a source table's shape and row counts | `psql` with the repository's dev connection |\n\n## 3. jupyter — the kernel runs whatever you send\n\n- Executing a cell is exactly as powerful as run_terminal: it can delete files, call the network and write to databases. Apply the same care, and never send credentials into a cell.\n- The kernel keeps state between calls. A result that depends on cells you ran earlier, out of order, is not reproducible — re-run from a fresh kernel before trusting it.\n- Exploration in the kernel is fine; anything the deliverable needs moves into the package with tests (rule no-notebook-only-logic, notebook-to-module). A notebook edited through the server is still committed the repository's way (stripped outputs, jupytext pair or marimo file).\n- The server talks to a JupyterLab a person started on this machine; it may have nothing to do with your task workspace. Check the notebook path and the kernel's working directory before reading results as evidence about your branch.\n\n## 4. duckdb — fast SQL for EDA and result QA\n\n- Point it at files in your workspace or `\u002Ftmp\u002Ftt-\u003Ctask key>\u002F`: `SELECT * FROM read_parquet('data\u002Fsample\u002F*.parquet')`, `SUMMARIZE ...`, join-count checks (analytics-result-qa).\n- Results come back capped (about 1,024 rows by default): aggregate and profile in SQL instead of pulling raw rows to read by eye.\n- The default database is in-memory — tables you create disappear with the session. Anything a pipeline needs is written by code in the diff, not by a tool call.\n- A DuckDB file or a MotherDuck (`md:`) connection may be configured instead: treat writes there like writes to a shared database — do not.\n\n## 5. dbt — prefer read-only and dev-target commands\n\n- Lineage, `list`, `compile`, `parse` and `show` read the project and are safe; use them to find what a change affects before editing (sql-analytics-and-dbt).\n- `run` and `build` (and `test`, which queries the warehouse) execute against the warehouse with the credentials in `profiles.yml`. Use them only against a development target, never production, and only for the models your task touches (`--select fct_orders+`). A person may have disabled them for the server — then use the CLI on the dev target, or say they could not run.\n- The tool result is evidence only for the target it ran on; name the target in your closing message.\n\n## 6. mlflow — read to compare, never delete\n\n- Use it to read experiments, runs and registered models: the champion's version and alias, its logged metrics, the data fingerprint and git SHA it was trained on. Compare your candidate against those numbers on the same data (model-training-and-evaluation).\n- Never delete or rename runs, experiments, models or traces, and never move a registry alias: promotion is deploy-time work for a person (rule deploy-runbook-fields).\n- Your own test and exercise runs log to a local store in `\u002Ftmp\u002Ftt-\u003Ctask key>\u002F`, not to the server behind this tool (experiment-tracking-and-reproducibility).\n\n## 7. huggingface and postgres\n\n- Hugging Face: search, then pin what you choose by commit revision in code (deep-learning-pytorch); check the model's license and whether it needs `trust_remote_code` before proposing it.\n- Postgres: queries run read-only. Use them to learn a source's columns, types and row counts; never paste production rows into fixtures, commits or comments.\n\n## 8. Evidence\n\nA tool call is not a test. Whatever you learned through a server is either encoded in code with tests in the diff, or reported in your closing message as what you looked at (server, target, query). The hand-off gate runs the repository's checks (data-build-check); MCP results never replace them.\n\n## Common Mistakes\n\n- Waiting for, or asking for, a server that is not connected instead of using the CLI.\n- Building the deliverable in the Jupyter kernel and never moving it into the package.\n- Running `dbt build` through the server against the production target.\n- Pulling thousands of rows through DuckDB to inspect by eye.\n- Treating the shared MLflow server as a scratchpad for test runs.\n\n## Red Flags\n\n- A conclusion in your closing message that only exists in kernel state.\n- A dbt or SQL tool call whose target you do not know.\n- A delete, rename or alias change on a shared MLflow server.\n- Production data copied from a tool result into a fixture.\n",{"name":5,"source":24,"category":25,"tech_stack":26,"description":7},"original; informed by the datalayer jupyter-mcp-server, motherduckdb mcp-server-motherduck, dbt-labs dbt-mcp, MLflow MCP server, Hugging Face MCP and modelcontextprotocol server-postgres documentation (cited, not reproduced)","tools","MLOps","main","910394a0b6db9fc8c9a8f36328ec421d3ebd8b38","catalog\u002Fagents\u002Fdata-scientist\u002Fskills\u002Fdata-mcp-tools",18,"https:\u002F\u002Fgithub.com\u002Fmakifbaysal\u002Ftasktrooper","https:\u002F\u002Ftasktrooper.ai","https:\u002F\u002Fgithub.com\u002Fmakifbaysal","Senior Software Developer","2026-11-15 04:50:29",{"source":8,"skillName":5,"commands":37,"prompt":47,"zipAvailable":48},[38,41,44],{"tool":39,"command":40},"npx","npx skills add makifbaysal\u002Ftasktrooper --skill data-mcp-tools",{"tool":42,"command":43},"bunx","bunx skills add makifbaysal\u002Ftasktrooper --skill data-mcp-tools",{"tool":45,"command":46},"pnpm","pnpm dlx skills add makifbaysal\u002Ftasktrooper --skill data-mcp-tools","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fmakifbaysal\u002Ftasktrooper\u002Fblob\u002Fmain\u002Fcatalog\u002Fagents\u002Fdata-scientist\u002Fskills\u002Fdata-mcp-tools\u002FSKILL.md",true,[],[51,57,64,71,78,85],{"skillNo":52,"name":53,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":54,"categories":55,"keywords":56,"publishedAt":19},"01M7CSVEPQW6J6TD6CS2T5FPNZ","visual-craft","https:\u002F\u002Fgithub.com\u002Fmakifbaysal\u002Ftasktrooper\u002Fblob\u002Fmain\u002Fcatalog\u002Fagents\u002Fui-designer\u002Fskills\u002Fvisual-craft\u002FSKILL.md",[],[],{"skillNo":58,"name":59,"nameZh":6,"description":60,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":61,"categories":62,"keywords":63,"publishedAt":19},"01M7CSVAQG91ETHQ0HBNZDG5FR","ux-copy","Use when you write any words a user reads in a mockup or a hand-off - content language and voice, buttons, labels, errors, empty states, confirmations, numbers and dates, and the longest-string check","https:\u002F\u002Fgithub.com\u002Fmakifbaysal\u002Ftasktrooper\u002Fblob\u002Fmain\u002Fcatalog\u002Fagents\u002Fui-designer\u002Fskills\u002Fux-copy\u002FSKILL.md",[],[],{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":19},"01M7CSV6H0J38659NDXHJQ8396","screen-mockup-html","Use when you draw screens as the design task's HTML mockup - ONE canvas document holding every screen and every variant side by side, the sanitizer-safe recipe, phone and wide frames, the states grid, static control states, and the loopback self-check with screenshots at 375, 768 and 1440 attached to the task","https:\u002F\u002Fgithub.com\u002Fmakifbaysal\u002Ftasktrooper\u002Fblob\u002Fmain\u002Fcatalog\u002Fagents\u002Fui-designer\u002Fskills\u002Fscreen-mockup-html\u002FSKILL.md",[],[],{"skillNo":72,"name":73,"nameZh":6,"description":74,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":19},"01M7CSV3JFRF3BM3XK52M7YEYR","performance-awareness","How the performance score moves for your role. Use when the score context shows a declining trend, or before handing off work you have not self-checked against every criterion.","https:\u002F\u002Fgithub.com\u002Fmakifbaysal\u002Ftasktrooper\u002Fblob\u002Fmain\u002Fcatalog\u002Fagents\u002Fui-designer\u002Fskills\u002Fperformance-awareness\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":19},"01M7CSTZDH00YRF5SCT7BRT9QD","design-variants","Use when a screen design needs variants - how many, variants as side-by-side rows on the one canvas, what makes two variants genuinely different, recommending one, how the human's choice is read, and narrowing to the chosen one in revision","https:\u002F\u002Fgithub.com\u002Fmakifbaysal\u002Ftasktrooper\u002Fblob\u002Fmain\u002Fcatalog\u002Fagents\u002Fui-designer\u002Fskills\u002Fdesign-variants\u002FSKILL.md",[],[],{"skillNo":86,"name":87,"nameZh":6,"description":88,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":89,"categories":90,"keywords":91,"publishedAt":19},"01M7CSTV6P3SM3PAE28DB0RF77","design-system-authoring","Use when you derive, change or extend a design system - where tokens live per stack, the DTCG token tree, role-based names for both themes, base vs layer vs drift, the DESIGN.md section order, the inventory format, the contrast check and the lint every proposal returns","https:\u002F\u002Fgithub.com\u002Fmakifbaysal\u002Ftasktrooper\u002Fblob\u002Fmain\u002Fcatalog\u002Fagents\u002Fui-designer\u002Fskills\u002Fdesign-system-authoring\u002FSKILL.md",[],[]]