> ## Documentation Index > Fetch the complete documentation index at: https://learn.kubit.ai/llms.txt > Use this file to discover all available pages before exploring further. # Kubit skills > Reference for the Kubit plugin skills in Cursor and other coding agents: /kubit, /kubit-connect, /kubit-report, /kubit-inspect, /kubit-blame and more. This page covers the skills the Kubit plugin adds to Claude Code and Cursor: what each one does, what it needs and examples to try. It also covers installing, updating and removing the plugin, full-dataset analysis and the investigation chain from a metric to a commit. A skill is a slash command, such as `/kubit-report`, that tells your coding agent how to use the [Kubit MCP server](https://learn.kubit.ai/ai-assistant/overview.md) for one kind of task. You can also describe the task in plain language, and the agent picks the matching skill. ## When to use | Use case | Skill | Example | | - | - | - | | Not sure which skill fits | `/kubit` | `/kubit why did errors spike on Tuesday?` | | Sign in or change workspace | `/kubit-connect` | `/kubit-connect switch workspace staging` | | Trends, totals, funnels and retention | `/kubit-report` | `/kubit-report daily failed traces` | | Specific traces, sessions, users or events | `/kubit-inspect` | `/kubit-inspect failed traces with intent Checkout since yesterday` | | The commit behind a regression | `/kubit-blame` | `/kubit-blame find the commit behind the checkout escalation spike last week` | | Send traces from a repo to Kubit | `/kubit-integrate` | `/kubit-integrate` | | What Kubit can do, or contact support | `/kubit-help` | `/kubit-help what does /kubit-blame do?` | | Update the plugin | `/kubit-update` | `/kubit-update` | ## Install the plugin The Kubit plugin installs the skills and adds the Kubit MCP server to Claude Code, Cursor or both. ```bash theme={null} npx @kubit-ai/agent-plugin ``` The installer asks for the runtime (**Claude Code**, **Cursor** or **Both**) and the location. Restart your client afterwards, then run `/kubit-help` to check that the skills loaded. [Connect your AI assistant](https://learn.kubit.ai/ai-assistant/connect.md) covers the sign-in. | Flag | What it does | | - | - | | `-r`, `--runtime ` | Skips the runtime question. `claude`, `cursor` or `both`. | | `--global` | Installs user-wide. This is the default. | | `--local` | Installs into the current directory only | | `-y`, `--yes` | Accepts the default location without asking | | `-u`, `--uninstall` | Removes the plugin from the selected runtime | ### Requirements | Requirement | Needed for | | - | - | | Node.js 18+ | Installing and updating the plugin with `npx` | | `uv` or Python 3 | Full-dataset analysis in `/kubit-report` and `/kubit-inspect`. Without either, the skills show the MCP server's summary or sample rows instead. | | A git checkout of your agent's repo with full history | `/kubit-blame` | ## /kubit `/kubit` is the plain-language entry point. It reads your question and hands it to one skill, or runs an investigation across several. Use it when you don't know which skill fits, or when the question is "why did this change?". ### How it works | Your question | What `/kubit` does | | - | - | | Matches one skill | Hands the question to that skill, which does the rest | | Asks why a metric changed or what regressed | Runs the [investigation chain](#investigation-chain) | | Unclear | Asks you one short question, then routes by your answer | ### Examples > `/kubit why did errors spike on Tuesday?` > "Ask Kubit to show failed checkout traces since yesterday" ## /kubit-connect `/kubit-connect` signs you in to the Kubit MCP server and picks the organization and workspace the other skills work in. Other skills call it for you when they need a workspace, so you run it yourself mainly to switch. ### How it works | Step | Details | | - | - | | Update check | On first use, prints a notice when a newer plugin version is on npm. It doesn't update anything. | | Sign-in | If you aren't signed in, it opens the Kubit sign-in link in your browser, or tells you to run `/mcp` in Claude Code | | Workspace | Shows your current user, organization and workspace, and lists the others you can switch to. With one organization and one workspace, it doesn't ask. | | Switch | `/kubit-connect switch workspace ` changes the workspace for this conversation | Example workspaces are read-only and show `[example: read-only, cannot mint api key]` next to their name. To send traces, use a workspace you created. `/kubit-connect` doesn't create workspaces. `/kubit-integrate` does. ### Examples > `/kubit-connect` > `/kubit-connect switch workspace staging` ## /kubit-report `/kubit-report` finds, opens and creates Kubit reports. It builds the report types the MCP server supports: Query, Funnel, Path, Retention, Inspection and Agent flow. [MCP tools reference](https://learn.kubit.ai/ai-assistant/tools.md#report-types) lists them. ### How it works | You ask to | `/kubit-report` | | - | - | | Open a report by ID | Opens it, such as `/kubit-report 10798` | | Find a report by name | Searches. One match opens. Several matches show a list to pick from. No match offers to create one. | | See recent reports | Lists your 10 most recent reports | | Create a report ("build", "create", "make", "new") | Builds it from your wording | | Re-run a report | Runs it again on current data, only when you ask | | Change a report | Builds a new report. Reports can't be edited in place, and the original stays as it was. | When your wording could mean an existing report or a new one, `/kubit-report` searches first. It doesn't build several reports in one turn without asking. Every report comes back as a link to the report in Kubit. When a report shows errors or a regression, `/kubit-report` suggests `/kubit-blame` but doesn't run it. When the rows are traces, sessions or users, it suggests `/kubit-inspect` to drill in. ### Examples > `/kubit-report build a funnel for user query, intent classification, tool call, response` > `/kubit-report create a weekly retention report for users whose first session had zero errors` ## /kubit-inspect `/kubit-inspect` finds specific users, sessions, traces and events, and moves between them. A user has sessions, a session has traces and a trace has events, so you can start at any level and drill down. Use `/kubit-report` for trends over time. ### How it works `/kubit-inspect` passes your wording to the MCP server's `inspect` tool, which returns the total number of matches, up to 5 sample rows and a CSV export of every match. | Your question | What you get | | - | - | | A lookup, such as "show me failed traces" | A one-line count ("Showing 5 of 47") and a table of the sample rows. A single match is summarized in prose. It then offers full-dataset analysis. | | An analysis, such as a percentile, average, top N, breakdown or "why" | [Full-dataset analysis](#full-dataset-analysis) of every match, without asking first | | A follow-up about "those" or "that set" | Analysis of the data already downloaded, without a new query | | Nothing matches | A note to widen the time range or check the filter values | After the results, `/kubit-inspect` offers the next level down, such as the traces in a session. When the results include failures, it suggests `/kubit-blame`. ### Examples > `/kubit-inspect sessions where token cost > $5 in the last 2 hours` > `/kubit-inspect p95 latency for failed traces today` ## /kubit-blame `/kubit-blame` finds the commits most likely to have caused a trace regression, such as an error spike, sentiment drift or more escalations. It works on traces you already found with `/kubit-report` or `/kubit-inspect`, and never fetches metrics itself. ### How it works 1. **Detects your tracing setup.** It scans your repo for the tools that shape your spans and asks you to confirm what it found. 2. **Maps traces to code.** It matches agent names, tool names and prompts from the traces to files and lines in your repo. You confirm any match it isn't sure of. 3. **Picks the time window.** It uses the window from the report or asks you for one. 4. **Ranks commits.** It runs `git log` over the mapped code and ranks commits in the window by how close they are in time, how much of the mapped code they touch and how large the change is. The result lists the top **5** suspect commits, each with a short summary of how it changed behavior. If no commit touched the mapped code in the window, it suggests a wider window. | Constraint | Value | | - | - | | Tracing setups it reads | Langfuse and Braintrust as destinations. Vercel AI SDK, OpenTelemetry GenAI, LangChain, LangSmith, OpenInference, Traceloop, Logfire and OpenAI Agents as span sources. | | LangChain with no Langfuse or Braintrust callback | **Not supported.** LangChain sends no spans on its own. | | LangChain through Langfuse below v3 (`langfuse.callback` or `langfuse-langchain`) | LangChain spans can't be mapped to code. The skill marks them unresolved and continues. | | Repo without git history | **Not supported** | | Shallow clone | Warns and offers `git fetch --unshallow` | ### Examples > `/kubit-blame why did trace t_abc fail? What changed?` > `/kubit-blame find the commit behind the sentiment drop in the report I just ran` ## /kubit-integrate `/kubit-integrate` wires Kubit tracing into a Python or Node repo. It detects your tracing setup, creates or picks a workspace, writes an API key to your env file and adds the Kubit SDK. [Agent plugin](https://learn.kubit.ai/agent-observability/integrate/agent-plugin.md) covers it step by step. ### Examples > `/kubit-integrate` > `/kubit-integrate turn on Kubit for this Next.js app` ## /kubit-help `/kubit-help` lists the Kubit skills, explains one in detail or suggests the skill for a task you describe. It works before you sign in. It also sends a bug report, question or feedback to Kubit support. The request includes context from your conversation and project, such as your language, framework, Kubit SDK version, workspace and recent errors. The skill waits about 20 seconds for a reply, then tells you to check your email. ### Examples > `/kubit-help what does /kubit-blame do?` > `/kubit-help report a bug: my funnel report fails with a validation error` ## /kubit-update `/kubit-update` updates the plugin. It reads your installed version, checks npm for a newer one, shows the changelog since your version and asks you to confirm before it runs the installer. Restart your client after the update. It works before you sign in. > [!WARNING] > > The update is a clean install. It replaces the plugin's skill files, so any edits you made to them are lost. ### Examples > `/kubit-update` > "Check if there's a new version of Kubit" ## Investigation chain The investigation chain answers "why did this metric change?" in three steps. `/kubit` runs it, starting at the step your question already covers: | Step | Skill | Starts here when you have | | - | - | - | | 1. Chart the metric and pin the time window | `/kubit-report` | A metric or trend question | | 2. Pull the traces in that window | `/kubit-inspect` | Trace or session IDs, or an export link | | 3. Find the commit | `/kubit-blame` | The failing traces, and the question is what changed | The chain carries the workspace, time window, report links and trace IDs from one step to the next. It asks before it runs `/kubit-blame`, and stops when a step answers the question or finds no results. ### Examples > `/kubit why did errors spike on Tuesday?` > "Ask Kubit what regressed after Friday's deploy" ## Full-dataset analysis Full-dataset analysis answers questions about every row of a report or inspection, not only the sample the MCP server returns. `/kubit-report` and `/kubit-inspect` run it when the result has a CSV export link. ### How it works The skill starts an analyst subagent. The subagent downloads the CSV export, analyzes it with pandas and returns a headline, a compact table and the notable findings. It runs Python through `uv` when it's installed, or in a temporary Python 3 environment otherwise. The downloaded data stays on your machine, under `/tmp/kubit-dataset/`, one dataset per workspace. A follow-up question about the same data reuses it, and the next new query replaces it. | Constraint | Value | | - | - | | Needs | `uv` or Python 3 | | Without `uv` or Python 3 | The skill shows the MCP server's summary or sample rows | | Result without a CSV export | No full-dataset analysis. The skill says so and shows the summary. | ### Examples > `/kubit-inspect top 10 users by total token cost last week` > "Which of those traces had the longest tool calls?" ## Related * [Connect your AI assistant](https://learn.kubit.ai/ai-assistant/connect.md): sign in to the Kubit MCP server in Claude Code and Cursor. * [MCP tools reference](https://learn.kubit.ai/ai-assistant/tools.md): the tools the skills call and their limits. * [Agent plugin](https://learn.kubit.ai/agent-observability/integrate/agent-plugin.md): wire tracing into a repo with `/kubit-integrate`. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.