Steroid_Kit
Steroid CLI›Agent Skills
Agent Skills
Steroid manages two skills in Local mode and four in Connected mode, so you can drive project indexing, conversation-memory recall and workflow automation from supported agents. In Claude Code and Cursor you invoke them as /steroid…; in Codex as $steroid…. They install automatically when the CLI connects each IDE, and are removed when you disconnect it.
NOTE
Codex uses $, not /. Codex reserves / for its own built-in commands, so Steroid's skills are invoked with a dollar sign — $steroid-recall, etc. The arguments after the name are identical on every platform.
What Are Steroid Skills
Skills are named commands Steroid ships into your agent — each is a folder-with-SKILL.md the IDE exposes as a command you type in the chat. They install automatically when you connect an IDE (part of the connect flow's “Install skills” step) and uninstall when you disconnect or run steroid uninstall. The separate steroid-onboard bootstrap can be installed with steroid install-onboard-skill and intentionally survives uninstall.
Recall uses MCP memory tools, index-project registers/builds local indexes, and the two workflow skills call server-backed Flows tools. Managed skills follow the Local / Connected split as everything else.
| Skill | Family | Runs | Mode |
|---|---|---|---|
steroid-recall | Chat memory | recall_memory, recall_turn, recall_diff | Local & Connected |
steroid-index-project | Project indexing | Register and build a local conversation-memory index | Local & Connected |
steroid-create-workflow | Workflows | create_workflow / repair_workflow MCP tools | Connected only |
steroid-search-workflow | Workflows | search_workflows / run_workflow MCP tools | Connected only |
TIP
Recall works in both modes (it's the flagship local capability). The two workflow skills require Connected mode — in Local mode the workflow MCP tools aren't registered, so those skills have nothing to call.
Platforms & Invocation
Mode-appropriate managed skills install into the IDEs you configure. Two details differ by platform — how you type the command, and where the files land.
| IDE | Invoke a skill | Installed to |
|---|---|---|
| Claude Code | /steroid-<name> | ~/.claude/skills/<name>/SKILL.md |
| Cursor | /steroid-<name> | ~/.cursor/skills/<name>/SKILL.md |
| Codex | $steroid-<name> | ~/.codex/skills/<name>/SKILL.md (honors $CODEX_HOME) |
NOTE
GitHub Copilot is a registered setup target and history source. Exact Copilot skill-directory wiring was not verified in this checkout audit; use steroid setup copilot --check and the installed agent's supported skill invocation. Do not infer that Copilot has no file-based skills.
How skills get installed, verified, and removed
- Installed/reconciled during IDE setup and TUI mode transitions: recall and index-project in both modes; create-workflow and search-workflow in Connected.
- Only connected IDEs get skills. If you've connected Claude Code and Codex but not Cursor, only those two receive skills; Cursor is skipped — never reported as a failure.
- Checked through
steroid doctorand setup diagnostics. Follow each reported failure or informational skip; installing a skill does not prove provider authentication or existing chat history. - Shown in each IDE's connection-status screen as “Steroid slash-command skills installed”.
- Removed per-IDE when you disconnect that IDE, and for every IDE when you run
steroid uninstall. - Skills carry a version stamp, so removal only touches directories Steroid created — a same-named skill you installed yourself is left untouched.
steroid-recall
Proactive fallback for missing context — searches your local conversation history across every coding agent on this machine: Claude Code, Codex, Cursor, Cursor CLI and Copilot. Storage/indexing are local and do not require the Steroid backend; answering inference uses an installed agent's configured provider.
| Invoke | /steroid-recall (Claude Code, Cursor) · $steroid-recall (Codex) |
| Arguments | <query> and the project path when needed. global_search is currently disabled/no-op. |
| When to use | You want context, decisions, or solutions from past conversations — or any time you're uncertain about prior work on this project. |
| Runs | The recall_memory MCP tool (and recall_diff for full diffs) — not the shell CLI. |
| Mode | Both Local & Connected |
What it does: Calls recall_memory with the query. If the index isn't built yet, recall can auto-index/register the project and reports doing so. Explicit steroid memory build is also available without LLM calls. Answering sources use node selection followed by synthesis, preferably through the calling agent. On success it lists hits as Session: title · date · source with an excerpt. Skipped-source notes are surfaced on total failure, not guaranteed on partial success. Use recall_turn for a complete cited exchange. When a hit references a truncated diff, use recall_diff to fetch the file path and full before/after. History is searched across supported source stores with relevant history; inference prompts may reach the executor's provider.
steroid-index-project
Register and index the current repository or an explicit path. This differs from IDE setup: wiring MCP does not imply an index or any historical conversations exist. The daemon maintains existing indexes for registered projects; this skill performs the initial registration/build.
steroid-create-workflow
Build a new Activepieces workflow for the user via the create_workflow MCP tool.
| Invoke | /steroid-create-workflow (Claude Code, Cursor) · $steroid-create-workflow (Codex) |
| Arguments | <description of what it should do> |
| When to use | You want a new automation/workflow built. |
| Runs | The create_workflow MCP tool (and repair_workflow on failure) — not the shell CLI. |
| Mode | Connected only |
What it does: Clarifies your intent (trigger, actions, order), then hands the whole request to the tool, which plans the workflow for you — you don't have to pick pieces or wire schemas yourself. It may first offer to reuse an existing matching workflow, then presents a plan for approval: a small terminal prompt plus a Cmd/Ctrl-clickable read-only HTML plan (summary, a diagram, an inputs/output table, the data flow, and copyable curl snippets for running the finished flow), with an Approve & build / Request changes / Reject choice. Once you approve, it asks for any missing credentials, reports progress as each step is built, and — on success — gives you the workflow's inputs and publication/indexing outcome. Draft tests call real integrations and may create records or send messages. Repair is bounded and runs in place; failed builds may remain resumable, and acceptance warnings can accompany publication. This skill only creates — finding or running an existing one is steroid-search-workflow.
NOTE
Agent builds cache credentials locally and upsert encrypted AP App Connections in the cloud. Flow definitions use connection references; do not paste raw secrets into prompts or nodes. See Steroid Flows for the full workflow docs.
steroid-search-workflow
Find and optionally run one of the user's own previously-published Activepieces workflows.
| Invoke | /steroid-search-workflow (Claude Code, Cursor) · $steroid-search-workflow (Codex) |
| Arguments | <what you're looking for> |
| When to use | You want to find (and maybe run) a workflow you built earlier. |
| Runs | The search_workflows MCP tool, then run_workflow if you choose to run it — not the shell CLI. |
| Mode | Connected only |
What it does: Calls search_workflows and lists the matching workflows (each scoped to you). If nothing matches, it says so and suggests steroid-create-workflow instead of guessing. To run a match, it collects the inputs the workflow expects and runs it, then reports the result — including any output. It only finds workflows you that have been indexed for discovery. A published UI flow is not automatically proof of SRP indexing. Search responses do not expose a webhook URL in the inspected contract; copy the generated Live URL from the trigger when you need a direct HTTP endpoint.