YOUR AGENT. A CLOSER LOOK.
Know what you’re about to install.
The hotskills scanner is an MCP tool for your AI assistant. It reads a skill’s instructions and source files, highlights execution and network patterns, and returns findings you can inspect before deciding to install.
This check runs Smokecheck, the scanner behind every security level on hotskills. What Smokecheck is →
check_skill · MCP tool
Ask once. Inspect the evidence.
You ask in your assistant. Your assistant calls check_skill. hotskills returns the report to the same conversation, with the relevant files and line numbers.
Use the hotskills check_skill tool to inspect owner/repo before I install it. Show its security signals, the relevant files and lines, and anything I should review.Example prompt for an assistant already connected to the hotskills MCP server. Replace owner/repo with a public GitHub repository.
Connect first. Then ask your agent.
Connect your assistant with your personal hotskills MCP key. Its tools/list response must include check_skill; general MCP access alone does not activate scanning.
One conversation. Four clear steps.
- You ask
Name the public repository you are considering.
- Your agent calls
The assistant sends a check_skill request through MCP.
- hotskills reads
The static engine inspects the supplied source without running the skill.
- You decide
Review the findings in your conversation before installing.
What deserves your attention?
Instructions
What does the skill ask the agent to do? The scanner looks for suspicious directives and relevant source patterns.
Local execution
Which scripts and executable steps are included? Findings help you locate commands and code worth reading.
Network access
Does the source contact remote systems, fetch installers or pipe content into a shell? Inspect the exact locations.
What the report means
Read the individual findings alongside the level. Each signal identifies its kind, severity, file and line so you can inspect the relevant source. Analysis facts describe what the scanner can establish from the supplied text.
- Instructions only
instructions - Instructions and content without a detected execution requirement.
- Local scripts
local-scripts - Includes local scripts or executable steps that deserve inspection.
- Network or broad shell
network-shell - Includes network or shell execution patterns that require particular attention.
These are automated source signals, not a safety guarantee or proof that a skill performs its advertised task. A level does not replace reading serious findings.
The MCP request and response
check_skill accepts a public repository with an optional ref, selected skill files, or a pending job id. The server needs the tool enabled and its workers running. If a response is pending, the agent uses that job id to retrieve the result. A personal key is required.
For developers: terminal and CI integration
These clients use the same remote MCP service and need an active MCP key. The commands below show the setup; this page does not run them.
Smokecheck: a terminal client for MCP
From a checkout containing the full Smokecheck client directory (including config.json and lang/en.json), use Python 3.10 or newer and Git. Set HOTSKILLS_KEY privately in the client environment; this client uses that name, while the MCP setup guide uses HOTSKILLS_API_KEY. Use the same account-issued key, never a GitHub token. The remote MCP service and check_skill must be activated before a check can complete.
These commands run from a source checkout containing Smokecheck. No public installer is supplied here.
python3 -B smokecheck/smokecheck.py path/to/your-skillpython3 -B smokecheck/smokecheck.py --github your-name/your-skill --ref main --jsonpython3 -B smokecheck/smokecheck.py --markdown --repo your-name/your-skill .Smokecheck is a remote client, including when you pass a local folder: selected file contents are sent to the service. Review them for private material before sending. The client applies Git ignore and file-selection rules; these are not a guarantee that all secrets are excluded.
Default client limits are 200 files, 256 KiB per file and 2 MiB total. The server also enforces its own limits. Reports can be requested as JSON or Markdown; the client is not an offline scanner.
CI integration from a source checkout
If your CI workspace already contains smokecheck/action.yml, a workflow can reference that local action. Store HOTSKILLS_KEY as a repository secret; never put its value in YAML or logs. The action calls the same remote service and needs an activated key, tool and workers.
- name: Check skill sources
uses: ./smokecheck
with:
key: ${{ secrets.HOTSKILLS_KEY }}
path: skills/my-skill
fail-on: network-shellExit 0 means the check completed; exit 1 means the requested fail-on threshold was reached (or the level could not be determined with that threshold); exit 2 means the check could not run. Without fail-on, a reported level alone does not fail the job.
Optional repository form
Checks use temporary private processing. Consumed reports are removed and jobs expire; submitted files and repositories are not added to the catalog. Web checks do not ask for email; MCP clients require a personal key.
Automated signals, not a guarantee
