SkillSpector
NVIDIA/SkillSpectorReviewed by AI
Checked by RepoPicks Radar on
What is SkillSpector?
SkillSpector is a terminal tool from NVIDIA that inspects the skills and MCP servers people install into AI assistants such as Claude Code, Codex CLI or Gemini CLI. You point it at a folder, a SKILL.md file, a zip or a Git address and it looks for 71 patterns in 17 groups (prompt injection, data exfiltration, credential access, obfuscated code, vulnerable dependencies), then prints a 0-100 risk score and a recommendation. A fast pass runs offline with `--no-llm`; the deeper pass sends the files to an AI model you choose. It can also run as an MCP server to check a skill before installing it.
How easy is it to install SkillSpector?
D · terminal, several tools or a specific graphics card · Windows
- Windows D
- macOS D
- Linux D
There is no installer: on macOS and Linux you install Python 3.12+ and uv (`uv tool install git+https://github.com/NVIDIA/skillspector.git`) or build the included Dockerfile, then run commands in a terminal and set environment variables for an AI provider, so D. Windows is D as well but rougher: every example in the README is written for a Unix shell (`source .venv/bin/activate`, `$PWD`), so Docker or WSL is the practical route. A graphics card is only needed if you choose to run the AI model locally with Ollama.
Coming soon: install in 1 click with the RepoPicks app.
SkillSpector on GitHub
- Stars
- 25,199on Nov 10, 2026
- Downloads
- 51,9086,659 this week
- Growth
- +850stars in 6 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Python, Docker
- To build from source
- Python, Node.js, Docker, NVIDIA graphics card
- Disk space
- 200 MB–800 MB · estimate
- Accounts or keys
- NVIDIA Build (NVIDIA_INFERENCE_KEY, the default provider), OpenAI, Anthropic, Azure OpenAI, Any OpenAI-compatible gateway (SKILLSPECTOR_COMPAT_API_KEY), LangSmith tracing (LANGCHAIN_API_KEY, optional)
- Cost
- Free, API costs extra · Apache-2.0 and free to run. The static pass (`--no-llm`) needs no account; the second, semantic pass needs an AI provider — NVIDIA Build, OpenAI, Anthropic, Azure OpenAI, Bedrock or Gemini all bill per use, and a local Ollama model is the free option.
- License
- Apache-2.0 (Permissive)
- Last change on GitHub
- Nov 10, 2026
SkillSpector risk signals (6)
- Attentionalpha software that issues install verdicts
pyproject.toml declares "Development Status :: 3 - Alpha", yet the output is a verdict people act on: scores 51-100 print "DO NOT INSTALL", scores 0-20 print a green recommendation, and the MCP tool returns a yes/no install field. A finding the scanner misses therefore reads as approval.
AI review
- Infoscanned files are sent to an ai provider
The deeper pass needs an LLM: "For the best results, configure an OpenAI-compatible LLM endpoint for semantic analysis", and the default provider is `nv_build` (build.nvidia.com). Whatever skill you scan — including private or client code — is sent to that provider unless you use `--no-llm` or a local Ollama model.
AI review
- Inforesearch numbers cannot be checked here
The README opens with "In the 31,132-skill analyzed subset of the research dataset, 26.1% of skills contain vulnerabilities and 5.2% show likely malicious intent." The dataset is not in the repository, so these figures are NVIDIA's own claim and cannot be verified from this code.
AI review
- Infomcp http mode has no login
The README's own security box: "The HTTP transport ships without authentication. Any caller that can reach the port can invoke scan_skill." It is limited to loopback addresses (127.0.0.1 or ::1) and local paths are rejected over HTTP, and it also notes that for the stdio transport "the initialize hang reported in issue #199 still applies".
AI review
- Infoit downloads what it inspects
`skillspector scan https://github.com/user/my-skill` clones repositories and unpacks zip files you point it at, so untrusted files land on your disk before analysis. Limits exist: `INGEST_MAX_BYTES` 100 MiB and `INGEST_MAX_ZIP_MEMBERS` 10,000, and the install notice warns "This project will download and install additional third-party open source software projects."
AI review
- Infoone provider is disabled for safety
.env.example and the provider table state that `codex_cli` is "registered but disabled: its read-only sandbox permits host-file reads", so that route is blocked on purpose.
AI review
Evidence is shown in its original language.
What Radar could not check about SkillSpector
- How often it misses a genuinely malicious skill, and how many harmless skills it flags
- What one scan costs at the paid AI providers, since a full workflow can run up to 18 minutes and many analysis batches
- Whether the documented install works on Windows without Docker or WSL
- What the chosen AI provider stores of the files you send for analysis
- Whether a low score means much for skills whose risky behaviour only appears while they run