Agent Skills for four tools: three open-source projects (Frame, adaptive-classifier and PTS) and the mlx-optiq package for Apple Silicon, plus rsi-skill, which measures and improves prompts, and ellora-skill, which fine-tunes models with LoRA recipes. Each skill is a folder with a SKILL.md in the open Agent Skills format, so it works in any agent that reads skills: Claude Code, Codex, Cursor, OpenCode, Gemini CLI, GitHub Copilot and others.
| Skill | What it does | Tool |
|---|---|---|
| frame-security-scan | Scan Python, JavaScript/TypeScript, Java, C# and C/C++ code for vulnerabilities, triage the findings, generate and verify fixes, re-scan | lambdasec/frame |
| adaptive-classifier | Build text classifiers that learn new examples and new classes without retraining, and evaluate them honestly | codelion/adaptive-classifier |
| pts-pivotal-token-search | Find the tokens, sentences and latent concepts that change whether a model solves a task, and turn them into datasets | codelion/pts |
| optiq-local-models | Pick, serve and check a quantized model on an Apple Silicon Mac with an OpenAI-compatible endpoint | mlx-optiq |
| rsi-skill | Score a prompt on your own test cases, then hill-climb it to a measurably better version with any open model. GSM8K example: 17% to 92% on held-out problems | OptiQ Cloud, a local optiq serve, or any OpenAI-compatible API |
| ellora-skill | Fine-tune a model with a LoRA adapter for one capability (4-bit accuracy recovery, reasoning, tool calling, long context, secure code, execution traces, a new domain from an invented dataset): generate the data, train, compare with the base model on held-out data | codelion/ellora (PEFT on CUDA; mlx-lm route for recipe 7 on Apple Silicon) |
With the skills CLI, for any supported agent:
npx skills add lambdasec/skills
npx skills add lambdasec/skills --skill frame-security-scan # just oneAs a Claude Code plugin marketplace:
/plugin marketplace add lambdasec/skills
/plugin install frame-security-scan@lambdasec-skills
Or install any one skill in a single step (Claude Code 2.1.275 or later):
/plugin install frame-security-scan --marketplace lambdasec/skills
/plugin install adaptive-classifier --marketplace lambdasec/skills
/plugin install pts-pivotal-token-search --marketplace lambdasec/skills
/plugin install optiq-local-models --marketplace lambdasec/skills
/plugin install rsi-skill --marketplace lambdasec/skills
/plugin install ellora-skill --marketplace lambdasec/skills
The frame-security-scan plugin also connects the frame MCP server (uvx frame-mcp, needs uv), so Frame's scan, verify and fix are available as tools. Outside Claude Code: claude mcp add frame -- uvx frame-mcp or uvx frame-mcp in any MCP client.
Or copy a folder from skills/ into your agent's skills directory, such as .claude/skills/ or .agents/skills/.
The skills are plain Markdown with no bundled scripts, except rsi-skill, which ships one standard-library Python script (scripts/rsi.py) that sends your prompt and test cases to the model endpoint you choose. ellora-skill ships Python scripts converted from the Ellora notebooks; they download models and datasets and train on your GPU or Mac, so the skill tells the agent to ask before running them. They do not install software: each one checks that its tool is already installed and points to the tool's own install steps if it is not. The one exception is the frame-security-scan plugin's MCP server entry: when Claude Code starts it, uv downloads frame-mcp and Frame from PyPI. They ask before anything downloads model weights or datasets, sends code to a network endpoint, changes files, or publishes to the Hugging Face Hub. Numbers quoted in the skills come from each project's README or notebooks, with their caveats.
Apache-2.0. See LICENSE.