← MCP server and skill grades

skills

JuliusBrussee/caveman/skills

C
Type
Skill
Scanned
Nov 13, 2026
Version
b4ee71d557d9
Files read
64

This project is a collection of skills for AI coding agents that shorten agent responses, compress markdown memory files, and wire repositories into Caveman Cloud for LLM cost tracking and optimization. The biggest risk is that the setup skill tells the agent to send live LLM provider API keys to a prompt-supplied gateway URL without asking permission first, while the compression script leaves Claude CLI read tools enabled in the working directory when compressing untrusted files.

What it can reach

Files
caveman-compress reads markdown and text files, overwrites them in place, and writes backups and locks under the user data directory; editing skills (caveman-setup, caveman-discover, caveman-learn, caveman-optimize) instruct the agent to read and modify repository source files, .env files, and agent memory files.
Network
caveman-compress sends the text of compressed files to Anthropic, OpenAI-compatible HTTP endpoints, or the claude, codex, or opencode CLIs; caveman-setup routes app LLM traffic through a Caveman Cloud gateway URL, sends a verification request with curl, and fetches remote integration docs.
Shell commands
caveman-compress spawns claude, codex, or opencode subprocesses, and several skills instruct the host agent to run python3, curl, project test scripts, and caveman CLI commands.
Secrets
caveman-compress reads ANTHROPIC_API_KEY and CAVEMAN_COMPRESS_API_KEY from environment variables, and caveman-setup writes CAVE_API_KEY to .env files and sends existing OPENAI_API_KEY or ANTHROPIC_API_KEY values to the configured gateway in bring-your-own-key mode.

Findings

1 high2 medium
  1. HighSetup skill tells agent to send provider API keys to unvalidated gateway URL without asking

    skills/caveman-setup/SKILL.md:147

    Unlike the other editing skills in this pack, caveman-setup can be invoked automatically by the model (disable-model-invocation is not set), accepts any GATEWAY URL from the conversation without checking that it belongs to gateway.caveman.so or uses HTTPS, and explicitly tells the agent not to pause to ask permission before sending a live verification curl request. In bring-your-own-key (byok) mode, that request includes the repository's existing OPENAI_API_KEY or ANTHROPIC_API_KEY in the x-cave-upstream-key header. If untrusted text in an issue, PR, or webpage tricks the agent with a fake setup prompt pointing GATEWAY at an attacker's server, the skill directs the agent to exfiltrate the user's live LLM API key and point repository LLM clients at that server.

    Fix: Add disable-model-invocation: true to the frontmatter so only the user can invoke /caveman-setup, restrict GATEWAY to https://gateway.caveman.so or loopback unless explicitly approved by the user, and require user confirmation before editing files or sending API keys over the network.

  2. MediumSetup skill fetches and follows unpinned remote instructions at runtime

    skills/caveman-setup/SKILL.md:134

    When wiring repositories that use google-genai, crewai, pydantic-ai, or openai-agents (and again when offering workflow discovery), caveman-setup tells the agent to fetch markdown pages from <docs origin>/docs/ over the network and follow them to edit repository code. Because <docs origin> is an unpinned placeholder inferred from prompt-supplied URLs, either a prompt-injected origin or a change to the remote webpage can deliver unreviewed instructions to an agent actively editing source files.

    Fix: Include the integration patterns for supported frameworks directly inside skills/caveman-setup/SKILL.md and delegate workflow labeling to the local caveman-discover skill instead of fetching remote markdown instructions.

  3. MediumCompression fallback leaves Claude CLI built-in tools enabled in the working directory

    skills/caveman-compress/scripts/compress.py:672

    When ANTHROPIC_API_KEY is not set, caveman-compress sends the target file's prose as a prompt to the claude --print CLI. Whereas call_codex_cli runs inside an empty temporary directory and disables all tools, and call_opencode_cli denies all tool permissions, call_claude_cli omits --tools "" and run_cli inherits the caller's working directory (which also lets opencode load repo-level opencode.json MCP servers if run inside a repository). Because Claude CLI allows built-in read-only file tools (Read, Glob, Grep) inside the workspace/git root in non-interactive mode and compress.py only scans for credentials before calling the model, prompt injection inside a markdown file being compressed can cause the inner claude process to read sensitive workspace files such as .env and embed their contents into the output or validation errors.

    Fix: Pass --tools "" in call_claude_cli to disable all built-in tools, execute run_cli with cwd set to an empty tempfile.TemporaryDirectory(), and run find_credential() on the returned text before staging or printing validation diffs.