beck

your agent's skills, at its beck and call.

Your agent loads every skill into its system prompt on every turn. 500+ skills = 21,000 tokens before you say a word.

beck indexes your skills and serves them on demand. 200 flat tokens.

Install

curl -fsSL https://raw.githubusercontent.com/notabotchef/beck/main/scripts/install.sh | sh

cargo install beck

Quick start

beck sync                      # index your skills
beck query "transcribe audio"  # search (BM25, 4ms)
beck load whisper              # load the match

MCP (Claude Code, Desktop, Cursor)

claude mcp add -s user beck /absolute/path/to/beck mcp

Two MCP tools: skills_query and skills_load. Session cost: ~200 tokens.

The math

Path	Tokens per turn
Inject everything	~150,000 (300 skills)
beck MCP	~200 (flat)

98% top-3 retrieval recall. Pure FTS5 BM25. No embeddings. Eval harness in tests/eval/. Run cargo test --features eval --bin eval.

Commands

Command	What it does
`beck sync`	Index SKILL.md files from `~/.hermes/skills` and `~/.claude/skills`
`beck query "<text>"`	BM25-ranked search across name, description, tags, body
`beck load <name>`	Print the full skill body
`beck mcp`	MCP server over stdio
`beck link`	Write skills once, install into every agent
`beck check`	Diagnose agents, orphans, collisions
`beck bench`	See your real token savings
`beck prompt`	Print the agent integration stub

2.0 MB binary. Zero network calls. Zero telemetry. macOS + Linux.

License

MIT OR Apache-2.0