beck-0.2.1 has been yanked.
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
|
Quick start
MCP (Claude Code, Desktop, Cursor)
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