I've been building agentic systems and kept running into the same problem:
agents either stuff entire tool docs into every prompt, or hallucinate
the wrong tool entirely.
CLI tools are 10–32× cheaper than MCP and nearly 100% reliable — but
there's no standard way for agents to discover and invoke them.
So I built CLIbrary: an open standard where each CLI tool has a
manifest.json with intent_triggers — natural language phrases
describing when to use it.
A routing layer does nearest-neighbor search over these embeddings
and returns a ready-to-execute tool call.
Results
Validated on 1,380 test cases using multilingual-e5-base, zero fine-tuning:
- Overall accuracy: 93.6%
- All 8 categories above 90%
Current state
69 manifests across 8 categories (ai-ml, devops, data, web, security,
media, productivity, finance). Schema and metadata only right now —
no CLI implementations yet. This is the standard layer first.
Manifest example
{
"name": "sql-runner",
"category": "data",
"intent_triggers": [
"query a database",
"run a SQL statement",
"get data from postgres"
]
}
Looking for feedback
- Does the manifest schema make sense for your use case?
- What CLIs or categories would you want to see first?
- Is anyone else building tooling for agent tool discovery? Would love to connect.
Top comments (0)