cli-registry
I’ve been thinking about agent tooling more.
CLI is all you need and similar social posts have the strong impression that the models were magic and would just know how to use any CLI it was presented with.
I just… I don’t think that’s true at all?
Models are clearly trained on popular CLI like GitHub’s gh so of course they know how to use that. What happens if you make their own? You think Claude will know exactly the interaction patterns necessary to navigate your terminal-driven flow? Doubt it.
This isn’t just vibes. Arize ran the same Claude through the same GitHub tasks four ways (the official MCP server, two different gh skills, and a bare shell) and bare Claude scored highest. But like… of course it did. Anthropic’s models likely read every tutorial about it during training. Mine (and more importantly yours) didn’t get that treatment.
You know what models do know how to use? Tools. They just love tools. The description is right there for them in their system prompt to describe how to use it!
MCP injects such tools. But wiring up a decent MCP server is annoying.
And closed-source harnesses don’t let you just add more tools to the harness.
So like any rational curious dev I thought what if you could just register a CLI and turn it into model tooling via an MCP proxy. Would that even work? Is it even a good idea?
The Scaffolding Matters More Than the Interface showed that the rest of the harness matters more than MCP vs CLI. Or at least that’s my read. So maybe it’s fine.
So I’m experimenting with cli-registry
The whole registration is a manifest:
{
"version": 1,
"tools": {
"tighten-lint": {
"description": "Run writetighter's deterministic prose lint on a file (requires writetighter on PATH).",
"command": ["writetighter", "lint"],
"timeout": "60s",
"inputSchema": {
"type": "object",
"properties": {
"path": { "type": "string", "description": "file to lint" },
"kind": { "type": "string", "enum": ["pr", "procedure", "description", "code-comment", "reference", "decision", "incident", "agent-instruction", "status-update"] }
},
"additionalProperties": false,
"required": ["path"]
},
"bindings": [
{ "property": "path", "style": "positional" },
{ "property": "kind", "style": "space", "flag": "--kind" }
],
"output": { "mode": "raw" }
}
}
}
And here’s what the model actually receives in its tools/list:
{
"name": "tighten-lint",
"description": "Run writetighter's deterministic prose lint on a file (requires writetighter on PATH).",
"inputSchema": {
"type": "object",
"properties": {
"kind": { "type": "string", "enum": ["pr", "procedure", "description", "code-comment", "reference", "decision", "incident", "agent-instruction", "status-update"] },
"path": { "description": "file to lint", "type": "string" }
},
"required": ["path"],
"additionalProperties": false
}
}
Still early days. Could be interesting. Could be context bloat.