ZenNotes file ops benchmark — MCP vs zen CLI vs plain Unix

[TomSchimana]

Hi all,

I'm putting together an AI substrate for a personal second brain on top of a Markdown vault, with ZenNotes as the editor. Before locking in which path the agent should use for routine file ops, I ran a quick benchmark and figured I'd share the numbers in case anyone else is wondering.

Setup: 10 Markdown notes per path, each with a tiny Lorem-Ipsum body and ten reciprocal [[wikilinks]] to the others. For each path (MCP, zen CLI, plain Unix) I ran the same four phases: create 10, move 10 into a sibling folder, read 10, hard-delete 10. Wall-clock is end-to-end. For MCP that includes the agent's tool-call round-trip per op, because that's what the agent actually pays in practice.

Phase	MCP	zen CLI	Unix
Create 10	82.7 s	0.78 s	0.005 s
Move 10	72.6 s	0.77 s	0.019 s
Read 10	~178 s (projected from 2)	0.76 s	0.015 s
Delete 10 (hard)	66.7 s	0.76 s	0.015 s
Total 40 ops	~400 s	3.07 s	0.054 s
Payload returned to agent	~30 kB	3.3 kB	3.3 kB

Roughly MCP / CLI ≈ 100×, CLI / Unix ≈ 57×.

Couple of things I noticed:

• The MCP wall-clock is mostly the agent's tool-call round-trip per op, not the server itself. But that's still what an agent pays when it runs a bunch in a row.
• The zennotes MCP also loads its tool defs into the agent's context every session, around 5-9 kB tokens whether you use them or not. The CLI loads nothing until you call it.
• Wikilinks after a move behaved the same on all three paths. Link text stayed unchanged ([[note-01]]), and ZenNotes resolves by title so backlinks kept working as long as titles stayed unique.
• Unix skips the ZenNotes index. Fine for normal file ops, but trash and archive lose their original-path metadata and restore breaks. Use MCP or CLI for those.
• For full-text search, grep / rg / find are still the fastest. This benchmark only covers mutate and read.

Tom