Measured 2026-09-18 on codemode 0.8.1. Method and environment: PROTOCOL.md, written before the numbers so the design could not be fitted to the result.

Aside ships a real ripgrep — runtime/native/bin/rg, 6,476,288 bytes, 15.2.0 with pcre2, first on the agent PATH — and no documented harness tool reaches it. That directory is the skill runtime's utility bin, holding pdftotext, pdftoppm, pdfinfo, python3 and node alongside rg, all vendored from Homebrew (runtime/manifest.txt). The documented agent tools are read_file, write_file, edit_file, bash and repl; none is a content search.

So an agent's actual options are POSIX grep and find through bash, or codemode. Comparing against rg would measure a route the agent cannot take. Ripgrep numbers are kept below as a ceiling, not a baseline.

All codemode numbers are from the resident MCP server. The CLI spawns a node process per call and pays 60–70 ms of startup that the MCP path does not; CLI figures are in evidence/bench-260918/local-search.md and are not product numbers.

N=7 after one discarded warm-up. Median with range. Result counts matched on every cell.

Find .js files containing throw new, take the first three by path, read them, return the names and total size. Both sides resolve to the same three files and the same 36,497 characters.

Wall clock is a tie. The difference is what lands in the transcript:

• grep -rl then cat: 39,154 bytes across 2 round trips
• codemode: 90 bytes in 1 round trip

435×. Not compression — placement. The shell pipeline puts the file list and the full contents of three files into the model's context, because reading a file is how an agent reads a file. Codemode does the same reads inside the sandbox and returns only the answer.

Counting function across ~/Developer is the interesting failure. grep -r -I did not run slowly; it ran past 128 seconds having emitted 3.23 GB and 3.7 million lines, still incomplete, and was aborted. Even grep -r -I -c reached 78 MB of per-file counts before the cap.

codemode answers the same question in 2,966 ms and returns 778 bytes: 502,963 matching lines across 57,403 files.

An agent holding only POSIX tools cannot ask this question at all. The output would bury the conversation before the answer arrived.

What a native engine does on the same tree, for reference. Not available to an Aside agent without shelling out.

Level, because search.* shells out to that same ripgrep. codemode is not a faster search engine; it is the same engine with the result filtered before it reaches the model.

Eight fixed URLs, three runs.

Parallel gain 2.69–2.87× across runs, against 8,354 ms of summed navigation. Native cost is three calls per page (open, snapshot, close); the accessibility trees are what a reading agent actually receives.

The MCP server exposes one tool. Its full definition is 2,433 bytes (2,034 description, 274 input schema) and covers 34 guest actions whose own descriptions total 44,330 bytes — 18.2×. browse.exec alone is 8,249 bytes described, more than three times the entire resident tool surface. Schema is paid every turn; action descriptions are fetched only when actions.describe is called.

Re-measured at 0.9.0, after the completeness work widened eleven signatures and the resident description named the 64 KiB return-path cap. The four numbers moved together and the note records the run: aside-mcp-attach-260918.md.

• Single small grep. Roughly a tie on time and larger on bytes: 43,948 against grep's 35,750, because rows carry structure and a completeness envelope. If one line answers the question, one line is the right tool.
• Plain file listing against ripgrep directly — 188 ms vs 183.5 ms. Against find it wins by 2.6×, but that is find losing, not codemode winning.
• Unpruned counting against ripgrep: 3,483 ms vs 3,045 ms.
• CLI transport: 93.8 ms for a search the MCP path serves in 12 ms. Use the MCP route.

One machine (Apple silicon, 12 cores, 24 GiB, macOS 27.0, node v26.5.0), warm filesystem cache, one repository, one network location. Different trees and cold caches will differ. Round trips are counted as agent tool calls, which is the unit an agent pays; a human writing one shell pipeline pays one.

complete: true is not a guarantee that every matching file was returned — see issue #41 and evidence/bench-260918/mcp-completeness.md. A default search on the fixture returned 8 of 15 known matches and still reported complete.