Published
Updated
JevScout ranks oversized MCP text and prints 38% less agent time on its own check
This topic was created 31 days ago, and the information it contains may have evolved or changed since then.
Tim Lin posted JevScout on September 29. The MIT hook leaves Claude Code's MCP result alone under 100 KB, and otherwise asks Jev which segments to keep verbatim. The README's check is 38% less agent time and 26% less agent cost, with TypeSafe's bill separate. One printed row is 5 of 5 against 1 of 5. The evaluation note does not list all seven tasks. On September 30, 2026 this desk scored seven stand-in tasks, three times each, on jev-1.13.0. The critical line stayed in all 21 filtered runs. Claude Code was not installed, so the 38% and 26% were not measured.
Tim Lin posted JevScout on September 29. It is a Claude Code hook, MIT licensed, in yctimlin/JevScout. When an MCP tool returns a large text result, Jev scores pieces of it. The hook keeps the chosen pieces verbatim, marks what it left out, and stores the original locally so a command can read it back. If there is no TypeSafe key, if the call fails, or if the best relevance estimate is below 0.5, the original result goes through unchanged. Images are not touched.
The default scope is narrow. The result has to be one Claude Code would already have saved as too large, and it has to be at least 100 KB. Smaller results pass through. The packet sent on is at most 6,000 bytes. The timeout is 8,000 ms. The README says the package is not on npm yet. Install is from a checkout, and the tested Claude Code version on the page is 2.1.281. A keyword-only mode exists and the page says not to use it for paraphrased requests. A Codex proxy and a shell hook are marked experimental.
The post says that across 7 tasks the hook cut agent time 38% and agent cost 26%, with no loss of correctness. The README status table uses the same two percentages. The category table prints four of those tasks, five runs each. A documentation paraphrase at 176 KB stays 5 of 5 and drops median agent time 43% and agent cost 29%. An issue search at 170 KB stays 5 of 5, at 51% and 42%. One hundred comments at 196 KB go from 1 of 5 without the hook to 5 of 5 with it, at 72% and 60%. An npm metadata response of 5.2 MB stays 5 of 5, at 45% and 29%. The evaluation note repeats 38% and 26% and does not name the other three tasks. The demo caption describes one clip as 169 KB, two turns instead of five, and the same 51% and 42% as the issue-search row.
TypeSafe’s bill is separate. The README prints 9,000 to 81,000 TypeSafe tokens on the condensed tasks around 170 to 196 KB, and about 300,000 tokens for the 5.2 MB response. The hook sends segment text and the request to api.typesafe.ai. Originals stay under ~/.cache/jevscout/outputs and are deleted after 7 days. We did not run the seven tasks in the README, and the evaluation note still does not name three of them.
On September 30, 2026 this desk wrote seven stand-in tasks, three repeats each, and scored them with jev-1.13.0 using the hook’s 0.5 keep rule. All 21 filtered runs kept the critical line and dropped the decoy. Decoy scores were 0.03 to 0.32. Critical scores were 0.87 to 0.97. The list cost was $0.001814 for 43,191 input tokens. A no-filter control also kept the line on 21 of 21 runs, at no TypeSafe tokens, because it passes the text through. Claude Code is not installed on this machine, so the 38% agent-time and 26% agent-cost figures were not measured. The planted lines were obvious, and the repository’s fixtures were not in the clone.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Post by @ycalintim, September 29, 2026, 05:41 UTC. Text: across 7 tasks, 38% less time, 26% less cost, no correctness loss. The repository is yctimlin/JevScout. LICENSE is MIT, copyright 2026 JevScout contributors.
The README status table repeats 38% and 26% for the Claude Code hook and says the package is not published. The category table prints four rows, each 5 runs: documentation paraphrase, 176 KB, 5/5 against 5/5, agent time −43%, agent cost −29%; issue search, 170 KB, 5/5 against 5/5, −51% and −42%; 100 comments, 196 KB, 5/5 against 1/5, −72% and −60%; npm metadata, 5.2 MB, 5/5 against 5/5, −45% and −29%.
docs/evaluation.md repeats the 38% and 26% and does not name the other three tasks. The demo caption says 169 KB, 2 turns instead of 5, and the same −51% and −42% on that one task.
TypeSafe usage on the condensed tasks is printed as 9K to 81K tokens at 170 to 196 KB, and about 300K tokens for the 5.2 MB response, billed apart from agent cost. Default floor is 100 KB and Claude Code's own oversize notice. Best estimate below 0.5 passes the original through.
Packet budget 6,000 bytes. Timeout 8,000 ms. Tested Claude Code version in the README is 2.1.281. Node.js 24+. We did not run the hook.
On September 30, 2026 this desk wrote seven stand-in tasks, because the README names four of seven and the repository does not ship the fixtures. Each task had one critical line and a decoy. The seven ids were redirect, timeout-ms, header, retry-count, sql-lock, clock, and deny. Three repeats made 21 runs. The call used the hook's Noul rule: keep a segment at 0.5 or above, and if the best score is under 0.5 pass the original through. The packet budget in that rule is 6,000 bytes. The request alias was jev-latest. Responses named jev-1.13.0. All 21 runs kept the critical line and dropped the decoy. Critical scores were 0.87 to 0.97. Decoy scores were 0.03 to 0.32. Input tokens were 43,191 and output tokens were 2,940. At $0.042 per million input tokens, output at $0, the list cost is $0.001814. That bill is only the TypeSafe calls for these stand-ins.
The no-filter control also kept the critical line on 21 of 21 runs, at 0 tokens, because it does not delete anything. Filtering did not drop the planted line. The lines were written to be obvious. Claude Code is not installed here, so agent time and agent dollar cost were not measured, and the README's 38% and 26% were not reproduced. The 100 KB floor, the 8,000 ms hook timeout, and Claude Code 2.1.281 were not exercised. The three tasks the evaluation note does not name are still unnamed.
Compare
jevgrep also asks Jev which files matter, then leaves the patch to another model. Its printed set is 10 SWE-bench tasks, 8 of 10 on both arms. JevScout's printed set is oversized MCP text inside Claude Code, and one of the four rows changes the correct count.
toolgate scores a shell call before it runs. This hook does not.
Terms
- 38%
- The README's aggregate drop in agent time for the default hook against Claude Code with no hook. Agent cost on the same sentence is 26% lower. TypeSafe's own token bill is not in that 26%.
- 5/5 against 1/5
- The comments row, 196 KB, five runs each. JevScout was correct on all five. The no-hook path was correct on one. The other three printed rows are 5/5 against 5/5.
- 0.5
- The README's bar. If Jev's best relevance estimate is below 0.5, the original MCP result is left unchanged.