Jev × Claude Code · updated 2026-09-27
Jev with Claude Code
Claude Code is the harness; Jev is the decision. The integrations we have read both replace a step where a model was writing something when what the loop needed was a choice: which tool calls survive a long session, which harness should take the next task, which memory is worth surfacing. None of them ask Jev to produce a sentence, and that is the pattern worth taking from them.
In short
- For: decisions taken *inside* a coding session — compaction, routing, reranking — at decision-model latency and price ($0.042 per million input tokens, no charge for output).
- Not for: writing code, summaries or explanations. Jev gives string generation up by design, so any integration that wants text has to hand it back to the LLM.
- Evidence: two repositories with our own dossiers (9,300+ stars between them, read 2026-09-23 and 2026-09-25) and two first-person posts. Neither repository publishes a before/after measurement of what it saves.
Who owns which job
The point of a pairing page is this table, not the two names in the title. Every row is a job that has to happen in the loop, the decision it needs, and which of the two should own it — including the row at the bottom, where the answer is "not the decision model".
| Job | What the decision is | Who should own it | Why |
|---|---|---|---|
| Choosing which tool calls survive a long session | one score per call and result; the stale ones are deleted or truncated | Jev | A summariser rewrites what it keeps, and a rewritten file path or error message is a different fact. Deleting is reversible; paraphrasing is not. |
| Routing a task to the right harness | one choice out of the harnesses already on the machine | Jev, before delegation | The rules were already written in the system prompt. A hook that decides before the hand-off is what turns a written rule into a deterministic one. |
| Reranking what an agent remembers | a relevance score per candidate memory | Jev, as an optional filter | It sorts candidates another system already stored and retrieved. It does not embed, store or own the memory — turning it on is one line of config. |
| Writing the code, the summary, the answer | text | the LLM — not Jev | Jev returns a typed decision with a probability. Asking it for prose is asking the wrong model, and the vendor's own launch post says so. |
What we hold on this pairing
Everything we hold on this pairing, newest reading first. Two repositories have a page of ours; the posts link to our write-up of the post.
-
6,703 stars · MIT · read 2026-09-23
tamaratran/fast-jev-compaction
Claude Code plugin that replaces the compaction summary with Jev decisions: every tool call and result is scored in one fast request, stale ones are dropped or truncated, everything kept stays verbatim.
-
2,845 likes · read 2026-09-24
The co-founder's own nod to the Claude Code workflow
A Chinese-language clip about long coding sessions where the context window fills up. The co-founder reposted it, which is how it reached this list.
-
2,651 stars · MIT · read 2026-09-25
zilliztech/memsearch
A persistent, unified memory layer for all your AI agents (e.g. Claude Code, Codex, DSH), backed by Markdown and Milvus.
-
839 likes · read 2026-09-24
Using Jev for local harness and model routing
A developer using a decision model to route between Claude Code, Codex and opencode: the practical version of the routing argument made elsewhere in this column.
What we have not checked
- We did not install either plugin into a session, so how much context compaction actually saves is unmeasured here.
- Neither repository publishes an evaluation of the reranker's effect on retrieval quality.
- Star and like counts are snapshots read on the dates printed beside them.
READMEs and licence files read 2026-09-23 and 2026-09-25 like counts read 2026-09-24Jev's price from the OpenRouter listing, 2026-09-18Jev × Claude Code, updated 2026-09-27