User repository on GitHub
Portable JevRouter setup for cost-aware Codex subagent model selection, with English and Russian instructions
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Codex Jev Router
A local MCP server that gives Codex a few cited excerpts instead of a whole noisy search result, with a deterministic default selector that needs no Jev account or key.
- Stars
- 7
- Language
- JavaScript
- License
- MIT
- Published by
- Community
- Author
- suenot
- Last verified
- Oct 2026
About this project
Recommended use: opt-in deterministic evidence selection for noisy-log investigations. A local MCP tool searches a workspace and gives Codex a few exact, cited excerpts instead of a large raw log result. The continuing Codex session keeps its configured model and reasoning effort. The default selector is deterministic and needs no Jev account, API key, or local model. Laya remains optional for experiments; it did not improve cost over deterministic selection in our holdout.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- JavaScript
Highlights
- Two tools: search_workspace_evidence returns short exact excerpts with relative paths and line numbers, read_selected_evidence widens one when it is not enough.
- The default selector is deterministic and local, so no Jev account, API key or local model is involved; a --selector laya flag is optional.
- Search and response sizes are bounded, and hidden files, known sensitive paths and anything outside the chosen workspace root are excluded.
- The installer only adds a codex-local-evidence server entry to the global Codex config, backs up what it changes, and touches no model or reasoning settings.
Watch out
The savings come from 16 tasks, and the four noisy incident logs used for the headline number were deterministic overlays on two pinned public repositories, not production logs. Codex made no MCP calls at all on the multi-file tasks, so the small difference there is run variation rather than a tool benefit, and it made none on exact lookups or edits either. Two earlier diagnostic attempts were excluded because they leaked a grading rule. Optional local Laya cost 8.8% more than deterministic selection on eligible tasks, which is why it is not the default. The tool does not promise to catch every secret in ordinary visible source, and selected excerpts still enter Codex context. Cost figures are API estimates that exclude local compute and are not a subscription bill. The old subagent-model router stays in the tree for reproducibility but is not installed and its own benchmark found Jev-picked children more expensive than a single xhigh agent. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
By default, no Jev is involved. Selection is deterministic and local, and the installer never selects hosted Jev. The routing code that used Jev for subagent models is still in the tree for reproducibility but is not installed or recommended; Jev only comes back if you explicitly invoke a legacy routing command.
Split by task type, and mostly yes for noisy logs: 41.5% less on noisy-log tasks and 29.9% on the eligible log and multi-file group, dropping to 19.8% across all 16 tasks. The honest catch is that Codex never called the tool on the multi-file tasks, so the 2.4% difference there is variation. Exact lookups and edits produced no MCP calls either.
The server is scoped to the workspace root you pass the installer and excludes hidden files, known sensitive paths and anything outside that root. It does not claim to detect every secret in ordinary visible source text, and whatever excerpts it selects still enter Codex context, so it reduces volume rather than acting as a redaction layer.
No. The installer adds one marked MCP server section to the global Codex config, backs up what it changes, removes a stale routing block from AGENTS.md if present, and leaves the model, reasoning effort, role files and skill hooks alone. Remove the marked section from config.toml to disable it. The server can appear in other workspaces even though its file access stays limited to the root you chose.
Node.js 20+, ripgrep (rg) on the path, and a Codex installation that supports local MCP servers. From the repository you run npm ci, then node scripts/install.mjs --root /absolute/workspace, optionally adding --dry-run to preview or --codex-home for a nonstandard Codex directory.
Ask Codex to use search_workspace_evidence on a broad code or log investigation and check that the returned result starts with "Selection: deterministic". Tool discovery alone does not prove a token or cost saving, and after migration Jev stays disabled unless you explicitly invoke a legacy routing command.
It only reorders sufficiently large candidate sets through a locally running Laya service and falls back to deterministic selection for small sets or failed decisions, stating which selector ran. In the holdout, 14 searches used genuine Laya selection and cost 8.8% more than deterministic, so it is not the default and hosted Jev is never selected.
More Jev resources like this
14GitHub stars
Opt-in Hermes plugin that routes each turn to one skill: a wide Jev Choice plus request checks gate a narrow fits round, injecting a single ignorable hint line.
Jev Agents & MCP#community#jevbest#catalog-import
7GitHub stars
MCP server exposing one evaluate tool for Jev Choice, Noul, and Score questions, plus a separate eval harness that compares Jev against LLMs on shared labeled sets.
Jev Agents & MCP#community#jevbest#catalog-import
Community
arunav25/jev-mcpJavaScript
arunav25/jev-mcpJavaScript9GitHub stars
MCP tool router for Codex that keeps large external tool inventories behind a small find_tool surface and asks Jev to select from names and descriptions at a 0.90 threshold, falling back to bounded search.
Jev Agents & MCP#community#jevbest#catalog-import
Community
jackbarunz/jev-tool-routerJavaScript
jackbarunz/jev-tool-routerJavaScript6GitHub stars
LangChain deepagents middleware that replaces SkillsMiddleware with per-turn judging, loading only the SKILL.md files a turn needs so large catalogs stay out of the prompt.
Jev Agents & MCP#community#jevbest#catalog-import
4GitHub stars
OpenClaw plugin adding a jev_route tool that classifies task type, complexity and risk with TypeSafe Jev, then maps the result to a route through testable JavaScript policy.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
trietphan/jev-clawJavaScript
trietphan/jev-clawJavaScript5GitHub stars
A Claude Code function hook that sends AskUserQuestion dialogs to Jev and draws a probability per option above the dialog, so you still choose while seeing how confident the model is.
Jev Agents & MCP#community#jevbest#catalog-import
Community
BeLazy167/typesafe-modTypeScript
BeLazy167/typesafe-modTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.