jev-routing
Archived Go harness finding Jev tool steering does not reliably cut agent tokens in Claude Code or Codex; context trimming helped.
Overview
jev-routing was a Go harness for Claude Code, Codex and Grok Build that tested whether Jev-based tool selection and steering could reduce the upstream tokens AI coding agents use while preserving quality. It was archived on 26 September 2026 with its conclusions and trial-and-error record kept public; Jev's own cost was excluded from the metric.
The main finding is negative: forcing Jev's tool choice while keeping the tool list reduced tokens by a median 19.3% on one Codex model but increased them by 37% and 51% on two others, sometimes roughly doubling upstream requests, while advice-only steering on Claude Code models had little effect. What did help was shrinking context: Anthropic-native clearing of old tool uses gated by Jev gave a median 11.1% net reduction in runs where clearing occurred (estimated about 6.3% overall), and deterministic truncation of large tool results gave a median 8.19% under load testing, with an interval crossing zero. Results come from preregistered six-pair comparisons on a fixed task.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which tool should the agent call next? | — |
| Yes / no | Should this old tool use be cleared from context?(inferred) | yes · no |
How it works
- Run baseline and intervention sessions in preregistered pairs
- Apply Jev tool steering or context clearing
- Measure upstream input and output tokens
- Compare medians per pair and totals
Strengths
- Honest negative results with data
- Preregistered comparisons
- Identifies context trimming as the useful lever
Caveats
- Archived; no further development
- Results specific to tested models and one task
- Small number of pairs
At a glance
- Type
- MCP server
- Status
- research
- Language
- Go
- Runtime
- Agent harness (archived)
- Has code
- yes
Use cases and integrations
How curated lists describe it
Go Jev harness for Claude Code, Codex, and Grok Build. No npx. Not an MCP server.
hellogumbo/awesome-jev
Frequently asked questions
Did Jev tool steering save tokens?
Not generically; it helped one Codex model and increased tokens on two others.
What did work?
Shrinking context through gated tool-use clearing or truncating large tool results.
Is it maintained?
No. It was archived on 26 September 2026.
Related projects
Local compaction proxy for Claude Code and Codex that compacts more often with local rules and keeps full history searchable.
ProxybetaGoLocal gateway for coding agents that asks Jev which tool to call next and passes everything else to your usual model.
ServicebetaTypeScriptClaude Code (and Codex) plugin that trims long Bash output chunk by chunk with Jev before it reaches the model, without summarising.
PluginbetaTypeScriptToolkit where Jev condenses large MCP responses for Claude Code and maps Codex requests to reviewed, fixed repository operations.
ProjectbetaTypeScriptAgent skill that has Jev pick the few relevant lines from big logs before your coding agent reads them, with a reversible reduce and ledger.
ProjectbetaPythonGives an agent one tool that uses Jev to search a nested capability tree and return only the documented commands a task needs.
ProjectprototypeTypeScript