User repository on GitHub
Grok Build with TypeSafe Jev routing tool selection once per turn: 22 to 40% cheaper on the same tasks
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Nitro
Grok Build wrapper that asks Jev one noul per tool before the first model call of a turn, then holds the shortened tool list so the prompt cache survives.
- Stars
- 5
- Language
- Rust
- Published by
- Community
- Author
- daniel-farina
- Last verified
- Oct 2026
About this project
Grok Build with TypeSafe Jev in front of tool selection. The command is nitro. Once per turn, before the first model call, Nitro asks Jev one question per tool: "will finishing this request need this tool at any step?" All the questions go in one request and come back in about 350 ms. The tools Jev says no to are left out of the chat model's request, and that shorter list is held for every step of the turn. The chat model still writes every command, path, and patch. Jev only decides which tools are in play.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- Rust
Highlights
- Keeps 8 of 25 tools on a coding task, cutting schemas from about 11K tokens to about 2.9K every request.
- Adds about 350 ms once per turn with a 0.30 keep threshold; core workspace primitives always stay.
- Re-applies the same keep-set on later steps without new calls, after a per-step version proved costlier.
- Leaves the full list on failed calls, timeouts, or missing answers, and never forces tool_choice.
Watch out
Two runs per cell is a small sample, not a distribution; Jev picks tools only and the chat model still writes every command, path, and patch. 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.
Both ran the same model, prompts, and outputs with means shown, and every run produced the asked files and a passing suite. The earlier per-step version cost 70% more because changing tools mid-turn broke the prompt cache, which the research note documents.
It is gone for that turn when scored under 0.30. The always-kept workspace primitives plus dependency rules cover the coding path, StructuredOutput is never removed, and the list is never emptied.
Use a TypeSafe key or an OpenRouter key, saved mode 0600 in the Grok home directory or via environment variables. Headless runs never prompt, GROK_TYPESAFE=off disables routing, and the TUI takes keys and toggles with nitro commands.
That was the first version and it lost: re-asking every step changed schemas and broke caching. One Jev call on the first step with the keep-set re-applied keeps schemas identical for the rest of the turn.
On a coding task Jev keeps 8 of Grok Build's 25 tools, which cuts the tool schemas carried in every request from about 11K tokens to about 2.9K. The call is one noul per tool, all in a single request answered in about 350 ms, and it lands only on the first model step of a turn.
A failed call, a 4-second timeout, or a response missing an answer leaves the full tool list in place. Nothing is forced with tool_choice, the list is never emptied, and StructuredOutput is never removed, so the failure mode is full capability rather than a silently narrowed one.
Clone and run ./install.sh with a Rust toolchain, a Grok Build login or XAI_API_KEY, and a TypeSafe or OpenRouter key. The binary lands in ~/.grok/downloads linked from ~/.grok/bin/nitro, leaving the grok command untouched, and install.sh is re-run after a pull. The benchmarks are ./bench.sh grok grok 1 against ./bench.sh nitro nitro 1, summarized with python3 summarize.py.
More Jev resources like this
35GitHub stars
Rust terminal tool that asks Jev noul, choice and score questions about piped text or files and returns calibrated probabilities with stable exit codes for shell scripts, CI jobs and agents.
Jev Agents & MCP#community#jevbest#catalog-import
Community
shaharia-lab/jev-cliRust
shaharia-lab/jev-cliRust25GitHub stars
Agent-oriented playbook for using Jev as a cheap, fast decision layer that triages, classifies, scores, and gates work before expensive browser, research, retry, or subagent steps.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
Bodila51/muse-jev-playbookPython
Bodila51/muse-jev-playbookPython14GitHub 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
6GitHub stars
Terminal coding agent on Pi where Jev shortlists skills, ranks retrieved excerpts, and triages tool failures while the coding model handles reasoning and edits.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
tonyzdev/pijevTypeScript
tonyzdev/pijevTypeScript5GitHub stars
CLI that routes a task to one specialist skill from a Tink library using TypeSafe Jev, verifies the winner with tink mount, then prints it inline or as a mounted path.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
jon-devlapaz/tink-routePython
jon-devlapaz/tink-routePython6GitHub 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
From the Jev community
Posts from builders shipping with Jev right now.