Skip to main content
Glama

Gemini + Jev Bridge for Codex

English | Русский

A local MCP server that lets Codex delegate substantial routine drafting, analysis and code proposals to Gemini. TypeSafe Jev checks the draft against explicit criteria. Codex plans the work, reviews and applies changes, runs tests and gives the final answer.

Gemini uses the included quota of a Google AI Pro subscription through Antigravity CLI. The bridge requires useG1Credits=false: when quota runs out, Gemini work stops until it resets. It does not use a Gemini API key, API billing, credit overages or a paid fallback. Model availability and limits depend on the signed-in Google account and Google's current plan terms.

Jev uses a separate TypeSafe account, API key and quota. It is not included in Google AI Pro. This is an independent project, unaffiliated with Google, OpenAI or TypeSafe.

EN/RU guides

Guide

English

Русский

Installation, usage and troubleshooting

User guide

Инструкция пользователя

Setup prompt and instructions for another GPT/Codex chat

GPT guide

Инструкция для GPT

Context selection, caching and metrics

Performance guide

Быстродействие

Preparing and publishing a clean source export

Publishing guide

Публикация

Related MCP server: codex-antigravity-mcp

Quick start

You need Node.js 22+, npm, Codex with local MCP support, Antigravity CLI signed in to your Google AI Pro account, and a TypeSafe key. These instructions use Windows/PowerShell; live service access on macOS/Linux has not been verified for this release.

In the repository folder:

npm.cmd ci --ignore-scripts
node setup.mjs

Edit the generated settings.json: select an available Gemini Flash model from agy models, set the CLI path and allow only the project folders you need in sourceRoots. Store the TypeSafe key separately in .secrets/TYPESAFE_API_KEY. Then:

npm.cmd test
node smoke.mjs --local-only
node install.mjs

Review the installation plan before applying it:

node install.mjs --apply

The installer registers the MCP server, adds delegation rules and restricts the global Antigravity CLI profile, including manual CLI sessions. It backs up existing files. See the user guide for the exact changes and removal steps, then restart Codex.

MCP tools

Tool

Purpose

delegate_task

Gemini drafts text or a patch; Jev checks 1–6 criteria; results are saved locally

jev_evaluate

Up to 12 compact routing, ranking or evidence questions per request

bridge_status

Local configuration status without contacting the services

Only task text and explicitly selected files are sent to the providers. Credentials, local settings and generated results stay out of Git. Read the data boundaries and delegation rules before selecting files. Jev scores are advisory; they do not replace review or tests. Delegation still uses Codex for task setup and review, so it does not guarantee a fixed quota saving.

Performance

Select inclusive line ranges with source_selections, give Jev a narrow review_context and mapped review_evidence, and keep the default compact response. Exact successful requests are cached for 10 minutes in memory. Identical concurrent Gemini tasks share one execution; distinct tasks return busy. Full drafts/reviews remain in local artifacts, with timing and actual provider-call metrics. See the performance guide.

Verification

npm.cmd test
npm.cmd run check:public

These checks and node smoke.mjs --local-only do not call Gemini or Jev. A live node smoke.mjs call uses their quotas and requires completed setup. Passing offline checks does not confirm account access or remaining quota.

License

MIT. Dependencies retain their own licenses.

Available Tools

3 tools
bridge_statusA
Read-only

Check local Gemini/Jev integration readiness without making paid requests or returning credentials.

Inputs: no named parameters.

delegate_taskA

Send one self-contained routine task to Gemini Flash, then batch-check its result with Jev. Supply selected absolute file paths and explicit acceptance criteria. Returns a short preview, full result file and review verdict. Cannot change source files, browse or execute commands. Uses Google quota and TypeSafe API credits.

Inputs: task (required), review_checks (required), files, context.

jev_evaluateA
Read-only

Ask Jev 1-12 narrow typed questions for routing, ranking or evidence checks on Codex/Gemini results. Supply context as text/JSON plus optional selected files. Confidence is advisory, not proof or authorization. Repeated identical evaluations are cached in this server process. Uses TypeSafe API credits.

Inputs: context (required), questions (required), files.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.8/5.0

    Scored across 3 tools

    Disambiguation4/5

    The three tools have largely distinct roles: bridge_status is a readiness probe, delegate_task runs a Gemini task with a Jev review, and jev_evaluate queries Jev directly. There is mild overlap since both delegate_task and jev_evaluate involve Jev evaluation, but the descriptions make the boundaries (full delegated workflow vs. narrow standalone questions) reasonably clear.

    Naming Consistency3/5

    All names use snake_case, which is good, but the structural pattern is mixed: bridge_status is noun_noun, delegate_task is verb_noun, and jev_evaluate is a namespace-prefixed noun_verb. Readable and not chaotic, but not a predictable verb_noun convention throughout.

    Tool Count4/5

    Three tools is slightly thin but appropriate for a narrow Gemini/Jev bridge whose scope is status checking, task delegation, and evaluation. Each tool earns its place without redundancy.

    Completeness4/5

    The surface covers the core bridge lifecycle: check readiness, delegate a task with review, and run standalone evaluations. Minor gaps exist (e.g., no way to cancel, poll, or manage in-flight delegated tasks), but agents can work around these for the stated purpose.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables Codex users to delegate engineering, refactoring, auditing, and test execution tasks to Google Antigravity CLI as a persistent autonomous subagent, with tools for managing sessions and running prompts.
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables Claude Code to delegate heavy coding and image-generation work to Gemini through the Antigravity CLI using a Google AI Pro subscription instead of an API key, while Claude writes briefs, reviews results, and sends corrections. It runs jobs asynchronously in a guarded work folder, continues the same Gemini conversation for fixes, and reports status, output, and token usage back to the caller.
      MIT