GODKILLER MCP 是一个面向 Google Antigravity 的 MCP 服务器,用“计划—门禁—编辑—磁盘验证—完成声明”的阶段内核,阻止没有证据就宣称完成。

Stars

7

7 天增长

暂无数据

Fork 数

2

开放 Issue

0

开源协议

MIT

最近更新

2026-07-31

AI 仓库情报摘要
FR-AI / ANALYSIS

为什么值得关注

它把 AI 编码代理的“完成”从口头承诺变成必须在磁盘上留下证据的受控流程,并提供了 A/B 竞技场协议和引擎关卡来衡量有/无该内核时的代理表现。

适合谁使用

  • 使用 Google Antigravity 的开发者
  • 依赖 MCP 工具链的 AI 编码工作流使用者
  • 关注 AI 代理可靠性与可验证性的工程团队
  • 做 LLM 编码代理评测的研究者

典型使用场景

  • 在 Antigravity 中强制 plan→edit→verify→claim_done 的阶段纪律
  • 运行 A/B 协议比较有/无 GODKILLER MCP 的代理修复效果
  • 用 /ultradeep 对单个文件进行单步、可控的深度修改
  • 用 verify bundle、fault probe 和 council 评审提高输出可信度

项目优势

  • 阶段机、plan lock 和 edit-safe 限制未经验证的修改
  • verify 环节要求磁盘证据(PASSING_TEST/BLAST_RADIUS/EDIT_SAFE),并能用 fault_probe 杀掉薄弱测试
  • 内置 11 维 A/B 评测框架与引擎关卡,报告 331 个包测试通过
  • 安全设计包含命令白名单、secret 不返回、严格模式可选

使用前须知

  • 主要面向 Google Antigravity,对其他代理的适用性未在 README 中展示
  • A/B 竞技场需要手动重置和分步评分,不是一条命令自动完成
  • 严格模式、浏览器/抓取/OCR 等属于可选或实验性功能,需要额外安装或显式开启

README 快速开始

GODKILLER MCP

Phase kernel for Google Antigravity.
Plan. Gate. Prove on disk. Then claim done — or get blocked.

pip install godkiller-mcp

Contact: Facebook · Instagram @Kayvin.th


Mission

Antigravity skips phases. GODKILLER does not allow that.

Failure modeKernel response
No /planplan / phase gates
Multi-file rush/ultradeep one-file think → plan → edit
Fake “done”verify_bundle + claim_done block (server-authored evidence only)
Session amnesiamarathon + memory graph
Weak reviewLLM council — Coder / Hacker / Optimizer, multi-round debate
goal → mode → evidence/plan → gated edit → disk verify → claim_done

Arena — Antigravity A/B (the real proof)

Pre-solved twins that both pass 516 prove nothing. The honest protocol:

  1. Reset arms to sealed buggy 1_ORIGINAL
  2. You open Antigravity on 3_WITHOUT_MCP with MCP off → agent tries → we score
  3. Reset / open 2_WITH_MCP with MCP on → agent tries → we score
  4. Compare 11 dimensions (fail-closed on missing evidence)
ControlValue
Arena rootGODKILLER_ISOLATED_ARENA / GODKILLER_ARENA_ROOT
Baseline1_ORIGINAL (never edit mid-run)
Bare3_WITHOUT_MCP
GODKILLER2_WITH_MCP
Oraclesealed hidden_oracle/
Protocolbenchmarks/ANTIGRAVITY_AB_PROTOCOL.md

11 dimensions

#Dimension
1Code correctness (oracle pass %)
2Oracle volume (≥516 collected)
3Output integrity (full pytest body)
4Delta from baseline (files actually changed)
5Reconnaissance / exhaustive read artifacts
6Phase discipline (task / plan / marathon)
7Blast radius + edit-safe
8Verify bundle + claim_done
9Council (Coder / Hacker / Optimizer)
10Security hardening signals
11UI visual gate (critic / screenshots)
$env:GODKILLER_ARENA_ROOT="C:\Users\ASUS\Desktop\GODKILLER_ISOLATED_ARENA"
python -m benchmarks.reset_arena
# → open Antigravity bare, then:
python -m benchmarks.score_11 --arm 3_WITHOUT_MCP
# → open Antigravity +

项目描述

Phase kernel for Google Antigravity — plan, gate, verify on disk, then claim_done.

相关仓库与替代方案

根据分类、Topic 和编程语言匹配的相似项目。

0xwilliamortiz
精选
0xwilliamortiz GitHub avatar

openclaude-improved

OpenClaude is an open-source CLI coding agent that runs on any platform and supports a wide range of LLM providers, offering the same tools and workflows as Claude Code.

AI 与机器学习大语言模型
577
slvDev
精选
slvDev GitHub avatar

esp32-ai

A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.

AI 与机器学习大语言模型
1,960
gavamedia
精选
gavamedia GitHub avatar

deltafin

Deltafin is a research project that runs the 2.8-trillion-parameter Mixture-of-Experts model Kimi K3 on a single Apple Silicon Mac (e.g., M1 Max with 64 GB) at about 16 seconds per token, using exact, reproducible inference with local or streaming expert loading.

AI 与机器学习大语言模型
304