Antigravity Codex MCP
A controlled Model Context Protocol bridge that lets OpenAI Codex delegate work to Google Antigravity CLI on demand, with project-scoped opt-in and read-only delegation.
README
Antigravity Codex MCP
A controlled Model Context Protocol bridge that lets OpenAI Codex delegate work to Google Antigravity CLI on demand.
The bridge stays idle until a user explicitly asks Codex to use AGY. It then authorizes only the exact current project, keeps a project-scoped conversation, synchronizes user-visible dialogue, and confines generated implementation changes to an isolated copy.
This is an independent community project. It is not affiliated with or endorsed by Google or OpenAI.
Highlights
- Explicit opt-in: Codex cannot enable or call AGY unless the user asks for it.
- Exact project scope: broad roots and system directories are rejected.
- Read-only delegation by default: no AGY write, command, URL, or MCP permission is granted.
- Project conversations: active conversation IDs and delegation history survive new Codex tasks.
- Visible transcript sync: messages entered through AGY CLI become available to Codex on the next sync.
- Private-reasoning filter: thinking/reasoning fields, system messages, and checkpoints are excluded before persistence or MCP return.
- Isolated implementation: validated full-file replacements are applied only to a disposable project copy.
- Auditable runs: responses, changed-file manifests, verification output, and sanitized tool events are retained locally.
Requirements
- Node.js 20 or newer
- Antigravity CLI
- OpenAI Codex with MCP support
Authenticate Antigravity CLI and verify that agy is available:
agy --version
agy -p "Reply exactly: AGY_OK" --output-format json
If agy is not on PATH, set AGY_BIN to its executable path in the MCP configuration.
Installation
git clone https://github.com/ustc-fyk/antigravity-codex-mcp.git
cd antigravity-codex-mcp
npm ci
npm test
Add the server to your Codex config.toml. Replace the example paths with the absolute clone path on your machine:
Linux / macOS:
[mcp_servers.antigravity]
command = "node"
args = ["/path/to/antigravity-codex-mcp/src/index.js"]
cwd = "/path/to/antigravity-codex-mcp"
enabled = true
required = false
startup_timeout_sec = 20
tool_timeout_sec = 900
default_tools_approval_mode = "prompt"
enabled_tools = [
"antigravity_health",
"antigravity_project_status",
"antigravity_enable_project",
"antigravity_disable_project",
"antigravity_start_session",
"antigravity_get_active_session",
"antigravity_list_sessions",
"antigravity_sync_conversation",
"antigravity_get_transcript",
"antigravity_ask",
"antigravity_continue",
"antigravity_review",
"antigravity_execute",
"antigravity_list_runs",
"antigravity_get_run",
]
# Optional when agy is not available on PATH:
# [mcp_servers.antigravity.env]
# AGY_BIN = "/path/to/agy"
Windows:
[mcp_servers.antigravity]
command = "node"
args = ['C:\path\to\antigravity-codex-mcp\src\index.js']
cwd = 'C:\path\to\antigravity-codex-mcp'
enabled = true
required = false
startup_timeout_sec = 20
tool_timeout_sec = 900
default_tools_approval_mode = "prompt"
enabled_tools = [
"antigravity_health",
"antigravity_project_status",
"antigravity_enable_project",
"antigravity_disable_project",
"antigravity_start_session",
"antigravity_get_active_session",
"antigravity_list_sessions",
"antigravity_sync_conversation",
"antigravity_get_transcript",
"antigravity_ask",
"antigravity_continue",
"antigravity_review",
"antigravity_execute",
"antigravity_list_runs",
"antigravity_get_run",
]
# Optional when agy is not available on PATH:
# [mcp_servers.antigravity.env]
# AGY_BIN = 'C:\path\to\agy.exe'
Install the included Codex Skill so Codex follows the opt-in workflow:
Linux / macOS (Bash):
mkdir -p ~/.codex/skills/agy-project-assistant
cp ./skills/agy-project-assistant/SKILL.md ~/.codex/skills/agy-project-assistant/
Windows (PowerShell):
$skillRoot = Join-Path $env:USERPROFILE ".codex\skills\agy-project-assistant"
New-Item -ItemType Directory -Force $skillRoot | Out-Null
Copy-Item ".\skills\agy-project-assistant\SKILL.md" $skillRoot
Restart Codex after changing MCP configuration or installing the Skill.
Usage
- Open any project directory in Codex.
- Work normally; the AGY bridge remains idle.
- Explicitly ask:
Load AGY and review this project. - Codex enables read access for that exact project and starts or reuses its AGY conversation.
- Codex delegates analysis, review, continuation, or isolated implementation as appropriate.
- Ask
Disable AGY for this projectto revoke the exact project permission while preserving local audit history.
To inspect or participate in the same conversation directly:
agy --conversation=<conversation_id>
After sending messages through AGY CLI, ask Codex to sync the AGY conversation. antigravity_continue also synchronizes before and after every follow-up. Synchronization is on demand, not a live push channel, so avoid simultaneous sends from Codex and AGY CLI.
MCP tools
| Area | Tools |
|---|---|
| Health and lifecycle | antigravity_health, antigravity_project_status, antigravity_enable_project, antigravity_disable_project |
| Conversations | antigravity_start_session, antigravity_get_active_session, antigravity_list_sessions, antigravity_ask, antigravity_continue, antigravity_review |
| Visible transcripts | antigravity_sync_conversation, antigravity_get_transcript |
| Isolated implementation | antigravity_execute, antigravity_list_runs, antigravity_get_run |
Model calls occur only through the start, ask, continue, review, and execute tools. Health, status, history, run inspection, and transcript synchronization do not consume an AGY model turn.
Project-local state
Each enabled project receives an ignored local state directory:
.antigravity-mcp/
├── project.json
├── sessions.jsonl
├── transcripts/
│ └── <conversation-id>.jsonl
└── runs/
└── <run-id>/
├── metadata.json
├── events.jsonl
├── response.md
└── workspace/
project.jsonrecords opt-in state and the active conversation.sessions.jsonlrecords delegated calls and token usage.transcriptscontains only visible user/assistant records and sanitized tool traces.runscontains isolated implementation workspaces and audit results.
This directory is excluded from the source copy and ignored by Git. Do not commit it manually because it may contain private project or conversation data.
Safety model
Project enablement atomically adds only the exact root to Antigravity trustedWorkspaces and read_file(...) allow rules. Disablement removes only those exact entries.
Isolated implementation rejects:
- absolute paths and parent traversal;
- file deletion;
- secret files, dependencies, build output, metadata, and symbolic links;
- more than 50 changed files;
- more than 2,000,000 replacement characters.
The transcript pipeline recursively removes private reasoning fields and excludes system/checkpoint records before writing the project mirror. See SECURITY.md for reporting guidance and limitations.
Development
npm ci
npm test
npm run smoke:mcp
Live checks require an authenticated AGY CLI and consume model quota:
npm run smoke:live
npm run smoke:execute
npm run smoke:project -- "/absolute/path/to/disposable-project"
npm run smoke:continue -- "/absolute/path/to/enabled-project"
npm run smoke:transcript -- "/absolute/path/to/enabled-project"
npm test is self-contained. smoke:mcp also checks the local AGY binary, while the remaining smoke tests may change project authorization or consume quota.
License
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。