Antigravity Delegate MCP
A local-STDIO MCP server that lets Codex invoke the Antigravity CLI from any project, with project-owned routing guidance and enforceable policies.
README
Antigravity Delegate MCP
A distributable local-STDIO MCP server that lets Codex invoke Antigravity CLI from any project. The global server is policy-neutral; each project owns its routing guidance and enforceable policy.
Version 0.4 adds a read-only first-run readiness tool. A new or renamed project can expose every missing prerequisite—CLI, authentication, model mapping, project policy, routing guidance, file permissions, and stale historical paths—without silently changing global or project configuration.
Version 0.4 also packages the MCP and gemini-delegation-router Skill as one Codex Plugin. After installation, Codex can implicitly route suitable subtasks when a user says phrases such as “过程中可以调用 Gemini MCP”, while project policy and Codex review remain in control.
Install the Codex Plugin from GitHub
After this version is merged to the repository's default branch:
codex plugin marketplace add silennsong/Antigravity-MCP-Tool
codex plugin add antigravity-mcp@antigravity-tools
Start a new Codex task after installation so the bundled Skill and MCP tools are loaded. A natural-language trigger is enough:
完成这个任务,过程中可以调用 Gemini MCP。
Explicit invocation is the most deterministic option:
使用 $gemini-delegation-router 完成这个任务。
The Plugin bundles its dependency-free Python MCP runtime, so its server does not depend on the repository checkout path or a globally installed Python package. It still needs Python 3.10+ and the Antigravity CLI. On first use, check_antigravity_readiness exposes copy-paste commands for any missing CLI installation, OAuth, model mapping, project policy, or scoped workspace permission. Interactive Google consent and project trust remain user-controlled.
For local development of this branch:
codex plugin marketplace add /absolute/path/to/Antigravity-MCP-Tool
codex plugin add antigravity-mcp@antigravity-tools
The marketplace definition is stored in .agents/plugins/marketplace.json; the installable package is under plugins/antigravity-mcp.
Install from a release package
Clone the public repository:
git clone --depth 1 https://github.com/silennsong/Antigravity-MCP-Tool.git
cd Antigravity-MCP-Tool
python3 bootstrap.py --install-agy
Or download and extract the source archive from the latest GitHub Release, inspect bootstrap.py, then run:
python3 bootstrap.py --install-agy
The bootstrapper:
- Creates an isolated venv under
~/.local/share/antigravity-delegate-mcp/venv. - Installs this package without modifying the system Python environment.
- Optionally installs
agyfrom Google's official installer. - Registers
antigravity_delegateglobally throughcodex mcp add. - Preserves unrelated Codex configuration and adds MCP timeout/tool settings.
- Runs
doctor.
For interactive onboarding as part of setup:
python3 bootstrap.py --install-agy --auth --configure-models
Google OAuth, terms, data-use choices, and workspace trust require the user to interact with Antigravity. The installer does not silently accept them.
See AGENT_INSTALL.md for a prompt that can be copied to a local coding agent.
Project homepage: https://github.com/silennsong/Antigravity-MCP-Tool
First project connection
Restart Codex after installation, open the target project, and ask the Agent:
Call the MCP tool check_antigravity_readiness for the current workspace before any delegation.
Report every failed or missing check and the exact suggested commands.
Do not change global permissions or project files unless I explicitly approve the relevant command.
The readiness tool is read-only. It reports two independent states:
ready_to_delegate: Antigravity CLI, authentication, and the requested model tier work.safe_project_ready: the project also has policy, AGENTS routing, and a read/deny-write permission pair.
For a complete explicit first-run setup from the terminal:
antigravity-delegate-mcp setup \
--workspace /path/to/project \
--auth \
--configure-models \
--init-project \
--profile read-only \
--deep
If the project already has AGENTS.md, inspect it first. Add --force only when you want the installer to append its clearly marked Antigravity routing block; existing content is preserved.
Commands
antigravity-delegate-mcp install --replace
antigravity-delegate-mcp auth --workspace /path/to/project
antigravity-delegate-mcp configure-models
antigravity-delegate-mcp init-project --workspace /path/to/project --profile read-only
antigravity-delegate-mcp validate-policy --workspace /path/to/project
antigravity-delegate-mcp permissions audit --workspace /path/to/project
antigravity-delegate-mcp permissions sync --workspace /path/to/project
antigravity-delegate-mcp permissions prune --stale
antigravity-delegate-mcp doctor --workspace /path/to/project
antigravity-delegate-mcp doctor --deep --workspace /path/to/project
init-project offers two profiles:
read-only: generates routing rules, a restrictive project policy, a local JSON Schema, and maps the mode toagy --mode plan --sandbox --disable-slash-commands.open: creates a minimal project policy without global task/model/mode restrictions.
Existing AGENTS.md files are not overwritten. Use --force to append a marked Antigravity section after reviewing the file.
The read-only profile also adds scoped Antigravity CLI grants in ~/.gemini/antigravity-cli/settings.json: recursive read_file(<project>) is allowed and write_file(<project>) is denied. Use --no-agy-permissions to skip that explicit initialization change.
Absolute-path permissions are global Antigravity CLI state. Renaming or moving a project can leave a stale historical rule. The lifecycle is deliberately explicit:
permissions auditreads current coverage and stale paths without changing anything.permissions syncadds only the supplied workspace's read/deny-write pair.permissions prune --staleis preview-only.- Rerun the prune command with
--yesonly after reviewing every listed rule.
Delegation never edits this permission file automatically. Permission or authentication failures return structured onboarding_required, code, and actions fields so an Agent can explain exactly what is missing.
Responsibility split
Global layer:
- Registers
delegate_to_antigravityplus the read-onlycheck_antigravity_readinesstool. - Resolves stored or environment-provided model names.
- Adapts to the verified
agyprint-mode flags. - Applies process timeout/output handling and returns structured metadata.
- Does not globally hard-code task categories, model routing, or read/write behavior.
Project layer:
AGENTS.mdtells Codex when delegation is useful..codex/antigravity-policy.jsondefines enforceable task, model, mode, prompt, CLI-argument, output, and environment rules..codex/antigravity-policy.schema.jsongives editors a local policy schema.- The closest policy found while walking upward from
workspaceis used.
Without a project policy, the global MCP applies no semantic task restriction.
Readiness may therefore report ready_to_delegate: true and safe_project_ready: false. This distinction preserves a policy-neutral global entrypoint while making incomplete project onboarding visible.
Authentication and model configuration
The official Antigravity CLI installer for macOS/Linux is:
curl -fsSL https://antigravity.google/cli/install.sh | bash
This project downloads the installer before executing it instead of piping it blindly. To complete Google OAuth and project trust:
antigravity-delegate-mcp auth --workspace /path/to/project
Then configure exact model IDs returned by agy models:
antigravity-delegate-mcp configure-models \
--flash gemini-3.6-flash-low \
--pro gemini-3.1-pro-high \
--non-interactive
Mappings are stored with mode 0600 in:
~/.config/antigravity-delegate/config.json
Environment variables such as ANTIGRAVITY_FLASH_MODEL override stored mappings. Exact model input is also supported when project policy permits it.
Headless permissions
Antigravity print mode cannot display interactive tool-approval cards. A project that needs file access must first complete workspace trust or configure appropriately scoped Antigravity permission grants. init-project --profile read-only handles the scoped read/deny-write grant automatically. Do not solve this with --dangerously-skip-permissions unless the project explicitly chooses that risk.
The MCP treats exit code 0 with empty stdout as an error and returns Antigravity's permission explanation instead of reporting false success.
It classifies common missing prerequisites for Agents:
antigravity_cli_missingmodel_mapping_missingauthentication_requiredworkspace_read_permission_missing
Each classified failure includes copy-paste remediation commands and confirms that no configuration change was made.
Tool input
{
"task": "Inspect the repository entry points and map the main modules.",
"task_kind": "repository_exploration",
"workspace": "/absolute/path/to/current/project",
"model_tier": "flash",
"mode": "read_only",
"max_output_chars": 12000,
"timeout_seconds": 600
}
Globally, task_kind, model_tier, and mode accept project-defined strings. Project policy may restrict them. Passing model uses an exact Antigravity model name instead of tier lookup.
Project policy fields
versionallowed_task_kindsallowed_model_tiersallow_explicit_modelallowed_modesmodel_tier_task_kindsmode_cli_argsmin_task_chars,max_task_charsmax_output_chars,max_timeout_secondsforbidden_task_patternsworker_instructionsrequired_output_sectionsforward_env
Use validate-policy or doctor before relying on a new policy.
Each forbidden-pattern entry may set ignore_negated: true when phrases such as “do not delete files” should not be treated as a request to perform the forbidden action.
Build distributable artifacts
python3 scripts/build_plugin_bundle.py
python3 -m build
Outputs:
dist/antigravity_delegate_mcp-<version>-py3-none-any.whl
dist/antigravity_delegate_mcp-<version>.tar.gz
Verification
python3 -m unittest discover -s tests -v
antigravity-delegate-mcp doctor --deep --workspace /path/to/project
codex mcp get antigravity_delegate
Runtime Python dependencies are intentionally empty. Python 3.10 or newer is required.
Plugin-specific validation:
python3 scripts/build_plugin_bundle.py
python3 ~/.codex/skills/.system/skill-creator/scripts/quick_validate.py \
plugins/antigravity-mcp/skills/gemini-delegation-router
python3 ~/.codex/skills/.system/plugin-creator/scripts/validate_plugin.py \
plugins/antigravity-mcp
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
Neon Database
MCP server for interacting with Neon Management API and databases
E2B
Using MCP to run code via e2b.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.