Skip to content

[Bug]: Antigravity ACP does not inherit user MCP/hooks config; Windows hook paths can block tools #37423

Closed as not planned

Description

@felipeassump22

Before submitting

  • I searched existing reports and found related prior work: discussion #13117 and issue #18112 already describe missing MCP servers in Antigravity ACP sessions.
  • This report adds a user-global configuration + lifecycle hooks reproduction and a Windows hook-command path failure, with an observed workaround.
  • I included reproducible steps and environment details.

Area

apps/server — managed Antigravity ACP provider / configuration handling. The Windows hook-command failure may be upstream in the Antigravity runtime.

Summary

T3 Code launches managed Antigravity (agy_acp_server) with an isolated provider home:

%USERPROFILE%\.t3\userdata\providers\antigravity\<id>\
  antigravity-acp\settings.json
  antigravity-cli\skills\
  config\skills\

In the observed T3 Code 0.1.9 installation, this profile's config\ has neither mcp_config.json nor hooks.json, and the Antigravity session does not appear to inherit the user's existing %USERPROFILE%\.gemini\config\ configuration.

The observations are:

  1. User MCP configuration is missing. The session exposes only T3 Code's t3-code MCP server, although additional servers work in standalone Antigravity.
  2. User lifecycle hooks are missing. Hooks configured for standalone Antigravity are not executed in T3 Code sessions.
  3. Windows hook-command paths may fail after manually copying the config. With the observed hook command containing Windows backslash paths, tool calls are rejected with an external hook execution error. Replacing path separators in the command with forward slashes made the hook work in the tested environment.

These are related but potentially separate faults: configuration discovery / isolation versus hook command execution. The third one has not been conclusively attributed to T3 Code rather than the managed Antigravity runtime.

Related reports (not claiming the MCP gap is new)

Steps to reproduce

A. User-global MCP configuration

  1. Register a working stdio MCP server in %USERPROFILE%\.gemini\config\mcp_config.json. Example configuration, with neutral placeholders (not a verbatim reproduction log):

    {
      "mcpServers": {
        "example-server": {
          "command": "node",
          "args": ["C:/path/to/server.js"]
        }
      }
    }
  2. Confirm the server's tools are available in standalone Antigravity.

  3. Start a new Antigravity thread in T3 Code and inspect available MCP servers / attempt to invoke a tool from the registered server.

  4. Observed: only t3-code is available; the user-configured server is not.

  5. Copy the working mcp_config.json to %USERPROFILE%\.t3\userdata\providers\antigravity\<id>\config\mcp_config.json and start a new Antigravity thread.

  6. Observed: the user-configured server's tools become available and calls succeed.

B. Lifecycle hooks and Windows command paths

  1. Start with a working %USERPROFILE%\.gemini\config\hooks.json that configures a command-based PreToolUse hook. The affected command used ordinary Windows executable and argument paths of the form C:\Users\<user>\...\hook.exe (details omitted to keep this report independent of any particular third-party integration).
  2. Start an Antigravity thread in T3 Code and invoke a tool.
  3. Observed: the user-global hook does not run.
  4. Copy hooks.json into the isolated provider's config\ directory and invoke a tool again.
  5. Observed: tool execution is rejected with denied by pre-tool hook: External hook '<hook-name>' failed to execute. (hook name redacted).
  6. Run that same hook command manually with the same stdin payload.
  7. Observed: it completes with exit code 0 and returns {"decision":"allow"}.
  8. Replace backslash path separators inside the hook command with forward slashes (C:/Users/.../hook.exe, including relevant arguments), then retry from a new Antigravity thread.
  9. Observed: tool execution succeeds and the hook runs.

Expected behavior

  • Clearly document whether managed Antigravity profiles are intentionally isolated from the user's global Antigravity MCP/hook configuration.
  • Either make user MCP servers and hooks available in T3 Code Antigravity sessions (subject to safe credential handling), or expose a supported, documented configuration mechanism.
  • When a hook cannot start, provide an actionable execution error instead of an opaque denial of tool calls.
  • Investigate whether Windows backslash path handling belongs in T3 Code or the managed Antigravity runtime, and redirect that part to the appropriate tracker if necessary.

Actual behavior

  • MCP servers configured in the user's global Antigravity config are absent until manually copied into the private provider profile.
  • Global lifecycle hooks are not run in the managed Antigravity session.
  • A copied hook command using Windows backslash paths fails to execute and blocks tool calls; equivalent forward-slash paths work in this tested setup.

Impact

User-configured tooling and lifecycle automation that work in standalone Antigravity are silently unavailable through T3 Code. Manually adding a hook can instead prevent the provider from using tools, with an error that does not explain why the hook failed.

Version or commit

  • T3 Code desktop 0.1.9
  • Managed Antigravity ACP runtime in .t3\tools\antigravity-acp\win32-x64
  • Runtime release ID: d24f4c73c9f96d4f44384e281c855a56c72d8eaa9c5b0fad7ccb1635c6b63f50

Environment

  • Windows 11 (x64)
  • T3 Code desktop with managed Antigravity ACP provider
  • User MCP server configuration using stdio / Node; command-based lifecycle hook
  • Other T3 providers were available for comparison

Logs (identifiers redacted)

# Before seeding the isolated MCP config
servers available: t3-code
invalid tool call error (invalid_args) tool <tool-name> is not enabled for server <server-name>

# After copying the hooks config with Windows backslash paths
denied by pre-tool hook: External hook '<hook-name>' failed to execute.

The latest file found in ~/.gemini/antigravity/log predated these runs. If the managed ACP runtime writes to a different log location, please advise where to collect diagnostics.

Workaround

  1. Place the desired MCP definitions in %USERPROFILE%\.t3\userdata\providers\antigravity\<id>\config\mcp_config.json. Avoid copying credentials or unrelated servers into a profile that does not need them.
  2. For the tested command-based hook, place its config in the same profile and use forward slashes in executable and argument paths.
  3. Start a new Antigravity thread and verify the servers and hooks.

The <id> profile directory is generated by T3 Code; the manual workaround is therefore fragile if the profile changes.

Activity

  1. juliusmarminge commented on Nov 16, 2026

    @juliusmarminge
    Member

    Thanks for writing this up. It asks for new behaviour rather than reporting a defect, so it does not belong in the bug tracker and I'm closing it as not planned here. Feature ideas are discussed in Discussions: please post it there.

  2. changed the title [-][Bug]: Antigravity ignores user MCP servers and hooks (isolated provider home); backslash hook paths block all tools on Windows[/-] [+][Bug]: Antigravity ACP does not inherit user MCP/hooks config; Windows hook paths can block tools[/+] on Nov 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions