gms-mcp 0.1.83

GameMaker CLI + MCP server toolset

GameMaker MCP Tools

Project Features

• gms: a Python CLI for GameMaker project operations (asset creation, maintenance, runner, etc).
• gms-mcp: an MCP server that exposes the same operations as MCP tools (Cursor is the primary example client).
• TCP Bridge (optional): live, bidirectional game communication (commands + log capture) via gm_bridge_install, gm_run_command, and gm_run_logs. See documentation/BRIDGE.md.
• Reliability-First Architecture: Custom exception hierarchy, typed result objects, and an execution policy manager replace monolithic exit calls and raw dictionaries. This enables structured error handling, consistent tool integration, and optimized performance (Fast assets, Resilient runner).
• Health & Diagnostics: gm_mcp_health provides a one-click diagnostic tool to verify the local GameMaker environment. gm_diagnostics provides structured, machine-readable project diagnostics (JSON, naming, orphans, references) compatible with IDE problem panels.
• Runtime Management: gm_runtime_list, gm_runtime_pin, and gm_runtime_verify allow precise control over which GameMaker runtime version is used for builds and execution.
• Cross-Platform Runner Defaults: gm_run / gm_compile now default to the host OS target platform (macOS, Linux, or Windows) when not explicitly provided.
• macOS Local Runner Behavior: local gm_run / gm_compile use Igor's run-based path for IDE-equivalent validation without Developer ID packaging, and macOS background run sessions now track and stop the real Mac_Runner process cleanly. Packaged temp-output runs still resolve .app bundles via Contents/MacOS/ when PackageZip is used.
• GML Symbol Indexing & Code Intelligence: gm_build_index, gm_find_definition, gm_find_references, and gm_list_symbols provide deep, fast, and filtered code analysis (definitions and cross-file references).
• Introspection: complete project inspection with support for all asset types (including extensions and datafiles).
• MCP Resources: addressable project index and asset graph for high-performance agent context loading.
• gms-mcp-init: generates shareable MCP config files for a workspace. Now auto-detects environment variables like GMS_MCP_GMS_PATH to include in the generated config.

Install (recommended: pipx)

pipx install gms-mcp

PowerShell equivalent:

pipx install gms-mcp

Claude Code Plugin

For Claude Code users, install the plugin for the best experience:

/install-plugin github:Ampersand-Game-Studios/gms-mcp

This provides:

• Skills: 18 workflow guides + 7 reference docs
• Hooks: Automatic update checks and error notifications
• MCP Server: Auto-configured via uvx (no pip install needed)

For Other Tools (Cursor, VSCode, OpenClaw, etc.)

pip install gms-mcp
gms-mcp-init --cursor  # or --vscode, --windsurf, --openclaw, etc.

For skill packs, OpenClaw users can install to user or workspace scope:

gms skills install --openclaw            # user scope: ~/.openclaw/skills/
gms skills install --openclaw --project  # workspace scope: ./skills/

Note: .openclaw/openclaw.json is for settings. Workspace skills are loaded from ./skills/.

For Codex

gms-mcp-init --codex

This writes a workspace .codex/mcp.toml file and prints the codex mcp add registration command.

Global config mode writes directly to ~/.codex/config.toml (merging server entries).

Use the printed command directly, or copy .codex/mcp.toml content into the [mcp_servers] section of your ~/.codex/config.toml.

Codex helpers:

• gms-mcp-init --codex-check prints detected Codex config paths and active server entry.
• gms-mcp-init --codex-check-json prints the same check output in machine-readable JSON.
• gms-mcp-init --codex-dry-run-only prints final merged payloads for workspace + global Codex config without writing files.
• gms-mcp-init --codex-app-setup runs one-shot Codex app setup: writes workspace config, previews global merge, then prints check + readiness summary.

Local Development Setup

If you are working on the gms-mcp codebase itself, follow these steps to set up a local development environment:

1. Clone and install in editable mode:

   git checkout dev
   python3.12 -m venv .venv
   source .venv/bin/activate
   python3.12 -m pip install -e ".[dev]"
   

   gms-mcp requires Python 3.10+; we recommend Python 3.12 for local development.

2. Run the full local test suite:

   PYTHONPATH=src python3.12 cli/tests/python/run_all_tests.py
   

3. Initialize local and global MCP servers for testing: We recommend setting up two separate MCP server configurations in Cursor to test your changes:

   • Global (gms-global): For general use across all your GameMaker projects.
   • Local (gms-local): Specifically for testing your current changes to the server.

   Run these commands from the project root (zsh/bash):

   # Global setup (names it 'gms-global' in Cursor)
   gms-mcp-init --cursor-global --server-name gms-global --mode python-module --python python3 --non-interactive
   
   # Local setup (names it 'gms-local' in Cursor)
   gms-mcp-init --cursor --server-name gms-local --mode python-module --python python3 --non-interactive
   

   PowerShell equivalent:

   # Global setup (names it 'gms-global' in Cursor)
   gms-mcp-init --cursor-global --server-name gms-global --mode python-module --python python --non-interactive
   
   # Local setup (names it 'gms-local' in Cursor)
   gms-mcp-init --cursor --server-name gms-local --mode python-module --python python --non-interactive
   

4. Verify in Cursor: Go to Cursor Settings > Features > MCP to see your new servers. You may need to click "Reload" or restart Cursor to see changes.

Publishing (maintainers)

Publishing is automated via GitHub Actions (PyPI Trusted Publishing) on every push to main and on tags v*. See RELEASING.md for the one-time PyPI setup and the first manual upload helper scripts.

CI Coverage

• Core CI runs on Ubuntu and Windows across Python 3.11-3.13.
• Runner/session regression tests also run on macOS across Python 3.11-3.13, including a mockless smoke test that builds a real .app bundle structure and validates executable path resolution.

Quality Reports

Quality reports are generated during CI and published as quality-reports-* artifacts.

• TEST_COVERAGE_REPORT.md
• MCP_TOOL_VALIDATION_REPORT.md
• coverage.xml
• pytest_results.xml
• quality_summary.json

You can regenerate these locally with:

python scripts/generate_quality_reports.py

Use --skip-test-run to regenerate from existing CI artifacts:

python scripts/generate_quality_reports.py --skip-test-run --junit-xml build/reports/pytest_results.xml --coverage-xml build/reports/coverage.xml

X (Twitter) posting on main

This repo can post to X automatically when main is updated.

• Personality / voice: .github/x-personality.md
• Tweet staging file: .github/next_tweet.txt

How it works

• When a commit lands on main, GitHub Actions reads .github/next_tweet.txt.
• If it contains the placeholder text (or is empty), it skips posting.
• If it contains a real tweet, it posts to X, records the post in cached tweet history, and leaves the staged text in git so later runs can safely dedupe it.
• Transient X API failures are retried automatically inside the posting script with bounded backoff so 503 Retry-After hints cannot exceed the workflow budget, and edge challenge pages are retried without misreporting them as credential failures.
• If X still fails with a known transient error after retries, the workflow now leaves the staged tweet or evergreen queue item in place for a later retry instead of marking the whole job failed.
• The X workflows also enforce job/request timeouts so a hung API call cannot block the queue indefinitely.
• You can also retry the release-post workflow manually with the Post to X workflow dispatch in GitHub Actions.

Maintainer flow (dev -> pre-release -> main)

Because this repo promotes changes dev -> pre-release -> main, prepare the tweet during the pre-release -> main PR:

• Update .github/next_tweet.txt with the tweet (following .github/x-personality.md)
• Merge to main

Use with a GameMaker project (multi-project friendly)

Run this inside each GameMaker project workspace (or repo) to generate config:

gms-mcp-init --cursor

This writes .cursor/mcp.json and attempts to auto-detect the .yyp location to set GM_PROJECT_ROOT.

For a one-time setup that works across many projects, write Cursor's global config instead:

gms-mcp-init --cursor-global

Generate a Codex config from the current workspace:

gms-mcp-init --codex

Generate a global Codex entry in ~/.codex/config.toml:

gms-mcp-init --codex-global

Global mode merges with existing entries so it is safe to keep multiple MCP servers in the same file.

Inspect current Codex config resolution:

gms-mcp-init --codex-check

Preview final merged Codex payloads for local + global without writing:

gms-mcp-init --codex-dry-run-only

Print Codex check output as JSON (useful for app automation):

gms-mcp-init --codex-check-json

One-shot Codex app setup (recommended for new workspaces):

gms-mcp-init --codex-app-setup

Codex App Quickstart

1. Run gms-mcp-init --codex-app-setup in your GameMaker workspace.
2. Confirm the output says Ready for Codex app: yes.
3. If needed, run gms-mcp-init --codex-check-json and verify active.scope is workspace.
4. Use gms-mcp-init --codex-dry-run-only before changing global config to preview merged TOML safely.

Canonical Client Workflow

All clients now support the same canonical action surface:

gms-mcp-init \
  --client <cursor|codex|claude-code|claude-desktop|antigravity|gemini|vscode|windsurf|openclaw|generic> \
  --scope <workspace|global> \
  --action <setup|check|check-json|app-setup>

Optional:

• --config-path /custom/path to override default config location
• --safe-profile to enforce conservative env defaults

Examples:

# Cursor setup + readiness check
gms-mcp-init --client cursor --scope workspace --action app-setup

# Codex machine-readable readiness
gms-mcp-init --client codex --scope workspace --action check-json

# Claude Desktop global plugin sync
gms-mcp-init --client claude-desktop --scope global --action setup

# Gemini alias (Antigravity path)
gms-mcp-init --client gemini --scope global --action app-setup

# OpenClaw app setup + workspace skills install
gms-mcp-init --client openclaw --scope workspace --action app-setup \
  --openclaw-install-skills --openclaw-skills-project

For parity status and supported defaults, see documentation/CLIENT_SUPPORT_MATRIX.md.

Generate example configs for other MCP-capable clients:

gms-mcp-init --vscode --windsurf --antigravity --openclaw

Set up Antigravity global config (recommended):

gms-mcp-init --antigravity-setup

This merges into ~/.gemini/antigravity/mcp_config.json, writes atomically, creates a timestamped backup on overwrite, and enables a conservative safety profile by default:

• GMS_MCP_ENABLE_DIRECT=0
• GMS_MCP_REQUIRE_DRY_RUN=1

Check Antigravity readiness:

gms-mcp-init --antigravity-check

Print Antigravity check output as JSON:

gms-mcp-init --antigravity-check-json

One-shot Antigravity app setup:

gms-mcp-init --antigravity-app-setup

Use a custom Antigravity config path:

gms-mcp-init --antigravity-setup --antigravity-config-path /path/to/mcp_config.json

Opt in to the conservative safety profile for Antigravity example configs too:

gms-mcp-init --antigravity --safe-profile

When GMS_MCP_REQUIRE_DRY_RUN=1 is set, you can allow specific destructive tools with:

export GMS_MCP_REQUIRE_DRY_RUN_ALLOWLIST=gm_asset_delete,gm_workflow_delete

Or generate everything at once:

gms-mcp-init --all

Monorepos / multiple .yyp

If multiple .yyp projects are detected in a workspace:

• gms-mcp-init will warn and (when interactive) prompt you to pick one.
• In non-interactive environments, it defaults GM_PROJECT_ROOT to ${workspaceFolder} (safe).

Force a specific project root:

gms-mcp-init --cursor --gm-project-root path/to/project

Preview output without writing files:

gms-mcp-init --cursor --dry-run

Code Intelligence & Introspection

The MCP server provides comprehensive project analysis capabilities:

GML Symbol Indexing (gm_build_index)

Build a high-performance index of all functions, enums, macros, and global variables in the project. This is required for advanced code intelligence tools.

Symbol Definition (gm_find_definition)

Find the exact location and docstrings for any GML symbol in your project.

Find References (gm_find_references)

Search for all usages of a specific function or variable across your entire codebase.

List Symbols (gm_list_symbols)

List all project symbols with filtering by type, name substring, or file path.

Asset Listing (gm_list_assets)

List all assets in your project, optionally filtered by type:

• Supported types: script, object, sprite, room, sound, font, shader, path, timeline, tileset, animcurve, sequence, note, folder, extension, includedfile (datafiles)

Asset Reading (gm_read_asset)

Read the complete .yy JSON metadata for any asset by name or path.

Reference Search (gm_search_references)

Search for patterns across project files with:

• Scopes: all, gml, yy, scripts, objects, extensions, datafiles
• Modes: literal string or regex
• Options: case sensitivity, max results

Asset Graph (gm_get_asset_graph)

Build a dependency graph of assets with two modes:

• Shallow (fast): Parses .yy files for structural references (parent objects, sprites, etc.)
• Deep (complete): Also scans all GML code for runtime references like instance_create, sprite_index, audio_play_sound, etc.

Texture Groups (gm_texture_group_*)

Create, inspect, and edit .yyp TextureGroups, plus bulk-assign assets (sprites/fonts/tilesets/etc) via textureGroupId.

Read-only tools:

• gm_texture_group_list: list texture groups + available configs (desktop/android/ios/etc)
• gm_texture_group_read: read a single texture group entry
• gm_texture_group_members: list assets in a group (top-level + ConfigValues overrides)
• gm_texture_group_scan: report missing groups referenced + mismatches (top-level vs config override)

Destructive tools (all support dry_run=true):

• gm_texture_group_create: clone an existing template group (default: Default)
• gm_texture_group_update: patch fields on a group (optionally per config via ConfigValues)
• gm_texture_group_rename: rename a group and rewrite asset references
• gm_texture_group_delete: blocks by default if referenced unless reassign_to is provided
• gm_texture_group_assign: bulk-assign assets by explicit list or filters

Config scope defaults:

• Assignment updates an asset's top-level textureGroupId only when it is a dict (null is left as-is).
• If configs is omitted, assignment updates only existing ConfigValues entries; pass configs=[...] to create explicit overrides.

MCP Resources

Pre-built, cacheable project data for agents:

• gms://project/index: Complete project structure (assets, folders, room order, configs, audio/texture groups, IDE version)
• gms://project/asset-graph: Asset dependency graph
• gms://system/updates: Returns a human-readable message if a newer version of gms-mcp is available on PyPI or GitHub.

Update Notifier

The server automatically checks for updates on startup and during common operations:

• Tool: gm_check_updates returns structured update info.
• Auto-check: gm_project_info includes an updates field.
• Resource: gms://system/updates provides a quick text status.

CLI usage

Run from a project directory (or pass --project-root):

gms --version
gms --project-root . asset create script my_function --parent-path "folders/Scripts.yy"
gms --project-root . texture-groups list
gms --project-root . texture-groups assign game --type sprite --folder-prefix sprites/ --dry-run