MCP Stockfish vs Godot MCP Runtime

attach_project

Inject the MCP bridge into a Godot process you launch yourself, then wait up to 20s for the bridge to start listening and up to 45s total once it has, so a large project's cold start is absorbed; a port that listens but answers no ping gives up sooner. Call BEFORE Godot launches - Godot reads autoloads only at process start, so a late call returns "bridge did not respond." Recommended pattern: kick off the Godot launch in parallel with this call so the wait absorbs startup. Prefer run_project unless MCP must not spawn Godot. Returns plain-text status with the resolved bridge port. Call detach_project or stop_project when done.

detach_project

Clear attached-mode runtime state and remove the injected McpBridge autoload. Does NOT stop the manually launched Godot process - that stays running. Use after attach_project when you are done driving the game from MCP. For spawned sessions (run_project), use stop_project instead. Mostly optional now: when the bridge disconnects (you closed Godot), the next runtime tool call probes once and ends the attached session itself, removing the autoload. Calling it afterwards still succeeds idempotently, wording the message to distinguish "an attached session existed and already ended" from "this server never attached to a project". Returns: message confirming detach plus externalProcessPreserved (always true here - that is the point of detach vs stop_project). Errors only when a spawned session is what is active; use stop_project for those.

get_debug_output

Get captured stdout/stderr from a spawned Godot project. Use whenever runtime tools fail unexpectedly - script errors, missing nodes, and crash backtraces all surface here. Still works after the process exits or crashes: the session clears itself on exit but the captured logs are retained until stop_project. Requires run_project (not attach_project; attached mode does not capture output). Returns: output/errors (last `limit` lines each, default 200), running (false after exit, null when attached), exitCode after exit, attached:true with empty arrays in attached mode.

stop_project

Stop the spawned Godot project and clean up bridge state. Call when done with runtime testing, even after a crash, and even if you closed the Godot window yourself: it frees the process slot and clears the flag blocking scene-editing tools. A process that exited on its own already removed the bridge autoload at that moment, and this still succeeds - it reports alreadyExited:true with the exit code and the logs captured before the exit, and leaves a finished profiler capture readable. Attached sessions detach without killing the external process. Returns: message, mode, externalProcessPreserved, alreadyExited, exitCode (already-exited case), and condensed finalOutput/finalErrors (capped at 200); get_debug_output has the full log. Errors only when there is no session and no exited process to report.

take_screenshot

Capture a PNG of the running viewport. responseMode: preview (default - saves full PNG, returns bounded inline preview at 960x540), full (full inline PNG; use for small text or pixel-level inspection), path_only (saved-path only, no inline image). Saved under .mcp/godot-runtime/screenshots/ (persists after stop_project). Returns: inline image block (full/preview modes), plus path and size of the saved PNG; previewPath/previewSize in preview mode; warnings for non-fatal runtime errors. Errors if no session or bridge times out (default 10000ms).

simulate_input

Simulate sequential input in a running project and report what each action did. Action `type`: key, mouse_button, mouse_motion, click_element, action, text, wait. For key/mouse_button/action, omit `pressed` to tap (press+release); set it to hold or release. click_element resolves by node path/name (see get_ui_elements), not visible text. Returns: results[] per action with ok, timing, signals fired, the Control hit, UI `changes` (appeared/disappeared/changed), `watch` samples, and `errors` from input handlers (spawned sessions only). Invalid batches inject nothing; a runtime failure stops the batch and skips the rest.

get_ui_elements

Walk the running scene tree and return all Control nodes with positions, sizes, types, and text content. Always call this before simulate_input click_element actions to discover valid element names and paths. Requires an active runtime session (run_project or attach_project). visibleOnly defaults true; pass false to include hidden Controls. filter narrows by class. Returns: elements[] with path/type/rect/visible plus optional text/disabled/tooltip.

run_script

Execute a custom GDScript in the live running project with full scene tree access. Requires an active runtime session. Script must extend RefCounted and define func execute(scene_tree: SceneTree) -> Variant. Return values are JSON-serialized (primitives, Vector2/3, Color, Dictionary, Array, and Node path strings). Use print() for debug output - it appears in get_debug_output, not in the result. In spawned mode, stderr runtime errors escalate to errors (when the script returns null) or surface as warnings. Returns: { success, result, warnings?, tip? } where result is the JSON-serialized return value of execute().

list_autoloads

List all registered autoloads in a project with paths and singleton status. Use first when diagnosing headless failures - broken autoloads crash all headless ops, so this tells you what is loaded. No Godot process required (reads project.godot directly). Returns: [{ name, path, singleton }].

add_autoload

Register a new autoload in a project. autoloadPath accepts "res://..." or a project-relative path (auto-prefixed). singleton defaults true (accessible globally by name). No Godot process required. Warning: autoloads initialize in headless mode - a broken script will crash every subsequent headless op; validate before adding. Returns plain-text confirmation with the registered name, path, and singleton flag. Errors if an autoload with the same name already exists; use update_autoload to modify.