AI and MCP

How the integration works

AI / MCP is a native Reckfell editor module. It exposes project tools to the built-in assistant and external MCP clients. A model discovers actions and invokes editor operations for scenes, resources, C# and the running game.

The model interprets a request and selects actions. MCP delivers those actions to the editor and returns results. Changing providers does not change the project format or the tools available in your editor build.

Editor requirements

Open a project in Reckfell. The workspace includes the AI / MCP dock. The module is included in the editor distribution. Configure a model connection in the dock before using the assistant.

Available actions depend on the modules in your build. Use the live catalog as the capability reference. A generic method invocation route does not mean every possible workflow has passed practical verification.

Connect a model

Configure the built-in assistant with these parameters:

External provider. Use your provider's compatible API and a model with tool-call support. The engine does not silently replace the chosen model or fall back to another provider.

Local model. Point the assistant at a server running on your computer or local network. For example, the standard local Ollama endpoint is:

http://localhost:11434/v1/chat/completions

Use the model identifier from your local setup. Tool support and task quality depend on the selected model.

Reckfell provider. Enter the endpoint, model identifier and key supplied with your access. This guide does not assign an unannounced public service URL or pricing plan.

The endpoint and model selection persist in local project settings. On Windows, keys are stored in Credential Manager rather than chat history.

Enable project editing

Select Allow AI to edit this project in the AI / MCP dock when the assistant needs to make changes. Without this permission, read operations remain available while writes return project_not_trusted.

Describe the expected result. For example: “Add a third-person camera to Player, reuse the existing InputMap, build the C# project and report errors.” The assistant and external clients use the same tool dispatcher.

Connect an external MCP client

1. Open your project and the AI / MCP dock.
2. Select Start server.
3. Select Copy connection and configure your client with the endpoint and Authorization header.
4. Begin with reckfell_catalog to discover tools.

The server listens on the loopback interface. Connection metadata also appears in .reckfell/editor/mcp-<pid>.json. It belongs to the current editor process and must not be committed to your project repository.

For stdio, invoke the editor executable directly rather than the console wrapper. Substitute the dock's port and provide the copied token through the MCP client's RECKFELL_MCP_TOKEN environment variable:

& ".\reckfell.windows.editor.x86_64.mono.exe" --mcp-stdio --mcp-port PORT

Discover an action

Pass these parameters to your client's tools/call request:

{
  "name": "reckfell_catalog",
  "arguments": {
    "action": "search",
    "query": "scene"
  }
}

Inspect the returned actions. For example, read the current document hierarchy:

{
  "name": "reckfell_scene",
  "arguments": {
    "action": "tree"
  }
}

The catalog groups editor, scene, object, resource, file, project settings, C#, runtime, spatial, workflow, voxel, connection and build operations. Use the class action on reckfell_catalog to inspect a native class's properties, methods and signals in your build.

Supply typed values

Engine values use an explicit type. A position vector is represented as:

{ "$type": "Vector3", "value": [0, 2, 0] }

Entity paths and resource references use these representations:

{
  "target": { "$entity": "Player" },
  "path": { "$type": "EntityPath", "value": "Player/Camera3D" },
  "material": { "$resource": "asset://Materials/Stone.rdata" }
}

These illustrate value encoding. Consult the catalog for each operation's fields: tools do not necessarily accept target, path and material together.

Check changes and results

Scene operations use Undo / Redo. Read undo_steps to determine the required undo steps. Resource changes can use snapshots. Writing an existing file requires its current expected_hash, and file checkpoints last for the editor session.

Running a game, building and other long operations return a job_id. This indicates submission. Poll reckfell_jobs with action get until completed or failed, then check result.ok and diagnostics. Runtime screenshots require rendering.

External tools

The reckfell_connections group configures external MCP services, discovers their tools and invokes them from the editor. Profiles include GitHub, Context7, Playwright and Blender. External process launches and tool calls present a confirmation for the specific operation.

Playwright requires your Node/npm installation; Blender requires its application-side bridge. These dependencies are not needed by Reckfell's native MCP control server.