Custom Roblox Studio MCP server (HTTP transport) bridging to a Rojo-built Studio plugin
Reaks MCP demonstrates solid tool coverage (46 tools) with consistent naming patterns (verb_noun: get_*, set_*, create_*, delete_*, search_*). All tools have descriptions (10-200 chars, within baseline 34-392 range). Input schemas are present and typed for all tools. However, parameter descriptions are minimal (often 1-2 words: 'Instance path', 'Property name'), falling short of the 72-char baseline for A+ tools. Output schemas are not documented, LLMs cannot predict return structure. Error handling is absent from descriptions; tools do not guide recovery. Security concerns: filesystem tools (read_file, write_file, delete_path) accept user paths without documented sanitization against traversal attacks. No tool annotations (readOnlyHint, destructiveHint) despite clear risk stratification (READ_ONLY, WRITE, DESTRUCTIVE, REVERSIBLE). Composition is strong: tools chain well (search_instances → get_properties → set_properties), and most are single-responsibility. Defaults are safe (depth=2, limit=100, mode='start_play'). Overall: good naming and schema presence, but thin descriptions and missing output documentation limit LLM reasoning quality.
Add a CollectionService tag to an instance.
Append content to a file on the host filesystem. Creates parent directories if needed.
Call a method on an instance, e.g. call_method('game.Workspace.Part','GetMass'). Args are coerced loosely; returns the (serialized) result.
Capture the Roblox Studio window as a PNG image from the host side.
Clone the instance at `path` (into new_parent, or its current parent).
Create an Instance of `class_name` under `parent`, optionally setting name/properties. Property values are coerced via the API dump (e.g. Size [4,1,2] -> Vector3).
Create a script. kind = 'Script' | 'LocalScript' | 'ModuleScript'.
Parameter descriptions are minimal (1-3 words: 'Instance path', 'Property name'). Baseline for A+ tools is 72 chars. LLMs cannot infer constraints, formats, or valid ranges from bare nouns. E.g., 'path' should be 'Instance path (e.g., game.Workspace.Part; use dot notation)'.
Output schemas are not documented. Tools return dicts/lists but LLMs cannot predict field names, types, or structure. E.g., get_tree returns 'nested JSON (name, class, children)' but no schema showing {name: string, class: string, children: [...]?}. This forces LLMs to guess and wastes tokens on discovery calls.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 68 | 2026-07-28+ | v2 |
Destroy the instance at `path`.
Delete a file or directory on the host filesystem.
FindFirstChild(name) under `path`. Returns the instance info or null.
All attributes of an instance as {name: value}.
Direct children of `path` as a list of {name, class, path}.
Recent Studio Output/console messages captured since the plugin connected.
A single instance's name, class, full path, parent and child count.
Place/game info: PlaceId, GameId, name, and the list of common services.
Read properties of an instance. If `props` is omitted, reads ALL scriptable properties for its class (via the bundled Roblox API dump). Roblox datatypes are returned tagged with __type (Vector3, Color3, Enum, Instance, ...).
Read a single property value from an instance.
Read the source code of a Script/LocalScript/ModuleScript at `path`.
Currently selected instances in Studio as [{name, class, path}].
List the DataModel services (Workspace, ReplicatedStorage, ...).
Current Studio run mode: 'stop' (edit), 'start_play', or 'run_server'.
CollectionService tags on an instance.
Instance subtree under `path` as nested JSON (name, class, children). e.g. 'game.Workspace'.
Insert an asset by id (game:GetObjects) under `parent`.
Search the marketplace for `query` and insert the top free model into Workspace.
List directory contents on the host filesystem.
List all scripts under `root` as [{name, class, path, lines}].
Create a directory on the host filesystem (with parent directories).
Reparent the instance at `path` under `new_parent`.
Move or rename a file/directory on the host filesystem.
Read a file from the host filesystem. Paths are expanded (~, env vars) and resolved to absolute. Large reads are truncated to keep responses sane.
Redo the last change (ChangeHistoryService).
Remove a CollectionService tag from an instance.
Rename the instance at `path`.
Start a play session, run `code` on the server datamodel, capture logs, then it stops. mode = 'start_play' | 'run_server'.
Run arbitrary Luau in edit mode; captures print/warn ([OUT]/[WARN]) and returns ([RETURN]).
Search for text in files on the host filesystem.
Find instances under `root`. by='name' (substring), 'class' (exact ClassName), or 'name_regex'. Returns [{name, class, path}].
Search script sources (and names) across the DataModel for `text`. Returns [{path, class, matches:[{line, text}]}].
Set a single attribute on an instance.
Set one or more properties on an instance (type-coerced via the API dump). Returns {applied:[...], errors:{prop:msg}}.
Replace the source code of a script at `path`.
Set the Studio selection to the given instance paths.
Control play-testing. mode = 'start_play' | 'stop' | 'run_server'.
Undo the last change (ChangeHistoryService).
Write content to a file on the host filesystem. Creates parent directories if needed.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk stratification in metadata (READ_ONLY, WRITE, DESTRUCTIVE, REVERSIBLE). Clients cannot determine which tools are safe to retry or require confirmation. E.g., delete_instance and delete_path should be marked destructiveHint=true.
Filesystem tools (read_file, write_file, append_file, delete_path, move_path) accept user-provided paths with only '~ and env var' expansion documented. No mention of path traversal sanitization (e.g., rejecting '../' sequences). LLMs can be tricked into reading /etc/passwd or writing outside intended directories.
Error handling is absent from tool descriptions. No guidance on recovery (e.g., 'If instance not found, try search_instances() first'). Tools do not document which errors are retryable vs. fatal. LLMs cannot self-correct on failures.