Skip to main content
Glama

The problem

Point three AI agents and two teammates at one Roblox game and you get chaos: one agent rewrites a file another is halfway through, nobody knows which remote name was agreed, and somebody stops a playtest that someone else was running.

Hive fixes this with a coordination layer that sits between your editors/agents and Rojo. Everyone, human or agent, reads and writes the same live project and sees the same context.

 Editors (humans) ──ws / CRDT──┐
 Agents (Claude, Codex…) ─MCP──┼──►  HIVE  ──files──►  Rojo (unchanged)  ──►  Roblox Studio
 Studio edits ◄────────────────┘   live docs · claims · ledger · playtest lease
                                        ▲
                                        └── StudioMCP: screenshots, play/stop, console

Related MCP server: Roblox Studio MCP

What makes it safe

Mechanism

What it does

CRDT documents

Every script is a Yjs Y.Text. Concurrent edits merge instead of overwriting.

Edit ops, not file writes

Agents change code with exact find-and-replace against the live doc. A stale or ambiguous edit fails with an error rather than clobbering a teammate.

Claims

An agent declares intent on a file or folder. Overlapping exclusive claims are refused, naming the holder and why. Claims auto-expire.

Shared ledger

One append-only log of claims, edits, decisions and playtests. hive_context gives any agent the full picture before it touches anything.

Decisions

hive_note kind=decision records conventions (remote names, API shapes) that every other agent then sees.

Playtest lease

One playtest at a time. Others can watch; only the owner can stop it (or force, explicitly).

Studio edits merge in

Changes made in Studio flow through Rojo to disk and are merged into the live doc as minimal diffs.

See and playtest the game

Open http://localhost:7777/ for a live viewer: edit-mode viewport, ▶ Playtest / ■ Stop, active claims, activity feed and the Studio console. Agents get the same abilities through MCP, so Claude can look at the game, run a playtest, read the console and report back.

Agent tools (MCP)

Tool

Purpose

hive_context

Read first. Claims, decisions, recent changes by everyone, file list

hive_read

Live merged contents of a script

hive_claim / hive_release

Declare / drop intent on a path or subtree

hive_edit

Exact str-replace on the live doc

hive_note

Share a note, question or decision with all agents

hive_screenshot

See the game (edit-mode viewport, optional camera pose)

hive_playtest_start / _stop / _status

Run playtests under the shared lease

hive_console

Read the Studio output log

Quick start

npm install
node src/server.js /path/to/rojo-project     # Hive on :7777
rojo serve                                   # in the same dir, for Studio sync

Register the MCP bridge with each agent (give each its own name):

{ "mcpServers": { "hive": {
    "command": "node", "args": ["/path/to/hive/src/mcp.js"],
    "env": { "HIVE_ACTOR": "agent-combat", "HIVE_URL": "http://localhost:7777" } } } }

Studio control uses Roblox's StudioMCP binary (macOS default path; override with STUDIO_MCP=).

npm test      # 13 tests

Layout

src/core.js     Hive: CRDT docs, claims, ledger, playtest lease (pure, no I/O)
src/server.js   disk mirror + WebSocket sync + HTTP API + viewer
src/studio.js   StudioBridge: screenshots (cached), play/stop, console
src/mcp.js      MCP tools for agents
src/viewer.html live viewer

Status

Working: core, server, MCP tools, Studio bridge, viewer, tests (Studio is faked in tests; live-Studio verification is pending).

Roadmap:

  • VS Code extension: Yjs binding, presence cursors, claims panel

  • Non-script instances (parts, properties) via a Studio plugin

  • Playtest-time viewport capture

  • Auth and multi-project hosting

  • Code-OSS fork once the extension proves out

Credits

Builds on Rojo (MPL-2.0) and Yjs. Not affiliated with Roblox Corporation.

License

MIT

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      C
      maintenance
      Enables AI agents to interact with Roblox Studio via a token-efficient MCP server, live two-way script sync, and zero-friction plugin setup.
      -
    • A
      license
      A
      quality
      A
      maintenance
      Enables AI agents to inspect, create, edit, debug, and playtest projects inside the Roblox editor via 29 lean tools, with push-based SSE transport, editor-safe script edits, and batched undoable writes.
      35
      1,088 npm
      14
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI assistants to explore, edit, and automate Roblox Studio sessions locally, including browsing instance hierarchies, reading and modifying scripts, manipulating properties, creating instances, building terrain, and running playtests.
      875 npm
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Enables AI agents to drive a running Roblox Studio in real time over a single WebSocket, executing whole Luau programs, installing in-engine controllers, hot-patching scripts during live playtests, and receiving errors, assertions and other events as pushes. It also reaches Roblox Open Cloud for the open place and answers visual questions via a vision sidecar, exposing all of this through nine MCP tools.
      9
      1
      MIT