remote-claude

Chat with Claude Code running on your own PC from the dev menu's Claude tab, inside a live dev-branch server. Claude can act on your running server ("jump me", "give me 100 coins") or change the code and propose a deploy.

• Dev-channel servers only; prod branches are refused with no override.
• Only the Roblox users you list can prompt, and only on servers on the session's branch.
• It runs on your Claude subscription only. API keys, Bedrock, Vertex and Foundry logins are refused.

Requirements

• Claude Code installed and logged in with your Claude.ai account (claude auth login).

• cloudflared (installed with winget on Windows when missing; macOS: brew install cloudflared). The tunnel is an account-free Cloudflare Quick Tunnel.

• The dev-server: Node 20+ or Bun 1.3+. The template has @typetorch/dev-server in its dev dependencies, so bun install gets it. Its range may hold you on an older minor version (^0.2.3 stays on 0.2.x): bun add -d @typetorch/dev-server@latest gets the current release. Without it: npx (see below), or a checkout next to your game (bun install in it).

• An Open Cloud key with universe-messaging-service:publish: OPENCLOUD_DEPLOY_KEY, else the shared OPENCLOUD_API_KEY (or ROBLOX_API_KEY, its alias). The dev-server tells game servers where the session is with it. It finds the key the same way the CLI does: the real environment, then the game repo's .env (next to typetorch.json; no parent folder's; --env-file <path> or TYPETORCH_ENV_FILE read another file instead). Keep the key in that .env, which is gitignored. The backend's keys (TYPETORCH_API_KEY, TYPETORCH_ADMIN_TOKEN) are never read here. When you play on another PC, screenshots arrive as CaptureService uploads, which OPENCLOUD_ASSETS_KEY (else the shared key) downloads.

• Experience settings: Allow HTTP Requests on. ServerScriptService.LoadStringEnabled on for run_luau, only in a place where you want it (a test place: kernel deploy --loadstring, CLI 0.7.5+); keep it off in a live game. kernel deploy patches leave it as it is; doctor reports it. Optional: Allow Mesh / Image APIs (and an ID-verified 13+ owner) for images Claude shows in the chat; Allow Loading Third Party Assets for Toolbox inserts.

Nothing else to set up in Roblox: no secrets.

Start a session

On a dev-channel git branch of your game:

git switch dev
bun run typetorch remote-claude --users 111111111

typetorch dev is the same command, shorter (bun run typetorch dev --users 111111111, or tt dev with the CLI installed globally).

• --users: the Roblox user ids allowed to prompt (no wildcard). It is remembered in .typetorch/remote-claude.json (git-ignored); omitted, the last list passed in this repo is used, else the owners in typetorch.json (members with the role "owner"; dev-server 0.3.3+).

typetorch remote-claude runs the dev-server with the same arguments. It finds it in TYPETORCH_DEV_SERVER, the game's node_modules (where the template's bun install puts it), next to the CLI, or a sibling ../dev-server checkout. Other ways to start it:

npx @typetorch/dev-server remote-claude --users 111111111                # from npm
npx -p @typetorch/cli -p @typetorch/dev-server typetorch remote-claude --users 111111111
bun ../dev-server/src/index.ts remote-claude --users 111111111           # a checkout

It prints a pairing code (and copies it to your clipboard):

pairing code: ABCD-EFGH-JKLM-NPQR-STUV-7QX2  (valid until 18:02, paste it into DEV > Claude in game)

In game: open a server on that branch (/tt new dev), open DEV > Claude, paste the code. That server is paired for up to 3 hours. Each code pairs one user on one server once; the next code prints right away.

Useful flags: --branch <name>, --max-prompts <n> (default 50), --no-deploy, --protect <globs> (files Claude may not edit), --code-ttl <minutes> (default 180), --model <name>, --max-budget-usd <n> (a cap per run on Claude Code's own cost estimate), --cli <entry> (the TypeTorch CLI used for deploys), --env-file <path>, --repo <dir>, --port <n>, --no-install (no bun install in a fresh worktree), --no-clipboard, --no-announce (local testing), --no-tunnel (no Cloudflare tunnel, for a local client only; what serve --claude uses), and, in dev-server 0.4.0 (not on npm yet), --deploy-branch <b> (where Deploy goes: a dev-channel branch; default the session's branch). Terminal commands while it runs: code, users, revoke <userId>, rotate, status, cancel <promptId>, quit.

It also runs in Docker (locally or as a Coolify app): the dev-server repo's Dockerfile and compose.yaml, with the users in TT_USERS and claude auth login inside the container. See the dev-server README.

Check: the Claude tab shows the session instead of "Not connected: start the dev server on this branch".

Live and Code modes

Pick the mode with the Live / Code pill next to "+". A conversation can switch modes; follow-ups continue the same Claude session.

	Live (default)	Code
What Claude does	acts on your running server	edits the branch in its own git worktree (../<game>-remote-claude)
Files	read only	edit, then bun run build
Game tools	all: run_luau (you approve each snippet), game_logs, inspect, find, game_status, screenshot	read-only ones (no run_luau)
Result	an answer, or a change on your server	a commit and a deploy proposal

• run_luau runs only on the server you prompted from, for you; you approve every snippet in the chat (once, always in this chat, or deny). Each run leaves an audit record. Prod servers never run it.
• A code run that changed files is committed and shown as a card: files, lines added and removed, a summary. Deploy runs typetorch deploy and your server hot-swaps to it; Discard undoes the commit. Only you decide; after 15 minutes it is discarded.
• If your approval policy needs a person for that branch ("approval": "all"), the deploy card says "Waiting for your approval on your PC": run bun run typetorch approve <id> in the game folder (the dev-server's terminal prints the id). With "approval": "prod" dev deploys go straight through.
• The dev-server never deploys to a prod-channel branch. Commits stay on its worktree branch; bring them home with git merge remote-claude/<branch>. Nothing is pushed.

Attach context

The "+" menu:

• Dex path: the instance selected in the Dex.
• My logs, Server logs, Player logs...: log history (about 64 KB each), passed as untrusted data and never logged.
• Screenshot: a capture of your screen, which you can crop and draw on before sending (pen in red, yellow, white or black, three sizes, Undo / Clear, Ctrl+Z / Ctrl+Y). When you play on the PC that runs the dev-server, it picks the capture file up directly.
• Toolbox: see below.

Tap any image in the chat to open it in big picture mode (zoom with the wheel or a pinch, drag to pan).

Toolbox (Creator Store)

Only when you ask for it. Pick Toolbox in the "+" menu: that one message lets Claude search the Creator Store (free assets, from your PC). The chip clears after you send; without it Claude has no Creator Store tools at all.

• In Live mode Claude can insert an asset you picked into your server. Every insert shows an approval card (what loaded, what will be removed, Anchor, Keep scripts off by default) with Insert / Deny, and inserted assets get a Remove button. Scripts are stripped. Needs "Allow Loading Third Party Assets" (Studio > File > Experience Settings > Security; it applies to every server of the experience, prod included).
• In Code mode it records the asset in toolbox.lock.toml instead.

From Studio

The TypeTorch Studio plugin has a Claude button that opens the same chat in Studio. It needs typetorch serve with --claude on the machine with the game checkout:

bun run typetorch serve --claude

Versions. serve from CLI 0.12.2+ starts remote-claude with --deploy-branch, which only @typetorch/dev-server 0.4.0 knows. 0.4.0 is not on npm yet (npm has 0.3.3, which exits with unexpected argument): until it is, set TYPETORCH_DEV_SERVER in the environment to a checkout's entry, for example TYPETORCH_DEV_SERVER=../dev-server/src/index.ts (run with Bun; the path is relative to the game folder) for a typetorch/dev-server checkout next to the game, with bun install run in it. Apply, Deploy to dev and the per-dev worktrees below are dev-server 0.4.0.

• serve runs remote-claude on this machine without a tunnel (remote-claude --no-tunnel --no-announce --no-clipboard --deploy-branch <serve's --dev-branch>) and pairs with it on the Studio user's behalf: nobody pastes a code. A remote-claude you start yourself in the repo works too (serve finds it through .typetorch/remote-claude.session.json). serve doesn't restart it when it exits: read serve logs, then serve restart claude.
• --claude-users <ids>: the Roblox user ids allowed (remembered). Without it, the owners in typetorch.json. Studio prompts as the Roblox user it runs as.
• Code mode only. Live mode works only from the in-game tab: Studio's Live pill says so. Claude edits a copy of the branch, then the tab shows a card with the summary and the changed files:
  • Apply puts the change into the checkout serve's Rojo syncs to Studio, uncommitted, so you review and commit it there. A change to lines that changed there too is refused, and the checkout stays as it was.
  • Deploy to dev runs typetorch deploy of that commit to serve's dev branch (dev, or serve --dev-branch). With "approval": "all" it waits for typetorch approve on your PC.
  • Discard drops the commit. With no answer in 15 minutes the change is discarded; only the dev who sent the prompt decides.
• Each dev gets their own worktree, ../<game>-remote-claude-<user id> on remote-claude/<branch>-<user id>, so several devs use Claude at once: each dev's prompts run one at a time, but never wait for another dev's. Apply changes the checkout one change at a time. Each worktree is brought up to the branch before every run, so a change another dev applied reaches your runs once it is committed.
• A prod-channel checkout works too (a local session only): Claude never commits to that branch itself, and Deploy goes to serve's dev branch.
• Claude reading the open place is planned. dev-server 0.4.0 offers Studio prompts the game tools, and the plugin (0.7.1+) answers inspect, find, game_logs and game_status on the place open in Edit, read-only. It never runs code it receives: run_luau is answered "not available in Studio". serve on CLI main doesn't relay these requests to the plugin yet, so today those tool calls time out and Claude works from the files.
• The "+" menu has Dex path (the Studio selection), My logs, Server logs and Screenshot (in Play only) and Toolbox. Recent chats are listed, and the chat follows you between Edit and Play.
• Replies stream to the tab over a WebSocket (/tt/claude/stream), so they show as Claude writes them.
• A helper on another machine pairs their Studio with serve --remote (A Studio on another machine); add their user id to --claude-users. Pair again after updating the plugin, so serve knows their Roblox user.

Security in short

• The tunnel URL is public, so the dev-server authenticates everything: single-use pairing codes bound to the tunnel, 5-minute access tokens bound to one user and one game server, rotating refresh tokens, replay protection, and a loopback-only HTTP server.
• Game servers keep the session URL and tokens in memory; no client ever receives them.
• Claude runs with a tool allowlist in a separate worktree; it can't push, can't use the web, and never sees your Open Cloud key or .env values. Build and tool configuration files are protected.
• Game data (names, logs, instance paths) reaches Claude labelled as untrusted data, never as instructions.

Full details: the dev-server README.