Changelog

Credits now go about 8% further than before October 30. Every build costs fewer credits; prices and plan credits are unchanged.

Scenery, buildings and props are designed new for each game; Creator Store models are found by search when you ask for ready-made or realistic ones.

Every free Starter account now gets 1 credit a day on DeepSeek V4 Flash, refilled at 00:00 UTC, with no card. There is no extra credit on the sign-up day, and free credits no longer need the RoCode Discord.

The Luau playground runs Luau, Roblox's scripting language, in your browser, with strict type checks and shareable links. It needs no account.

With plugin 3.9.14 or newer, RoCode can place models from its own library of hand-picked free Creator Store models, such as trees, rocks, lamps, fences, crates, houses and stalls. It scales each one to the height it needs, and the plugin removes any script inside. It cannot search the Toolbox or insert a model you name.

With plugin 3.9.14 or newer, a Build run takes screenshots in Studio of what it built, in edit mode, before it finishes. An AI model that reads images checks them for clear problems, such as floating parts, gaps or the wrong scale, and RoCode can fix what looks wrong. It runs at most once a run, on every plan.

Four free tools for Roblox creators, with no account: a game idea generator, a DevEx calculator, a Robux tax calculator and a thumbnail size checker. They are all on the free tools page.

Roblox Studio scripts collects the main scripts from our guides, sorted by what they build. Each one is type-checked, with the date of its check.

The free Starter plan now gives 1 credit a day, 2 on the day you sign up and 2 on the day you join the RoCode Discord, refilled at 00:00 UTC, on DeepSeek V4 Flash. After a few builds, free accounts are asked to join the RoCode Discord to keep getting free credits. Paying never needs Discord.

Superseded on October 24, 2026: Starter is 1 credit a day with no extra credit on the sign-up day and no Discord step.

GPT-OSS 120B is no longer in the model picker. DeepSeek V4 Flash is the free model, and a chat that used GPT-OSS 120B now runs DeepSeek V4 Flash. GPT-OSS 120B stays as a backup model: if DeepSeek V4 Flash does not respond, it can finish the run.

With plugin 3.9.13 or newer, Studio connects to the chat by itself when it is signed in to the same Roblox account you use on RoCode. Otherwise click Use a pairing code in the RoCode panel and type the 8-character code into the chat, as before.

GLM 5.3 is retired. GLM 5.3 Flash, the default on paid plans since 12 October, takes its place, and a chat that used GLM 5.3 now runs GLM 5.3 Flash.

The free Starter plan now gives 2 credits a day, refilled at 00:00 UTC, still on DeepSeek V4 Flash or GPT-OSS 120B. Paid plans are unchanged.

Superseded on October 17, 2026, and again on October 24, 2026: Starter is 1 credit a day on DeepSeek V4 Flash, and GPT-OSS 120B is no longer in the model picker.

Studio is now $19.99 a month with 320 credits, or $167.99 a year. Ultra is now $29.99 a month with 500 credits, or $251.99 a year. Pro stays at $9.99 a month with 150 credits. Models by plan are unchanged.

Max is no longer offered. If you already subscribe to Studio, Ultra or Max, you keep your price and credits until you change plan.

Credit packs are now 50 credits for $4.99, 220 for $19.99, 580 for $49.99 and 1,200 for $99.99. Pack credits never expire and are used after your plan credits.

The prices, plans and free credits in this entry were replaced on October 11, 2026 (Studio and Ultra prices, Max withdrawn, credit packs) and the free allowance again on October 17 and 24, 2026. Annual plans came back at 30% less. See Pricing.

Usage is now counted in credits, not tokens. A credit measures what a run costs to compute. Stronger models and longer runs use more. The chat shows the credits a run is using while it works and the total under every reply.

Pro is $9.99 a month with 150 credits. Studio is $24.99 with 400 credits. Ultra is now $49.99 with 900 credits. The new Max plan is $99.99 with 2,000 credits. Starter stays free with 3 credits a day, refilled at 00:00 UTC, which is the same daily allowance as before. Unused plan credits do not carry over.

If you already subscribe, nothing changes for you. You keep your price and your monthly allowance, now shown in credits, until you change plan yourself. Annual plans are no longer sold. If you have one, it stays as it is and its credits reset every 30 days.

Models now go by their real names: RoCode v4.0 is GLM 5.3, now the default on paid plans. RoCode v3.3 and v3.6 are retired, and a chat that used one moves to the model that replaced it. The free model is now DeepSeek V4 Flash, run on US servers. Three new models join them: Claude Sonnet 5.5 on Pro and up, Claude Opus 5.5 on Studio and up, and GPT-6 Astra on Ultra and up. The model menu shows what a task usually uses on each model. Every account gets 3 free credits a day for DeepSeek V4 Flash, on any plan.

Effort for paid models. Pick Low, Medium or High next to the model. Higher effort thinks longer and uses more credits.

If a paid model is slow to start, RoCode first tries the same model through another provider. If GLM 5.3 still fails, DeepSeek V4 Flash can finish a text task, charged at its lower price, and the line under the reply says so. On other models the run stops, and a run that fails on our side gets its credits back. Runs on your own key never switch.

Runs that you stop count up to where they stopped. When a run fails because of a problem on our side, its credits go back to your balance automatically. One run pauses after a set number of credits and asks before it goes on. The limit depends on the model, and it can be lower on smaller plans. You can continue or let it finish up.

Usage packs are now credit packs, from 50 credits for $4.99 up to 800 credits for $54.99. Pack credits never expire and are used after your plan credits.

The account page has a new Usage section. It shows your balance, your reset date, the last 14 days and every run with the credits it used. You can download the list as a CSV file.

Using your own OpenRouter key for Claude, GPT and other models is now part of Pro, Studio and Ultra. Runs on your own key are billed by OpenRouter and do not use your plan's quota. The Free plan keeps RoCode v3.3 every day.

Saved OpenRouter keys are now encrypted at rest. Keys that had not been used for 30 days were removed, so paste yours again in the model picker if you need it.

Added September 30, 2026: the model you pick must support tool calling, and not every model completes runs with RoCode's tools yet, so try yours on a small task first.

Connecting Studio now uses a one-time 8-character code. The plugin shows it, and you type it into the chat. Before, any Studio on the same network could connect to an open chat, so on a shared school or office network someone else's Studio could take your connection.

The agent keeps working until your whole request is done instead of stopping after its first change. Several changes to the same object in one step now each apply once and report correctly, and deleting numbered copies (Door[1], Door[2]) removes the ones you meant.

Builds no longer come out black when the model sends colours as decimals or hex codes.

Starting a second chat no longer stops an agent that is still working in another one.

If a run fails after it has already changed your place, the chat offers Continue instead of silently starting the whole prompt again.

Links in billing emails now go to your account page after you sign in, and the plan change window shows the current annual prices.

Annual plans are now $83.99, $167.99 and $251.99 a year for Pro, Studio and Ultra, up to 30% less than paying monthly. Annual subscribers who joined earlier keep their price.

RoCode v4.0 is in the model menu for every paid plan. It is the smartest included model RoCode offers and it thinks longer before it acts, so it suits big multi-script systems more than quick edits.

RoCode v4.0 uses your monthly budget about 10 times faster than RoCode v3.6, and the menu says so. RoCode v3.6 stays the default, so nothing changes unless you pick RoCode v4.0. The tokens left line in the sidebar adjusts to whichever model is selected.

The chat remembers exactly which model you picked, the quota meter counts down from 100% remaining, and the chatbox now stretches when you zoom the page.

New Ultra plan at $29.99/mo, with three times the Pro agent budget. It is for builds that run into the Pro or Studio ceiling before the month is out. Annual billing is $209.99/yr, the same 42% saving as Pro and Studio. Annual prices changed on September 27, see v4.0.1.

Nothing changes on Starter, Pro or Studio. Existing subscribers stay on their current plan and price.

The design language is now SpringGlass 2. The static background grid is gone; blooms and film grain stay.

RoCode v3.6 is fixed. Every Pro job was failing after the Sep 6 model switch: the new engine rejected a forced “reasoning off” setting. Both Free and Pro now use the model’s default reasoning, so Pro runs again.

RoCode v3.6 now runs as the updated paid model.

Pro (and Studio) agent budgets go further on the same plan. You get more quota per month because the new model is cheaper to run, and the dollar budget did not change.

The whole site is remastered to SpringGlass 1, same design language as Oracle and StackLess. Near-black chassis, lilac type, skeuo glass rims/sheen, Nunito, and the purple accent family replace the old Rocode-blue shell.

Buttons and press/hover no longer use CSS easing for interaction. Tilt and scale run on real underdamped springs; the glass face and the label tilt together (5°) and grow ~3.5% on hover.

That tilt-and-grow hover is now Spring Glass on Oracle and StackLess too, last Spring Glass 1 pass.

The WebGL starfield / background particles are gone. SpringGlass 1 keeps the static grid, film grain, and blooms only.

After successful Studio mutations without complete_task, the relay synthesizes a one-shot finish, think/query-only turns no longer auto-complete. Scripts must still get check_script on the changed path before complete (prompt + tool description enforce it).

Early-dispatched tools that grow on the same step re-paint correctly. Step text always lands before that step’s tools (even when tools arrive first over SSE), so streaming prose no longer ends up under the wrong tool block.

New reliability telemetry: agent_llm_turns, richer tool_invocations (path, turn, recovery, soft-skip, error detail), job finish_reason / synth flags, plus product events for Studio batches, LLM turns, playtest blocks, check_script findings, plugin pair/session, and plugin diags. Ops SQL pack + migration included.

Studio plugin: whitespace/CRLF-tolerant edit_script recovery (unique matches only, with re-read hints), soft-skip invalid property members instead of red-X’ing the turn, longer playtest-stop wait with retries, rate-limited diag emit over the plugin WebSocket, and clearer UDim2 string parsing.

Hitting 100% mid-job no longer cancels the in-flight prompt or tears down Studio presence (that looked like “plugin offline”). The job finishes; the paywall / composer lock applies on the next send. Relay billing notes overage for telemetry only.

Continue from checkpoint now posts only Continue the job from the last checkpoint., no duplicated error/tools dump in your prompt (that context is already in the turn above; the relay still gets the recovery brief off-meta).

Playtest, Script Injection, edit mismatch, and path-not-found no longer pause the live job behind a Continue panel, failures stay inline in the tool bubble and feed back to the model. Transient tips still show when useful.

Transport failures classify from the message first (timeout / Studio offline / playtest) so a wrong hint like permission_denied on a timeout cannot win. check_script findings track separately from transport crashes.

Hero remastered with a Studio-chrome live demo frame, clearer subtitle, and muted-autoplay caption, then the redundant “plugin launch” promo section was removed so the first viewport stays one composition.

Checkout honors a client referral code if claim failed earlier (network blip), creates a checkout_started attribution when needed so the referrer still gets credit, and idempotent claim returns the promo for the already-attributed code. Self-referral rejects are recorded once; login clears permanent claim failures so they aren’t retried forever. Referral landing visitor key aligns with analytics UUID.

Agent job wall clock set to 5 hours (plus client wait buffer), stream auth tokens stay long-lived separately.

Unsubscribe / preference links return a clear 503 when the preference secret is misconfigured instead of crashing the handler.

Fail-fast when the Studio WebSocket is down before dispatching a batch; plugin pair/session outcomes recorded; sync-relay-overrides expanded for completion + telemetry wiring. Tests cover message order, quota settle, completion synthesis, tool failure codes, and run telemetry.

Chat now starts agent jobs on Railway with your session JWT by default, no Vercel /api/agent-session hop on the happy path. One-shot fallback to the legacy session path remains if Railway rejects the JWT (set ROCODE_DIRECT_AGENT_JOB=0 to force rollback).

Railway loads full conversation history from Supabase by conversation id. The browser no longer re-uploads megabyte histories on every send/retry. Context stays unlimited, only the existing compact-when-huge behavior applies (no new message/think caps).

Chat profile fetches skip Stripe reconcile (skipStripe=1). After a job, one delayed refresh replaces the old 3 to 4× poll storm; quota meter prefers X-Quota-* headers from /agent-job when present.

Healthy Studio presence SSE is no longer torn down when the quota meter redraws. Presence and plugin state share one long-lived stream, no duplicate /plugin-state/stream while presence is up.

Agent job SSE sends step deltas instead of replaying the full steps array every tick; the client merges patches into the live steps cache (less bandwidth/CPU on long tools runs).

Empty agent replies no longer burn three automatic full-job restarts, UI Retry still works. If the relay already ran in-job LLM recovery, the client does not immediately stack another full restart for the same failure class. Quiet one-shot stall retry (no Studio tools yet) is unchanged.

Idle/hover conversation prefetch loads metadata only (id, role, timestamps), not full HTML message blobs ×12. Opening a chat still hydrates full content.

Plugin connect code (rc_live_…) is cached in session storage with a TTL so boot does not hit /api/generate-api-key every time. BYOK send trusts cached settings; the relay gate still rejects bad config.

New chats get a local fallback title immediately; AI title upgrade stays async and never blocks send. Next-prompt chips generate lazily (when visible / composer idle) with a short in-memory cache, same feature, fewer wasted OpenRouter calls when chips are never seen.

Analytics track-event rate limit uses a token-bucket/RPC counter instead of COUNT(*). Session heartbeat is visibility-aware and less frequent. Checkout paid-tier polling is lighter. BYOK save validates key auth + model id format without downloading the full OpenRouter /models catalog. Unused /api/activation-state removed.

Reload always returns you to the exact chat you left, including a blank New chat. It no longer jumps to the newest thread when you were on a new chat or a different conversation.

Studio no longer needs a paste-and-Connect code. Keep RoCode chat open while logged in, open the plugin, and it pairs automatically on the same network via event-driven IP presence (SSE from chat, WebSocket pair from Studio, up to ~2 minutes wait if chat opens later). No polling heartbeats.

Studio plugin remastered for auto-pair: pairing mode UI (“Waiting for chat…”, “Pairing with chat…”), saved-key reconnect vs fresh pair, clearer connecting/reconnecting status, Reset restarts auto-pair, and HTTP permission checks for both www.rocode.app and relay.rocode.app with actionable allow-list guidance when stuck on Connecting.

Connect modal is now Get the Studio plugin: download, keep chat open, allow HTTP. Header CTA is Download plugin (Creator Store link). Connect-code copy/reset UI is gone from the flow; activation checklist says download the plugin; agent-session errors tell you to open chat for auto-pair.

After a successful agent turn, chat shows 2 to 3 one-click follow-up chips under the completion card. Generated from the turn summary via a dedicated API (/api/generate-next-prompts), with rate limits, retries, and a local fallback if the model is empty, chips fill the composer so you can send immediately.

Finished runs get a richer completion card with a short “what changed in Studio” wow summary (up to six human lines from tool steps), then the next-prompt chips.

Empty agent replies are retryable: Retry this prompt re-sends the same user message (alongside Continue from checkpoint). Error copy and analytics updated; Retry/Continue still work after reload when error panels are restored from history.

Chat switches and reloads paint from a session message-row cache, then revalidate from Supabase. Recent-window fetch + background older-row backfill, idle sidebar prefetch, in-memory conversation view cache, and optimistic message painting so switching chats feels instant.

Agent-session responses are warm-prefetched and buffered so the first send skips cold RTTs. Stripe tier reconcile on submit runs in the background instead of blocking latency. Stale free-model ids migrate to the current FLASH alias (~deepseek/deepseek-v4-flash-latest).

Site-wide: Nunito trimmed to weights 400/600/700; Lucide CDN dropped on static pages in favor of local icons.js; homepage hero video is lazy (poster + IntersectionObserver); pricing/billing libs load when the pricing section nears the viewport; script loader uses idle scheduling, cleaner retries, and deferred non-critical work; JS/CSS cache headers extended to 1 day + long SWR.

Empty-state hint “Describe what you want to build” removed, title alone when connected, connect guidance when not. Character counter stays hidden until you are near the 6000-char limit. Message action chrome (copy/etc.) lazy-binds for faster history paint.

Feedback page fully removed (nav, sitemap, home/login links). Old /feedback URLs permanently redirect to Info. Discord join lives as an icon in the Configuration panel (bottom-right). Info “Get in Touch” is founder email support@rocode.app. llms.txt updated with founder + Discord and no Feedback link.

Lifecycle welcome and plugin-not-paired emails rewritten for auto-pair, keep chat open, open Studio plugin, no paste/Connect step.

Unused growth Template system fully removed from Supabase (templates, ratings, verifications, build receipts, template funnel analytics, template loop/page kinds) and related tests, never shipped as a product surface.

Rotating a plugin API key now updates an existing row, or creates one if missing, no silent no-op on first rotate.

Button spring-click and glass-hover no longer crash when the event target is an SVG/text node inside a button (safe closest lookup).

Getting-started UI still boots if rocode:ready already fired before the deferred activation script loaded.

Relay plugin presence + WebSocket pair path, sync-relay-overrides wiring, agent-session messaging for auto-pair, OpenRouter/platform FLASH model alias, pricing-mounted event for lazy home pricing, and assorted chat/plugin polish from the auto-pair and speed work.

Added September 30, 2026: auto-pair was replaced on September 27, 2026 (v4.0.1) by a one-time 8-character code that the plugin shows and you type into the chat. Since October 16, 2026 (plugin 3.9.13 or newer), Studio connects by itself when it is signed in to the same Roblox account you use on RoCode, and the code stays for every other case.

Superseded. There is no trial. Plans and prices changed on October 2 and October 11, 2026; see Pricing for the current ones.

Chat shows a Getting started checklist (connect Studio → first completed job → keep going). It tracks progress, then hides forever once onboarding is done.

Free accounts get a reverse trial of 3 completed agent jobs before the upgrade nudge. Annual billing unlocks after three weeks of completed jobs.

Lifecycle emails for welcome, unfinished pairing, connected-but-no-build, first completed job, day-3 recipe, dormant winback, and cancellation rescue, with a preferences page to turn them off. Auth, security, billing, and payment-recovery mail still always send.

Account → Invite builders: after a completed agent job, share your link. Friends get 20% off their first Pro month; you earn 20% commission on that first paid checkout (45-day attribution, self-referrals blocked).

Cancel flow asks why, then offers keep / pause next renewal / manage plan before confirming. Abandoned Stripe Checkout sessions get recovery links; promotion codes and automatic tax work on checkout.

Studio tool batches use clearer timeout contracts: mutating tools vs checks, safer retry signals when a timeout leaves outcome unknown, and fewer duplicate/unsafe retries after a partial Studio write.

Revenue control plane tracks visitor attribution (UTM, referrer, country), activation milestones, checkout attempts, churn reasons, and daily growth-loop scores. Paid ads and localization stay gated until quality thresholds are met, no scaling unproven channels.

Privacy policy updated for lifecycle email delivery records, referrals, experiment assignment, and coarse country attribution used for scale measurement.

Removed Plan for complex tasks entirely: the plan panel, toggle, plan tools (make/read/edit/remove), planning/executing phases, and related relay gating are gone. The agent runs the normal tool workflow only.

Tool execution is more reliable with fewer errors.

Bring Your Own Key is significantly simpler and much more robust.

Double-Enter and rapid sends no longer start duplicate agent jobs: a per-conversation lock blocks another submit until the job session exists. Continue restores cleanly if a submit aborts early.

Studio create/edit tools handle more agent quirks: class aliases (Wedge, Cylinder, Ball), material aliases (Leaf→LeafyGrass, Stone→Slate, etc.), GUI size/position tables, and invalid layout props are skipped instead of failing the whole batch. Pending tool batches queue instead of dropping when Studio is busy.

check_script and other structured failures now surface real error text from nested data.errors, not a generic “Tool failed.” Studio disconnects are distinguished from generic timeouts.

Quota and budget settlement clamps to what SQL actually applied, so there are no repeat debits when you hit the limit mid-run. Missing OpenRouter usage estimates use a 500-token floor instead of 10k. Pro model downgrade no longer double-applies the 70% budget cap.

Reserved domains (example.com, test.com, etc.) are rejected at signup with a clear message. OTP rate limits fail closed when the store is down, and failed sends release their slot so you are not locked out. Disposable emails are blocked on verify too.

Missing plugin API keys now prompt you to open the connect modal and generate one, instead of a vague “Missing required fields” error. Agent session returns a validated key after server check.

Custom model settings lazy-load so chat boots faster for platform-model users. Saves no longer require a separate encryption secret.

Profile load skips a redundant Stripe round-trip; upgrade buttons keep their Processing… label until checkout finishes.

Relay preconnect runs only on the chat page (not login/account). Relay health retries once on cold-start blips. Shared profile and BYOK fetches dedupe across bootstrap.

Telemetry FK noise, RLS hardening on checkpoints and internal tables, submit-token secret fallback for Vercel↔Railway alignment, and assorted agent-prompt guidance for shapes, materials, and UI layout.

The agent’s think tool now shows as its own collapsible thinking bubble in chat, focus + reasoning, saved with the turn, and restored when you reopen the conversation.

Thinking is used only when the next move is genuinely unclear, routine create/edit/check/complete packs into one turn. After successful Studio tools, a missing complete_task is finished automatically so runs don’t stall on a nudge loop.

Agent turns persist faithfully: thinking bubbles, tool calls, and interleaved text survive reload, stop, and Continue, without dropping mid-run history.

Stopping a run pins a clear Stopped card on the bubble (tools kept, live chrome cleared). Enter only sends or queues, it never stops generation; use the stop button for that.

You can still queue the next prompt while an agent is busy even if the plugin briefly disconnects, cold sends still require a connected plugin.

Checkpoint recovery briefs list tools already run and tell the agent to skip successes, verify only if needed, then call complete_task.

check_script Output is opt-in (include_output), scope guidance prefers single/root over full-game unless you ask, and long Studio tool batches no longer get cut off by hard plugin timeouts.

Pro/Studio token budgets reset on the paid window even for annual plans; Stripe re-checks on profile load; checkout handles async payments; mid-cycle plan changes no longer wipe used budget; checkout shows a clear “still activating” message when provisioning lags.

OTP verify rate-limiting fails open on DB blips so a brief outage can’t lock every sign-in, while a valid code is still required.

Assorted fixes across job completion detection, chat save races (thin Stop stubs upgrading to full turns), BYOK/model reasoning config, and relay pricing catalog caching.

While an agent is running, send another prompt (with or without files) to add a queued bubble above the composer, preview of the first words, pencil to edit in-place, trash to remove. Queued prompts run one after another when the current job finishes.

Attach any text file to a normal send (paperclip or drag-and-drop), markdown, code, configs, logs, and more. Binary files are skipped. Attachments are part of the prompt, not a separate “file queue.”

While generating: empty composer shows a stop square; typing text (or adding an attachment) switches it back to send so you can queue the next message.

Failed runs no longer auto-retry. You get a clear error and a single Continue button that resumes from the last durable checkpoint, without re-running attached files or draining the prompt queue by accident.

Done sound, completion card, and queue advance only fire after a real complete_task, stall recovery no longer pretends the job finished.

Agent jobs now save durable checkpoints so Continue can pick up mid-run after a pause, disconnect, or tab switch.

Background / multi-tab banners name the conversation that finished or needs attention, so you can jump straight back.

Paths with dots or spaces in instance names work via quoted segments (e.g. game.Workspace["Colt .45"]), including script-relative requires.

Queue handoff no longer drops a prompt if send is blocked (quota, plugin, concurrency); pending-chat queues migrate when the conversation id is created; Continue ignores composer attachments; paused error panels + Continue survive refresh; prompt queues persist in localStorage.

Your plan now re-syncs with Stripe when your profile loads, missed webhooks no longer leave Pro/Studio stuck on or off incorrectly.

Playtest Output from Server and Client is mirrored into Edit mode, so the agent can read the same Output window lines you see during and after playtests.

Failed and stopped agent turns now save properly, and reverting a message no longer leaves leftover assistant stubs in the thread.

Fixed voice-to-text not landing in the chat input, transcripts append correctly and the send button updates as expected.

OTP verify is rate-limited more reliably against brute-force attempts, and login resend Turnstile works more consistently.

Clearer tool status badges and check_script results, fewer false warnings, and Output lines show their Server/Client/Edit source.

The agent is less censored on language, it can swear freely without tone-matching, softening, or content warnings.

Modals are snappier: expensive backdrop blur is gone so zoom and overlays stay smoother, especially on lower-power devices.

Assorted reliability fixes across Studio tools, plugin connections, and chat persistence for a more stable agent experience.

Updated the site UI across home, pricing, and account, clearer plan cards, cleaner layout, and a more polished look.

Pages and the app load faster, less wait when opening the site and moving between views.

Switching between chats is significantly faster, conversations load with less wait when you jump around your sidebar.

Many agents can run much more efficiently at once, so concurrent jobs stay responsive under heavier load.

You can now log in with Discord or Google in addition to email OTP.

The whole app was debloated, leaner code and less overhead for a snappier experience.

Many bugs were fixed across chat and auth, including flaky Google/Discord sign-in, chats that failed to load with no error, sends that silently did nothing when the profile was missing, and agent resume leaking across account switches.

Superseded. Plans, prices and the free allowance changed on October 2, 11, 17 and 24, 2026; see Pricing for the current ones.

Pro is now $9.99/mo (annual $69.99/yr ≈ $5.83/mo, up to 42% off). Studio is now $19.99/mo (annual $139.99/yr ≈ $11.67/mo). Pro includes a $6.99/mo API budget (~4.7M monthly tokens).

Existing Pro and Studio subscribers were moved to the new prices automatically, no mid-cycle charge; the lower rate applies at your next renewal.

Starter quota rolls back to 100,000 tokens every 15 hours, the lifetime allowance is gone. The free countdown shows hours, minutes, and seconds; Pro and Studio still show days through seconds.

Pricing pages, upgrade CTAs, and schema markup now match the new Pro / Studio amounts and the annual “up to 42% off” badge.

Renamed the profile usage column to tokens_used so free and paid tracking share one generic counter name.

More agents can run at once, when the relay is busy, jobs wait in a queue instead of failing, then start automatically as capacity frees up.

Quota counting is more accurate for multi-step agents, every LLM turn is billed reliably, so usage matches what you actually used.

Starter lifetime quota set to 100,000 total tokens (down from 150,000). Usage is tracked against your lifetime allowance, no reset window.

Fixed a bug where quota would not count properly, usage is now tracked accurately against your lifetime allowance.

Many changes to the user interface were made across the app, cleaner layout, better readability, and a more polished overall experience.

Various bugs were fixed across the site and app.

Starter lifetime quota set to 150,000 total tokens.

Removed hierarchy/codebase syncing and relay-side indexing. All Studio inspection tools now run live in the plugin, lower RAM use on relay and Studio, faster agent job start, always-fresh reads.

RoCode plugin v3.4, no background sync crawls, dirty watchers, or periodic hierarchy uploads. Plugin status stays in the Configuration panel.

Removed the Studio Explorer panel. Use Configuration → Plugin Status for online/offline.

Fixed a bug where your messages disappeared when reopening an old chat, only the agent's replies would load. User messages are now saved reliably and show up again when you return to a conversation.

Improved message indexing so the agent has an accurate record of the full thread, what was sent, what was saved, and what was not, instead of working from a partial or out-of-order history.

Added hide buttons to the Conversations, Studio Explorer, and Configuration panels so you can tuck away panels you are not using and get more room for chat.

Added resize handles to the Conversations, Studio Explorer, and Configuration panels, drag the edge to set the width you want.

Panel hide and resize preferences are saved automatically, so your layout is restored the next time you open chat.

Added check_script, a new syntax tool for the agent. After writing or editing scripts, it validates Luau compile errors, delimiter balance, require paths, GetService names, deprecated APIs, and Output-window errors before finishing the task.

Fixed several OTP login bugs, rate-limit cooldowns now release correctly, resend works with Turnstile, double-submit on verify is blocked, and invalid or expired codes no longer spam error logs.

Removed many redundant tools that overlapped with cleaner alternatives, for a faster agent.

Stripped out lots of legacy code paths that were no longer in use, reducing complexity across the backend and plugin.

The agent is faster: less overhead from unused tools and dead code means quicker responses and shorter wait times.

Fixed several bugs uncovered during the cleanup for a more reliable experience overall.

Free-user quota now counts properly, usage is tracked accurately against your lifetime allowance.

Search tools were fixed and improved for more reliable results when the agent looks up docs and the DevForum.

The agent is much faster: it now wakes up and starts instantly instead of waiting on polling delays.

Renamed config options for clarity: “Provide code prints()” → “Allow code prints()” and “Provide code comments” → “Allow code comments”.

Starter quota is now 200,000 total lifetime tokens, no 15-hour reset. Use them across as many sessions as you need until they’re gone, then upgrade to Pro for monthly usage.

The sidebar meter for free users now shows usage as a percentage only, no token counts below the bar. Pro and Studio plans still reset on their monthly billing cycle.

Fixed several bugs across billing, subscriptions, and the chat experience for a more reliable upgrade and usage flow.

Updated plan pricing across the site and checkout:

• Note

  Superseded. Plans and prices changed on October 2 and October 11, 2026; see Pricing for the current ones.

• Pro monthly: $9.99/mo → $14.99/mo
• Pro annual: $59.99/yr ($5.00/mo) → $119.88/yr ($9.99/mo)
• Studio annual: $179.99/yr ($15.00/mo) → $239.88/yr ($19.99/mo)

Studio monthly stays at $24.99/mo.

All existing Pro subscribers were automatically moved to the new Pro pricing, no action needed on your end.

docsearch is now roblox_search, one tool that searches official docs, DevForum, and GitHub. Results show source badges in chat so you know where each hit came from.

Read-only tools run in parallel for faster inspections. Polling speeds up to 100ms while a job is active and relaxes to 1s when idle. Large script payloads can use base64 encoding. The Sync button is now Reset, reloads the plugin to recover from a stuck state.

Streamlined system prompts with clearer safety rules, mistake handling, and an action-before-questions protocol so the agent inspects your game instead of asking you first. Tool definitions moved to a manifest with auto-generated schemas.

Delete, remove tag, and clear terrain actions now show lime-green bubbles so removals are visually distinct from creates and edits.

Agent jobs request a screen wake lock and show a banner if you switch tabs mid-run. Polling catches up immediately when you return. Plugin status checks every 15s when online and every 2s when offline.

Daily free quota doubled from 50,000 to 100,000 tokens per 15-hour window.

Removed the manual hierarchy refresh button from the Studio Explorer, hierarchy stays in sync automatically through the relay.

Old chat mode is gone. Everyone now uses the new plugin mode.

Refreshed the interface with a cleaner layout and better visuals.

The agent is faster with streamlined responses and less overhead.

New tools reads Studio output so the AI can see errors and fix them.

Fixed duplicate messages broken code text and agents stopping too early.

Attach paths from the explorer with chips in chat and the input box.

Daily limit banner and tool bubbles look cleaner and work better.

Added a button to record your voice and transcribe it into the chat box.

Lots of bugs fixed across the app.

The AI now actively explores your 3D workspace, reading properties and navigating descendants dynamically before writing code.

Upgraded from 11 basic tools to 55 full-spectrum actions to manipulate Physics, Lighting, Terrain, Attributes, and Camera properties natively.

The AI captures its own print() and warn() outputs to test its own code and read terminal errors instantly.

Full ChangeHistoryService integration lets you safely hit Ctrl+Z to undo massive AI operations in Studio.

Removed the 2-second short polling loop in favor of Long Polling for instantaneous, lag-free command execution.

Added smooth cinematic animations and clean, glowing teal bubbles for workspace query tasks in chat.

RoCode Plugin v2: A complete rewrite of the Studio plugin from the ground up. It is now much more reliable, faster, and stable.

Chat Titles: Say goodbye to generic chat names! Conversations now automatically generate nice, descriptive titles.

New Prompt System: Implemented a brand new system for handling prompts behind the scenes to give you better results.

Model Upgrades: All AI models have received a backend update to improve speed and code quality.

Configuration Cleanup: We've streamlined the Configuration panel by removing outdated and unnecessary options.

The usage meter now shows Quota as a percentage for both free and paid users. Cleaner and easier to understand at a glance.

The app has a brand new user interface.

Added new AI models to improve code generation.

Many bugs were fixed to improve overall stability.

A big update to the RoCode plugin is coming soon!

Upgraded to the latest AI models for significantly better code quality, faster responses, and improved understanding of Roblox Luau syntax.

Switched to OpenRouter for more reliable API access, better rate limiting, and access to a wider range of models.

The native Roblox Studio plugin is finally here. You can now use RoCode without leaving your script editor. No more alt tabbing to a browser.

Added hierarchy awareness. The AI now sees your script names and where they are located in the explorer to give much better answers.

One click apply functionality. Insert or fix code directly into your active script with a single button press.

Trained specifically on the latest Roblox APIs. Unlike generic AI tools like Gemini it knows the difference between client and server logic and uses modern task library methods.

Free users can now send up to 15 messages a day within the plugin to help with their projects.

Historical. Since then you describe changes in the web chat at rocode.app and the plugin makes them in Studio. RoCode runs third-party models through OpenRouter with its own Roblox instructions and tools, and has no model of its own. For the free allowance today see Pricing.

Fixed a critical bug where upgrading to Pro sometimes failed to attach the user ID. The payment pipeline is now rock solid.

Optimized the background canvas rendering to prevent lag in session replays. The site should feel smoother on lower end devices.

Improved the Upgrade to Pro button logic to prevent crashes when loading user profiles.

Roblox Studio Plugin: I am finalizing a native plugin to bring RoCode directly into Studio. Coming very soon

Significantly reduced Model Overloaded errors. I implemented a smart retry system to handle high traffic automatically.

Enhanced backend security protocols and prepared infrastructure for plugin connectivity.

Visual refresh: Adjusted transparency for click particles and starfield bloom to improve visual clarity and reduce distraction.

Initial release of RoCode, an AI powered Roblox coding assistant.

Real time chat interface for generating and explaining Luau code.

Dedicated code viewer panel with copy and paste functionality.

Configuration panel to customize AI responses for example adding comments or focusing on performance.

Secure user authentication and conversation history powered by Supabase.