Release notes

​

Highlights

• File attachments in Clark Knowledge. Upload documents like design guideline PDFs, CSS stylesheets, internal docs, or security guidelines; Clark uses them to build out knowledge optimized for his own consumption when building. No more copying and pasting content in manually.
• Start building from superblocks.com. Type a prompt directly on the marketing site and land in a working app. You can start building even before you log in.

​

Bug fixes

• Fixed a bug where Clark could time out when testing APIs
• Improved Clark API generation accuracy by fixing an SDK documentation regression — Clark now generates working APIs more reliably on the first try, with fewer correction loops
• Fixed UI flickering that occurred when Clark reads system skills during prompting
• Fixed certain read-only APIs not correctly triggering approval prompts when Clark attempts to test them
• Fixed cleared Clark chats reappearing after a checkpoint refresh
• Fixed queued prompts causing horizontal scrolling in the AI chat sidebar
• Fixed an issue causing AI model errors for some users
• Fixed a handoff bug that could leave apps uneditable when switching between users

​

Highlights

• Clark is now available to everyone. Clark AI is out of private beta and open to all. New users can sign up for free at superblocks.com, start a free trial, and upgrade to the Teams plan entirely self-serve. No more waitlist.
• New Snowflake integrations: Cortex AI and Postgres. Superblocks now connects natively to both Snowflake Cortex AI and Snowflake Postgres. If your team is in the Snowflake ecosystem, you can bring AI-powered queries and Postgres-compatible data sources into your apps seamlessly.
• A new homepage for building. The new Clark homepage puts a prompt box front and center. The first thing you see when you log in is the Clark prompt box and your most recently edited apps, so you can start building immediately.

​

Onboarding

• New onboarding survey flow helps tailor the experience when you first sign up and start using Superblocks
• We added a banner to the apps list to help existing Low Code users easily navigate between Clark AI and their legacy apps and workflows

​

Clark improvements

• The Org Knowledge editing experience now includes an accept/reject diff workflow. You get an explicit review step before any changes are saved, and the Knowledge panel refreshes immediately after.

​

Bug fixes

• AI/Code Mode requests no longer fail when app screenshots are very large. They’re now scaled down automatically.
• Fixed a bug where saving an app knowledge fact could put the app into a broken state
• Fixed a bug where Clark would stop responding after you replied to a multiple-choice question about app knowledge
• Fixed a routing issue where Clark’s app view would be stuck in the center of the screen when routes were defined in a variable
• Prompts containing UUIDs are no longer incorrectly blocked by chat guardrails
• Downloaded app file names now reflect the actual app name
• Fixed the API panel getting stuck open with no way to exit in code mode
• Fixed a visual flicker and improved save UX in the Edit Data Tag modal
• Fixed a bug where the REST Base URL was silently discarded when saving an integration
• Fixed a package installation failure in cloud edit mode caused by npm being overridden by pnpm
• Default data tag selection is now pre-populated for a smoother setup experience

​

Highlights

• Clark in Organization Knowledge. You can now chat with Clark directly in the admin panel to set up your organization knowledge. Clark helps you articulate your org’s conventions, integration nuances, and personal preferences through conversation, making it faster and easier to get your knowledge dialed in. As a reminder, knowledge works at three levels that Clark follows by default:
• App Knowledge: Clark remembers context per app. You can now ask Clark to remember facts, architecture decisions, and context specific to an app — just prompt it during a session. Clark stores these as app-level skills and uses them in future sessions, so the more you build, the better Clark gets. No setup required — just tell Clark what to remember.
• Environment access controls. Admins can now configure which data tags are available in each Superblocks environment — edit, preview, and production — directly from the new Environments admin UI. This makes it easy to lock down production data so builders and Clark only work against the right environments.

​

What’s next

• Knowledge promotion. Soon, you’ll be able to save knowledge directly to your organization from chat while building apps. Today, you can save app knowledge during a build session, and org knowledge when chatting with Clark in the admin. Knowledge promotion bridges the gap, so insights you surface while building flow back to your org knowledge without leaving the app editor.
• Improved onboarding experience. We’re working on making it easier to get started with Superblocks, including a new home page, the ability to set up integrations by talking to Clark, and a more seamless commit and deploy flow.
• APIs as code. We’re moving to code-based APIs written entirely in TypeScript — a new format that Clark can generate more accurately and is easier for you and your agents to review and maintain.
• Source Control for Clark Apps. We’re adding the ability to sync your Clark app code to remote Git repositories for code review and automation.

​

Other improvements

• Web exploration links are now clickable in the results panel for easier access to referenced sources
• Open share modal immediately when you create an integration for easy access control management
• Restored manual JavaScript/Python/REST/GraphQL step creation in API editor
• Increased UI timeout for OPA healthchecks to reduce “unable to connect to on-premise agent” errors
• App debug downloads now use zip format for easy sharing

​

Bug fixes

• Fixed issue where Clark would hang indefinitely when testing APIs — the system now properly handles timeouts and error states
• Fixed bug where reference images uploaded with the first prompt were lost and couldn’t be included in plans
• Fixed bugs with API runs showing error exclamation points and green checkmarks on a single run

​

Highlights

• Queue prompts and steer Clark mid-generation. You can now send messages while Clark is working — queue up follow-up instructions as ideas come to you, or interrupt and redirect Clark mid-generation if it’s heading the wrong direction. No more waiting for Clark to finish before giving your next thought.
• Organization Knowledge beta is now open to all customers. We’ve opened up the Org Knowledge beta to everyone. You can define knowledge at three levels that Clark follows by default:
  • Organization — design system and brand guidelines, coding standards, team conventions, and other org-wide context
  • Integration — nuances about your data sources, schema quirks, naming conventions, query patterns, and other details that help Clark work with your APIs correctly
  • Personal — your individual preferences and how you’d like Clark to work with you

​

Clark improvements

• Screenshots now capture full pages and wait for content to fully load before analyzing, so Clark no longer reasons over incomplete UI states
• Improved context summarization to better track build completion — Clark no longer asks about next steps after successfully making proposed changes
• Added a warning when users enable too many Knowledge entries

​

Data exploration in Clark Chat

• Enhanced JSON view inside results tables for better readability
• Improved handling of truncated API outputs with clear messaging and partial result display

​

Bug fixes

• Fixed retry UI showing incorrect success messages when AI model requests were paused or failed
• Fixed success toasts incorrectly showing for failed API calls
• Clark no longer incorrectly requests permission to test read-only APIs
• Clearing context now properly clears in-memory data to prevent recurring inference errors in edge cases
• Fixed checkpoint restoration to avoid a rare race condition where users could not restore to past checkpoints

​

Highlights

• Clark is now powered by Claude Opus 4.6. Clark now runs on Anthropic’s Claude Opus 4.6 — the world’s best coding model. You should see improvements across the board in code generation quality, reasoning, and Clark’s ability to handle complex app-building tasks.
• Clark automatically optimizes for large datasets. Apps generated by Clark now automatically implement pagination, virtualization, and server-side aggregation when working with large datasets. This means fewer browser slowdowns and better performance out of the box — no extra configuration needed.
• Agent Memory is in Beta. We’re rolling out the first piece of Clark’s memory system: organization-level knowledge. Admins can now define facts about your org, integrations, and users that Clark follows by default — things like coding standards, integration quirks, and team conventions. Reach out for access!

​

Clark improvements

• Clark now handles API call failures gracefully — errors display in the UI instead of appearing to hang indefinitely
• You can now skip testing specific APIs and still complete generation when Clark encounters API test rejections
• Improved Clark’s ability to inspect historical API runs
• Fixed false positives in safety checks that were blocking legitimate internal security audits

​

Editor experience

• Added mobile view support on the editor canvas
• Fixed profile switcher being disabled when chat is in fullscreen mode
• Returning to the editor after idle time no longer shows a fullscreen connection error — it automatically reconnects
• Fixed chat sidebar opening before the app has fully loaded
• Fixed pages opening in new tabs getting stuck on a loading screen
• Fixed full page reloads when navigating between pages in deployed apps

​

API & integrations

• Added Salesforce OAuth 2.0 Token Exchange support for authentication via Login Identity Provider
• Added BigQuery Workforce Federation authentication
• Fixed API errors silently failing — Clark now properly throws errors when an API run fails
• Fixed Databricks SQL queries getting corrupted when manually editing code
• Improved handling of integrations with large OpenAPI specs — metadata search no longer hangs
• Fixed Python code with escape sequences getting corrupted

​

Highlights

• Clark can now explore data when planning without asking for permission. Instead of prompting for permission to read, Clark can explore your integrations automatically at your direction to figure out the right approach. Query results appear right in the chat so you can see exactly what Clark is working with, drill in, and guide it toward the right queries.
• Clark tests APIs faster. We cut API test latency significantly — previously this could take 30+ seconds. Things should feel noticeably snappier when Clark is building and validating APIs for you.
• New safety protections. Clark now detects and blocks malicious or unsafe prompts, and we’ve added protections to prevent system prompt extraction attempts. These safeguards run automatically in the background.
• New integration: Databricks Lakebase. You can now connect to Databricks Lakebase directly from Superblocks.

​

Clark improvements

• Clark can now offer up to 6 options for multiple choice questions
• Fetched links are now clickable in web exploration
• Improved retry experience when Clark falls back across models and inference providers
• Clark now knows it cannot navigate between pages to take screenshots, reducing unnecessary loops

​

Deployed apps

• Opening app pages in new tabs now uses the correct URLs
• Cleaner deployed app URLs (removed deployed_at timestamp)

​

Editor experience

• Fixed issue where typing “ms” quickly would accidentally trigger preview mode
• Keyboard shortcuts like Cmd+Click now work correctly in fullscreen mode

​

Highlights

• Clark now stays smart over long sessions. Previously, Clark would lose context as conversations grew. Now Clark maintains a working memory—automatically retaining what was accomplished, key decisions, and constraints discovered.
• API generation is significantly faster and more accurate. We rewrote core parts of how Clark understands and generates APIs—full support for control flow features, a completely rebuilt parameterized SQL system, and improved handling of API documentation. Fewer mistakes, faster results.
• Inference errors no longer block you. Intermittent Vertex/Bedrock errors are now handled gracefully—if you see one, just send another message and Clark will recover.
• Clark can search the web. Clark can now perform real-time web searches. This is especially useful for using third-party libraries and looking up documentation for integrations in real time.

​

New capabilities

• Embedded apps with two-way communication. Superblocks apps can now communicate with parent apps and vice versa. Ask Clark to set up the interface for properties and events.
• Windows local development. Powershell support—no WSL required.

​

Clark generation improvements

• Clark can now build file upload features (e.g., send a file to an AI model or upload to bucket storage)
• Clark can now build apps with data tag switching
• Clark can now modify REST API body content types reliably
• Clark is better at targeted UI edits thanks to improved context about UI component layout
• Clark tests APIs faster
• Clark no longer hits timeouts when testing APIs

​

API generation

• Added support for all control flow features
• Rewrote parameterized SQL system—Clark can now always write parameterized SQL accurately and safely queries when parameterization isn’t possible
• Improved handling of API documentation and integration OpenAPI specs
• Fixed API transformation so Clark always sees user modifications
• Fixed missing bodyType field in REST API type declarations that caused TypeScript errors
• Added missing integration class declarations for 8 data sources (S3, MongoDB, GraphQL, and others)
• Prevented Clark from reading redundant JavaScript files in API directories

​

Chat experience

• Improved performance and reduced flickering for chats with many messages
• Chat no longer appears stuck thinking—now distinguishes between model thinking and time between tool calls
• Keyboard shortcuts for multichoice responses
• Fixed newlines not rendering correctly in AI-generated summaries
• Removed duplicate reconnection messages

​

Plan mode

• Plan mode always presents a new plan when you give feedback
• Clark has better context about targeted UI components and can make targeted edits more accurately

​

Reliability

• Edit mode loads reliably even after bad code generations, allowing Clark to fix issues
• Fixed chat crashes when clicking Clark’s interactive buttons (Build it, multiple choice) too quickly
• Fixed chat getting stuck centered on screen after generation
• Deployed app links now work correctly when opened in new tabs
• Syntax error indicator in top bar now clears after Clark fixes the issue

​

Versions & checkpoints

• Restore to any checkpoint from the Versions panel
• Restore to past checkpoint if you pause a generation midway
• Reset now works from the history panel even when unpinned

​

Other improvements

• Clone existing apps from the homepage
• Homepage search clears when entering folders
• Step navigation preserves previous run outputs
• Run Superblocks APIs locally without disabling browser local network access checks
• Changed “Think…” to “Starting development server” for clearer status during startup