A few weeks ago, we introduced Tinybird Code, our agentic CLI inspired by Claude Code but specialized for real-time data engineering with ClickHouse. Since then, we've been listening to your feedback and shipping improvements that make Tinybird Code even more powerful and reliable.
Why we kept improving it
When we first launched Tinybird Code, we knew we had built something useful, but we also knew there was room for improvement. The initial version was good at handling Tinybird-specific tasks, but users wanted more flexibility and reliability. We heard three main requests:
More kinds of files: Tinybird Code can now read markdown instructions or analyze other code in your project. Many of you asked about expanding Tinybird Code's ability to build context beyond just
.datasourceand.pipefiles so that you could better guide and integrate it into your full development workflows.Better plan execution: The step-by-step approval process was too slow for complex multi-file operations.
Fewer hallucinations and bug fixes: The agent sometimes made up commands or tried to run things that didn't exist
So we got to work.
What's new
Here are the key improvements we've shipped based on your feedback:
1. Universal file access
Tinybird Code can now read any file in your project, not just Tinybird-specific files. This means you can:
- Read, write, and update rules for your project or current plan.
- Work with your entire application stack
- Create API documentation
- Write configuration files
- Build deployment scripts
- Generate client SDKs in multiple languages
We've seen users create complete SDKs, generate API clients, and even build entire microservices using Tinybird Code. The agent now understands your full project context, not just the data layer.
Tinybird Code cannot directly write non Tinybird related files, but it is as easy as copy pasting the suggestion.
2. YOLO mode for faster plan execution
Sometimes you trust the agent enough to execute a complex plan without clicking "enter" after every step. That's what YOLO mode is for.
When Tinybird Code presents you with a plan, you'll now see a new option: "Yes, continue and implement all". This lets the agent execute the entire plan automatically without stopping for approval at each step.
This is perfect for:
- Creating multiple related datasources and pipes
- Setting up entire data pipelines
- Generating comprehensive test suites
The agent will still validate each step and show you what it's doing, but it won't wait for your approval between steps.
tb (ai_telemetry) >> add a visits by device endpoint and delete blog_analytics endpoint
Plan description:
Add visits by device endpoint and remove the existing blog analytics endpoint.
Steps:
1. Endpoint: visits_by_device - Endpoint to analyze blog visits grouped by device type with filtering capabilities
2. Delete file: endpoints/blog_analytics.pipe - Replace with more specific device-focused analytics
Do you want to continue with the plan? (Use arrow keys)
Yes, continue
> Yes, continue and implement all
No, tell Tinybird Code what to do
Cancel
3. Better error handling and reliability
We fixed several issues that were causing frustration:
- After cancelling a command the agent tried to run it again
- Command hallucinations fixes
- Fix incorrect date formats in generated example requests
The agent now handles interruptions gracefully, generates more accurate commands, and produces properly formatted example requests. These might seem like small fixes, but they make a huge difference in day-to-day usage.
4. Cleaner project organization
We improved how Tinybird Code organizes and displays your project structure, making it easier to focus on your actual data sources without getting distracted by system-level resources.
5. Routing all AI commands through the agent
Previous existing AI commands like tb create, tb mock, tb test create... now use the Tinybird Code agent under the hood.
The result: A more capable agent
These improvements have made Tinybird Code significantly more powerful. Users are now:
- Building complete applications: Not just data pipelines, but full-stack applications with APIs, clients, and documentation
- Working faster: YOLO mode eliminates the friction of step-by-step approvals for complex operations
- Trusting the agent more: Better error handling and fewer hallucinations mean users can rely on the agent for more critical tasks
Pro tip: Use Tinybird Code with Claude Code
If you're still using Claude Code for general development but want Tinybird Code's specialized data engineering capabilities, you can integrate them. Add this to your Claude Code configuration:
# When working with Tinybird follow these rules when using Tinybird CLI:
- Use tb -p="my_prompt_here" when interacting with Tinybird in any way.
- Use `tb --cloud deploy` to deploy the changes to production.
- Use `tb info` to get the tokens for Tinybird cloud (production) and Tinybird local (dev) environments.
When using tb -p be as descriptive as possible so the agent can understand the context of the command.
Claude Code can delegate tasks to Tinybird Code using one-shot mode, giving you the best of both worlds: Claude Code's general programming capabilities and Tinybird Code's specialized data engineering expertise.
Try the new features
Ready to see what Tinybird Code can do now? Update to the latest version and try:
curl https://tinybird.co | sh
tb
Or if you want to see the new file access capabilities in action:
tb --prompt "Generate a Rust SDK for my Tinybird project with proper error handling and type hints"
The agent will now read your entire project structure, understand your data sources and pipes, and generate a complete, production-ready SDK.
We're excited to see what you'll build with these new capabilities. Let us know what you think!
