Tidewave Web: in-browser coding agent for Rails and Phoenix
tidewave.ai239 points by kieloo 10 hours ago
239 points by kieloo 10 hours ago
Hi folks! Great to see our work on the front page, and I'm happy to answer questions.
Regarding next steps, we're currently working on React integration, with Python and JavaScript server-side web frameworks coming next. If you're interested, here's our survey/waiting list: https://forms.gle/8MeXwGjpBFDeGNQw9
Very cool, we use it as MCP server when we develop things using Claude Code... how would using Tidewave Web change our workflow?
You can think of Tidewave MCP as integrating the agent with the language runtime, while Tidewave Web integrates the agent with your actual app and its interface. It knows your views and templates, it can correlate it with the rendered page, and then automate the browser to implement and validate features.
Our announcement post goes more into detail and explains why moving the agent to the browser means you (the developer) can spend more time out of the loop: https://tidewave.ai/blog/tidewave-web-phoenix-rails
As someone who builds AI products and having used agentic coding tools since they came out (often with Rails projects), I don't get this. There was a similar project called Rails MCP Server which said:
> "This Rails MCP Server implements the MCP specification to give AI models access to Rails projects for code analysis, exploration, and assistance."
And again I don't get the value. I can see some slight benefits to having a tight browser integration but I don't think that's worth leaving the IDE / CLI tools and the workflows they bring. You can also use Playwright MCP or just screenshot easily for more context. Claude Code can now run your server in the background and view logs as well. In a perfect world where LLM's can one shot whole features, maybe. But I can't let Claude Code go 10 minutes without it introducing a bad pattern and having to stop it. Reducing that visibility even further with this does not seem like a good combo.
I'm not wanting to tear down others projects either, just giving my perspective. I should try it to see how it does in the wild but the Copilot license or Anthropic API key requirement also deters me as well as having to have a project specific dependency.
As you said, the best would be for you to give it a try and see the benefits in practice (if any). Beyond that, I can add to two points you brought up:
First, Playwright MCP is really not enough. For example, imagine you are building a project management software and you are working on a feature for transferring tasks between projects. Testing this feature requires at least a user who is admin and at least two projects. When it is time for Playwright to test this feature, you need to automate the account creation, then it has to set up two projects, add the task, etc, and it often gets stuck.
When you are using Tidewave, you naturally do this setup as part of your development, which Tidewave has direct access to during tests. So it just interacts with the same page. Furthermore, if you see any bugs or places you could improve, you can't really tell Playwright or regular coding agent to fix them, you have to translate what you want. With Tidewave you just point and click and it works across browsers. I talk more about this in the announcement post: https://tidewave.ai/blog/tidewave-web-phoenix-rails
Other than that, I am really curious about the handoff between Tidewave and the editor that you mentioned. At some point, you will need the editor indeed, but how to best do this transaction? Should you be able to review and do tiny tweaks within Tidewave first? Regardless, moving the agent to the browser for web applications bring enough benefits to justify going one lever higher. :)
Playwright MCP has a mode where it can run as a Chrome Extension, which allows you to use it on your active browser session. Not sure if you can do point and click to communicate but it covers the development setup and login bit.
I've been using it for a few hours and I find that it struggles at times with some of the instructions, it completely obliterated a modal and overall I didn't felt like I trusted it too much with larger tasks, but then I started working with Claude Code on a console, and Tidewave on my browser and the workflow is pleasing, I only give UI portions of work to Tidewave and it works well, it's nice to be able to select elements and pass instructions for refactoring UI... this is only from like 2 or 3 hours of testing...
I spent a couple of hours last night testing it out.
The login process didn't work well, but it's due to my browser having a seperate container for Github.
I asked it to add a feature using Stimulus.js to my very simple gallery page written in Rails, when you hover over an image in the gallery, it should make an icon appear, and when clicking the icon, a modal should pop-up that allows the user to do cropping.
It picked up on my database models and it produced a very good result, however, it seemed to have issues verifying its solution, it particularly seemed to struggle with "on-hover".
The solution it produced was a lot better than what claude-code produced, and I liked the interface.
I also ended up hitting the limits on the Anthropic API, and it wasn't obvious to me what I should do in that case, but that's likely not the fault of Tidewave, and the Context Window Usage having a maximum of 200.0k also seemed very high, but that probably contributed to me hitting a limit.
Thank you `kawsper` for the feedback!
We are going to introduce more visibility into the context limit and make it easier to summarize.
Regarding the on-hover, do you still have the snippets around that it tried? We instruct the model to dispatch events using the proper browser APIs and I wonder what it tried to do. I would love to hear more! You can ping me on Tidewave's Discord (https://discord.gg/5GhK7E54yA) or send me an email (see github.com/josevalim).
I'm super-conflicted about this.
I've been loving Tidewave MCP since the day it launched and I wish I had that for all my work. I also fight the pain of driving web browsers through the playwright MCP every day, and while it's a huge help, it's also massively slow and kills my context window.
I want these problems solved. But I want them solved in a way that isn't unique to one stack, because I don't have the luxury of sticking to my preferred Elixir stack all the time. I also want them solved in a way that doesn't force me to take a huge backward step in my AI usage by bypassing my all-you-can-eat Max subscription.
The $10 subscription isn't that expensive, but then again, since this isn't a complete solution, it's going to be on top of other subs like Cursor/Max/Phoenix.new etc etc and it becomes death by a thousand cuts.
Maybe I just need to calm down a bit and not look at it as all-or-nothing.
Like, sure, switching to this full-time would be impossible. But maybe as a tool in my toolbox, where I pay $10 and use it once-in-a-while, when I have a tricky thing to do and it's worth me burning $10-20 on API tokens just to unblock me.
I'm just concerned that software development is fast becoming pay-to-play / pay-to-win, and that can quickly lead to chickenization, where prices are carefully titrated to shift most of the value to the tool providers.
We've seen this play out in infrastructure and tooling, and most orgs now "donate" a huge chunk of their profit margins to AWS.
> I'm just concerned that software development is fast becoming pay-to-play / pay-to-win
The people did it to themselves.
Instead of investing time learning these things, people (like always) want to take a shortcut. Big co's are more than willing to take advantage of that laziness and will indefinitely into the future—whether diet pills, AI, or a brain chip that occasionally makes you drool but is 100x more accurate with directions than Google Maps.
I hear you. I'd love to integrate with whatever model subscription is available but it seems using Max outside of Claude products is against their terms. I suggest reaching out to Anthropic and letting them know you would like to use your Max subscription with other coding agents.
I've been using the tidewave plugin for a bit now, this browser addition seems pretty cool. I'm curious about the pricing though. It says we need to provide our own github copilot or anthropic keys, but there's then a limit on the number on usage before you need to pay.
Is this because data is going through a tidewave server or something, or is it just a way to create a bit of a free trial vs "now you need to pay us"?
It is both. There is a Tidewave server (which will play a more meaningful role once we introduce agent coordination) but we also hope the limit will be enough for people to see value in the tool and convert into paying users (so we can continue improving it).
The details on what the architecture of this aren't clear, but I can see that the big feature is deeper integration into the browser.
To me this is an obvious direction for all coding agents to go in. Right now the cursor chrome browser MCP server doesn't work very well, but it's very obvious to me that this is the direction things need to go in.
I'm not sure why focusing on a framework would fundamentally make this any better than the generic approach to getting good context directly from the browser.
There are two main reasons why being framework specific can be a big boost:
1. We annotate and understand your template language. So when you select an element, we not only know its tag and attributes, but we also know which template file rendered it. The more structured the template language is (JSX, HEEx, etc), the better job we can do, which means LLMs don't have to guess.
2. We also run code within the web framework. This means database access, built-in documentation using the exact versions in your project, access to language and framework reflection/introspection APIs, etc. So the agents have more, err, agency.
I hope this clarifies it a bit!
When using browser-tools-mcp I can point to an element and ask Agent to "change color of this element to red". It would copy the part of the HTML and search the codebase for it.
I can imagine having an in browser, framework level tool would know exactly which controller and which template generated this element and could target it more precisely while using less tool calls and less tokens. I haven't tested it yet but that is my expectation.
Looks cool, I'll wait for some sort of local LLM support eg ollama before buying it
Hey Jose, How does this differ from Phoenix.new?
Phoenix.new is about remote agents, while Tidewave integrates with your app running on your machine!
Cool. That's actually interesting. I found with Phoenix.new that I would start there, but later ended up running everything locally - with the Tidewave MCP server of course!
Why does it forward to https://tidewave.ai/ when I type http://localhost:4000/tidewave ?
You are first redirected for authentication purposes. Once authenticated, it redirects back to your app.
I would love to have a React Copilot that has access to the console, network logs, the actual html elements, computed styling etc. + my code.
This would be such a game-changer. I am also convinced that monorepos will become the de facto standard, since it is way easier for LLMs to navigate / implement features across the whole stack.
We are working on React integration, first within Rails and Phoenix, but then also standalone. How are you serving/running your React apps in development?
Is it safe to say that the Tidewave MCP server will remain FOSS, but anything Tidewave Web related (localhost:4000/tidewave) will be proprietary yet optional?
This looks promising. I will definitely be giving it a shot. I have been getting pretty good results with Claude Code and Rails for a while now, I'm kind of excited to see how much things can improve with this.
Looks pretty neat, and certainly addresses a missing element in the current AI workflow.
Question: What happens to our data - i.e. the code and context sent to your service?
We log basic request metadata (timestamps, model used, token counts). Prompts and messages are not logged unless you explicitly opt-in. We don't store tool results. Note the underlying model provider you use may store data separately depending on your user agreement with them.
Hard for my team to get excited about this, because we haven't been able to use Tidewave at all, as neither it nor the Rust-based proxy support local HTTPS (which we need for our integrations with several third-party services, e.g. oauth providers).
Can you ping me on Tidewave's Discord or send me an email (you can find on github.com/josevalim)? We did test it on https recently and it worked fine, so there is something more going on. I'd be glad to see this ironed out!
Look, competing against Claude Code will be hard.
yet this is Jose Valim we are talking about. creator of Elixir, Plug, contributor in Phoenix and Rails
Guy is a machine https://github.com/josevalim
I believe he will deliver, will check it time to time, interesting approach
> Guy is a machine https://github.com/josevalim
Indeed. He is a big inspiration for Brazilian nerds.
Tidle >>>
<<<< Wave
I think opinionated frameworks like this are a good fit for ai.
Usually only one correct and idiomatic way to do things, and rails in particular really leans on convention over configuration.
Opinionated frameworks, strict types, and whatever non-ai tools for correctness validations (typescript, linters, compilers) are all helpful.
Hack