Even Graphs - probably are more useful for the translation between the agent and humans. I mean - more useful for understanding the codebase, understanding patterns, explain the codebase - those sort of features. I am not yet convinced that graph databases can make the code generation to be more accurate, authentic or optimal. Finally, what do we look for in the generated code? Same expectations as from human - the code should reuse already tested code as much as possible. It should be idiomatic and maintainable. It should be tested. Finally, it should consume less tokens. 
For any of these goals - will a graph database help? in which way? I am not yet clear. On the otherhand, we can already see that these dynamic intervention with something like Tidewave is leading to some real benefits with development. If not in generation - atleast in debugging. Maybe it should be enhanced more by integrating something like PlayWright MCP. But, beyond that probably I am not sure if we need more interventions.
This GitHub - sysread/fnord: Fnord is an AI code archaeologist. is an elixir project I followed with great interest. But it moved more towards humans than towards agents. So, what we need now is more of a tool that is tilted towards agents rather than us - in two facets - static and dynamic - both only in the generation side. We can debug and fix things. But, generation is what we have to make it more intelligent.
1 Like
We just announced the next iteration of Tidewave, Tidewave Web. I have opened up a new thread here: Tidewave Web: in-browser coding agent for Phoenix (and LiveView)
1 Like
For Zed I think the correct config is:
{
"context_servers": {
"tidewave-mcp": {
"command": "mcp-proxy",
"args": ["--transport", "streamablehttp", "http://localhost:4000/tidewave/mcp"],
"env": {},
"enabled": true,
"source": "custom"
}
}
}
}
I’ve submitted a PR for the documentation update.
I take that back, I was using the old recommended mcp-proxy.
