How to give Codex access to VSCode Shift+Enter Julia eval?

Tooling VS Code
1

I’m currently using my own custom module that starts server and exposes socket to Codex Agent so it can call it and execute code in my VS Code julia session.

I wonder if there’s simpler way - given that Julia VS Code extension already does exactly that with Shift + Enter.

So, can AI Agent use Juila VS Code extension API instead of my custom server?

2

Are you looking for GitHub - kahliburke/Kaimon.jl: MCP server giving AI agents full access to Julia's runtime via a live Gate — code execution, introspection, debugging, testing, and semantic search · GitHub

3

Thanks, but julia is very slow to startup, so I would like to avoid loading any large package.

My current NetEval is just one page script and it works, but ideally I would like to remove it too and just use already existing VS Code without adding any extra dependency.

Maybe I’m missing something, I don’t know what advantages MCP server has over simple REPL access.

4

I created VS Code extension that patches standard Julia VS Code extension GitHub - al6x/julia-bridge: Gives AI Agent access to Julia VS Code julia terminal · GitHub

Ideally something like that should be part of standard Julia VS Code extension.

Ideas to improve/simplify welcomed.

P.S.

Also here’s AI instructions how to use it

Don't run Julia unless explicitly asked, and never start a fresh process. Use static inspection by default. When asked,
run `/projects/julia-bridge/client.js --timeout 120 'code'` from the project root. It uses the existing VS Code REPL,
returns JSON, and interrupts Julia on timeout or Ctrl-C.


For nontrivial work, replace the body of `AgentEval.run()` in `./AgentEval.jl`, then evaluate
`import AgentEval; agent_result = Base.invokelatest(AgentEval.run)`. Revise reloads it before each request. After
changing `const` or `struct` definitions, explicitly run `Revise.revise(AgentEval)`. Return compact values, preferably
named tuples, and use files for large output. Use plain code and
don't use `eval` to monkey-patch package methods.
5

There is also GitHub - marius311/julia-vscode-unofficial-mcp: Julia VSCode REPL for agents · GitHub by @marius311, which exposes the Julia REPL in VSCode to an AI agent through MCP.

See also the thread in Letting AI agents use the VSCode Julia REPL.

6

Running KaimonGate.serve() in the VSCode REPL is very quick, taking less than 0.5 seconds before the connection with the Kaimon server is established. The Kaimon server itself, installed at ~/.julia/bin/kaimon, takes longer to start, but you only need to do that once per boot and keep it running.

There’s also an option that appears more lightweight than Kaimon, though I haven’t tried: