Set Up Python Tool Execution

Add client-side Pyodide execution for $ARGUMENTS. Skip this if the app uses the Atlas / EOS sidebar (integrate-fusion-agent).

Prerequisite: src/atlas-agent/ from integrate-atlas-chat, plus @sinclair/typebox. Copy python.ts, pyodide.ts, pyodide-react.ts, pyodide-runtime.ts from integrate-atlas-chat/code/ into src/atlas-agent/.

CDF runPythonCode tools arrive as toolConfirmation + clientTool. Wire usePyodideRuntime and pass pythonRuntime to useAtlasChat — no PythonToolConfig entries. Pyodide is ~30MB, cached after first load.

Step 1 — Understand the app

Read these files before touching anything:

• package.json — detect package manager and existing deps
• The component that calls useAtlasChat — understand current tools/config

Step 2 — Install Pyodide

Install exactly [email protected] using the app's package manager. This version must match the CDN artifacts loaded at runtime — installing a different version will cause errors.

• pnpm → pnpm add [email protected]
• npm → npm install [email protected]
• yarn → yarn add [email protected]

@sinclair/typebox should already be installed from integrate-atlas-chat. Add it if missing.

Step 3 — Set up usePyodideRuntime

In the component that calls useAtlasChat, add the Pyodide runtime hook:

import { loadPyodide } from "pyodide";
import { usePyodideRuntime } from "./atlas-agent/pyodide-react";
import { useAtlasChat } from "./atlas-agent/react";

function MyChat() {
  const { sdk, isLoading } = useDune();

  // Initialize Python runtime (loads Pyodide, installs packages, sets up Cognite SDK)
  const {
    runtime: pythonRuntime,
    loading: pythonLoading,
    progress: pythonProgress,
    error: pythonError,
    isReady: pythonReady,
  } = usePyodideRuntime({
    loadPyodide,
    client: isLoading ? null : sdk,
    requirements: ["pandas", "numpy"],    // optional — additional packages
  });

  // ... useAtlasChat below
}

Hook API reference

Loading state UI

Place the loading indicator above the chat input, not in the message list. Keep it compact — a pill/badge showing stage text and percent. Show an error badge separately. First load is ~30-60s (downloads ~30MB); subsequent loads are <2s from browser cache.

{/* Loading — shown above the input while Pyodide initializes */}
{pythonLoading && (
  <div className="flex items-center gap-2 rounded-lg border bg-muted/50 px-3 py-2 text-sm text-muted-foreground">
    {/* Optional: <IconBrandPython /> from @tabler/icons-react */}
    <span>{pythonProgress.stage || "Initializing Python..."}</span>
    {pythonProgress.percent > 0 && pythonProgress.percent < 100 && (
      <span className="text-xs opacity-70">({pythonProgress.percent}%)</span>
    )}
  </div>
)}

{/* Error — shown if init fails (after loading finishes) */}
{pythonError && !pythonLoading && (
  <div className="flex items-center gap-2 rounded-lg border border-destructive/30 bg-destructive/10 px-3 py-2 text-sm text-destructive">
    <span>Python runtime failed to load</span>
  </div>
)}

Step 4 — Wire into useAtlasChat

Pass the runtime to useAtlasChat. That's all — no tool configs needed:

const { messages, send, isStreaming, progress, error, reset, abort } = useAtlasChat({
  client: isLoading ? null : sdk,
  agentExternalId: "my-agent",
  tools: [renderTimeSeries],   // regular client tools (declared to agent), if any
  pythonRuntime,               // from usePyodideRuntime — enables Python tool execution
});

Note: Python tools are NOT declared to the agent via tools. The agent already knows about them from its CDF config. The library fetches the code automatically when needed.

Step 5 — Disable input while Python loads

The user shouldn't send messages before the runtime is ready. Disable the entire input area (not just the send button) so the state is unambiguous:

<ChatInput
  onSend={handleSend}
  disabled={isStreaming || pythonLoading}
  // ...
/>

If you have a home page with suggestion chips, disable those too:

<ChatHomePage
  onSuggestionClick={handleSuggestionClick}
  disabled={pythonLoading}
/>

Done

The app can now execute Python tools client-side via Pyodide. When the agent calls a Python tool, the library automatically fetches its code from the agent config, runs it in the browser, and returns the result to the agent.