Agent examples over AG-UI: LangGraph, Pydantic AI, Google ADK

.github/dependabot.yml

@@ -16,6 +16,16 @@ updates:
       dependencies:
         patterns: ['*']
         exclude-patterns: ['@aparte/*']
+  - package-ecosystem: pip
+    directories: ['/langgraph-ag-ui', '/pydantic-ai-ag-ui', '/google-adk-ag-ui']
+    schedule:
+      interval: weekly
+    ignore:
+      - dependency-name: '*'
+        update-types: ['version-update:semver-major']
+    groups:
+      dependencies:
+        patterns: ['*']
   - package-ecosystem: github-actions
     directory: /
     schedule:

.github/workflows/ci.yml

@@ -15,10 +15,13 @@ jobs:
     runs-on: ubuntu-latest
     outputs:
       examples: ${{ steps.find.outputs.examples }}
+      agents: ${{ steps.find.outputs.agents }}
     steps:
       - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
       - id: find
-        run: echo "examples=$(ls -d */package.json | xargs -n1 dirname | jq -R . | jq -sc .)" >> "$GITHUB_OUTPUT"
+        run: |
+          echo "examples=$(ls -d */package.json | xargs -n1 dirname | jq -R . | jq -sc .)" >> "$GITHUB_OUTPUT"
+          echo "agents=$(ls -d */requirements.txt | xargs -n1 dirname | jq -R . | jq -sc .)" >> "$GITHUB_OUTPUT"
 
   example:
     needs: list
@@ -42,11 +45,33 @@ jobs:
       - run: npm run check
       - run: npm run build
 
+  # Every folder with a requirements.txt holds a Python agent: it is installed and run, with its scripted model.
+  agent:
+    needs: list
+    name: ${{ matrix.example }} (agent)
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        example: ${{ fromJson(needs.list.outputs.agents) }}
+    defaults:
+      run:
+        working-directory: ${{ matrix.example }}
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+      - uses: actions/setup-python@6d4fd22b20465ccc0f7354476588e8df7c9ca0e0 # v7.1.0
+        with:
+          python-version: '3.13'
+          cache: pip
+          cache-dependency-path: ${{ matrix.example }}/requirements.txt
+      - run: pip install -r requirements.txt
+      - run: python smoke_test.py
+
   # The one check main requires: it passes when every example does, however many there are.
   all:
     name: all examples
     if: always()
-    needs: [list, example]
+    needs: [list, example, agent]
     runs-on: ubuntu-latest
     steps:
-      - run: test "${{ needs.list.result }}" = success && test "${{ needs.example.result }}" = success
+      - run: test "${{ needs.list.result }}" = success && test "${{ needs.example.result }}" = success && test "${{ needs.agent.result }}" = success

.gitignore

@@ -1,4 +1,7 @@
 node_modules
+dist
+.venv
+__pycache__
 .svelte-kit
 build
 .output

README.md

@@ -1,10 +1,13 @@
 # aparté examples
 
-Apps built on [aparté](https://github.com/apartejs/aparte), an MIT chat UI for AI made of Web Components. Each folder is a whole app: open it in StackBlitz, or copy it and run it.
+Apps built on [aparté](https://github.com/apartejs/aparte), an MIT chat UI for AI made of Web Components. Each folder is a whole app: copy it and run it, or open it in StackBlitz when it runs there.
 
 | Example | Stack | What it shows | Try it |
 |---|---|---|---|
 | [sveltekit-ai-sdk](./sveltekit-ai-sdk) | SvelteKit, Vercel AI SDK | A tool the model can only run once the user approves it, drawn by a Svelte component; Markdown formatted as it streams | [StackBlitz](https://stackblitz.com/github/apartejs/examples/tree/main/sveltekit-ai-sdk?file=src/lib/chat.ts) |
+| [langgraph-ag-ui](./langgraph-ag-ui) | LangGraph, AG-UI | A Python agent streaming into the chat, and a tool that waits for a person through `interrupt()` | On your machine |
+| [pydantic-ai-ag-ui](./pydantic-ai-ag-ui) | Pydantic AI, AG-UI | A Python agent streaming into the chat, and a tool that waits for a person through `requires_approval` | On your machine |
+| [google-adk-ag-ui](./google-adk-ag-ui) | Google ADK, AG-UI | A Python agent streaming into the chat, a tool that waits for a person through `require_confirmation`, and ADK's confirmation tool kept off the transcript | On your machine |
 
 ## Run one on your machine
 
@@ -15,13 +18,13 @@ npm install
 npm run dev
 ```
 
-With no API key, an example answers with a scripted model, which is how it runs on StackBlitz. Each folder's `.env.example` names the key that switches it to a real model.
+With no API key, an example answers with a scripted model, which is how it runs on StackBlitz. Each folder's `.env.example` names the key that switches it to a real model. An example with a Python agent starts that agent first: its README has the commands.
 
 ## Adding an example
 
 - **One folder, one app.** Nothing is installed at the root, and no folder imports from another: a folder copied alone must run.
 - **Name it after its stack**, framework first: `sveltekit-ai-sdk`, `nextjs-ai-sdk`. Add the topic only when two examples share a stack.
 - **aparté comes from npm**, at a published version, with the lockfile committed: the example installs what a reader's app would.
 - **It runs without a key.** A scripted model answers when the key is missing, and its answer says it is scripted.
-- **It has `check` and `build` scripts.** CI runs `npm ci`, `npm run check` and `npm run build` in every folder that has a `package.json`, and Dependabot updates each folder's lockfile.
-- **Its README opens with a StackBlitz link**, and it gets a row in the table above.
+- **It has `check` and `build` scripts.** CI runs `npm ci`, `npm run check` and `npm run build` in every folder that has a `package.json`, and Dependabot updates each folder's lockfile. A folder with a `requirements.txt` holds a Python agent: CI installs it and runs its `smoke_test.py`.
+- **Its README says how to try it**, with a StackBlitz link when it runs there (a Python agent does not), and it gets a row in the table above.

google-adk-ag-ui/.env.example

@@ -0,0 +1,3 @@
+# Optional. Without MODEL, a scripted model answers (scripted_model.py).
+MODEL=
+GOOGLE_API_KEY=

google-adk-ag-ui/README.md

@@ -0,0 +1,34 @@
+# Google ADK + aparté
+
+A [Google ADK](https://google.github.io/adk-docs/) agent served over [AG-UI](https://docs.ag-ui.com), with [aparté](https://apartejs.dev) as its chat. Replies stream, tool calls show as they run, and one tool waits for a person before it runs.
+
+## Run it
+
+The agent runs in Python, so this one runs on your machine and not on StackBlitz. Tested with Python 3.13 and Node 24.
+
+```bash
+npx degit apartejs/examples/google-adk-ag-ui my-agent
+cd my-agent
+python -m venv .venv
+source .venv/bin/activate   # on Windows: .venv\Scripts\activate
+pip install -r requirements.txt
+uvicorn agent:app --port 8000
+```
+
+In a second terminal:
+
+```bash
+npm install
+npm run dev
+```
+
+Open http://localhost:5173 and ask it to send an email.
+
+With nothing configured, a scripted model answers, and says it is scripted. For a real model, copy `.env.example` to `.env` and set `MODEL` to a Gemini model and `GOOGLE_API_KEY` to your key.
+
+## What to look at
+
+- [`agent.py`](./agent.py): the agent and its FastAPI route. `require_confirmation=True` on the tool pauses the run for a person, and `emit_interrupt_outcome=True` ends that run the way AG-UI 1.0 describes it.
+- [`src/main.ts`](./src/main.ts): the whole page. ADK asks for a confirmation through a tool of its own, `adk_request_confirmation`, so `registerToolDisplay` hides that row and has the question name the real tool. The rest knows nothing about ADK: `@aparte/runner-ag-ui` reads the agent's AG-UI stream.
+- [`scripted_model.py`](./scripted_model.py): the stand-in model.
+- [`smoke_test.py`](./smoke_test.py): what CI runs against the route: a reply, a pause, and a resume.

google-adk-ag-ui/agent.py

@@ -0,0 +1,41 @@
+"""A Google ADK agent served over AG-UI, with a tool that waits for a person."""
+
+import os
+
+from ag_ui_adk import ADKAgent, add_adk_fastapi_endpoint
+from dotenv import load_dotenv
+from fastapi import FastAPI
+from google.adk.agents import LlmAgent
+from google.adk.tools import FunctionTool
+
+load_dotenv()
+
+
+def send_email(to: str, subject: str) -> str:
+    """Send an email. A person has to approve it first."""
+    return f"Email sent to {to}: {subject}"
+
+
+def make_model():
+    if os.environ.get("MODEL"):
+        return os.environ["MODEL"]
+
+    from scripted_model import ScriptedModel
+
+    return ScriptedModel()
+
+
+agent = LlmAgent(
+    name="assistant",
+    model=make_model(),
+    instruction="You help the user. Send an email only when asked to.",
+    tools=[FunctionTool(send_email, require_confirmation=True)],
+)
+
+app = FastAPI()
+add_adk_fastapi_endpoint(
+    app,
+    # emit_interrupt_outcome: a pause ends the run the way AG-UI 1.0 describes it, which is what the chat reads.
+    ADKAgent(adk_agent=agent, app_name="example", user_id="reader", emit_interrupt_outcome=True),
+    path="/agent",
+)

google-adk-ag-ui/index.html

@@ -0,0 +1,15 @@
+<!doctype html>
+<html lang="en">
+  <head>
+    <meta charset="utf-8" />
+    <meta name="viewport" content="width=device-width, initial-scale=1" />
+    <title>Google ADK + aparté</title>
+    <style>
+      body { margin: 0; }
+    </style>
+  </head>
+  <body>
+    <aparte-chat style="height: 100dvh"></aparte-chat>
+    <script type="module" src="/src/main.ts"></script>
+  </body>
+</html>