Workshop quick-start

August 12, 2026 · View on GitHub

This guide gets you from git clone to a working chat UI in under 30 minutes, with zero GCP credentials required.

Audience: anyone running this codebase locally for the first time — workshop attendees, university students using the platform as a template, or developers exploring the protocol stack (ADK, AG-UI, A2UI, MCP, MCP Apps, A2A) without standing up cloud infrastructure first.

TL;DR

git clone https://github.com/sunholo-data/ai-protocol-platform.git
cd platform
cp .env.example .env

# Edit .env and uncomment these two lines:
#   LOCAL_MODE=1
#   NEXT_PUBLIC_LOCAL_MODE=1

# In one terminal:
cd backend && make install && cd ..
make dev

# In another terminal:
cd frontend && npm install && npm run dev

Open http://localhost:3456 — you should see a yellow LOCAL_MODE banner across the top and a "Welcome to Aitana" page. Click a demo skill and send a message. The agent streams a reply via AG-UI from a local ADK runtime; no Firebase, no Firestore, no Vertex AI.

What you get in LOCAL_MODE

Three demo skills, seeded into the in-memory Firestore at startup:

SkillWorkshop moduleWhat it demonstrates
Demo ResearcherW2 (ADK)Plain ADK agent, AG-UI text streaming, no tools
Demo Form BuilderW6 (A2UI)Emits A2UI form definitions the frontend renders
Demo Map ExplorerW7 (MCP Apps)Placeholder — describes what cloud-mode would render

The yellow banner at the top of every page reminds you what's stubbed (Firestore, Firebase auth, Vertex AI Search, Cloud Trace, etc.). Data is in-memory and ephemeral — it resets on every backend restart unless you set LOCAL_MODE_PERSIST=1 (writes to ~/.aitana-local/firestore.json on shutdown).

What gets stubbed

GCP serviceLOCAL_MODE replacement
FirestoreInMemoryFirestoreClient (drop-in for the methods v6 uses)
Firebase Admin AuthStub dep that accepts only local-mode-stub-token
Firebase web authFrontend bypasses sign-in, uses a fixed workshop identity
ADK sessionsInMemorySessionService (resets on backend restart)
ADK artifactsInMemoryArtifactService
Vertex AI SearchTool returns a "disabled in LOCAL_MODE" message
Cloud Trace / LoggingNo-op (logs still print to stdout)
MCP App sandboxFalls back to "describe what I'd render" responses

Everything else — ADK agent orchestration, AG-UI streaming, A2UI rendering, the chat UI, document workspace, skills layer — runs identically to cloud mode. The protocols don't care what backs the persistence layer.

Tier 1 — LOCAL_MODE (you are here)

Effort: 0 minutes. What's real: backend, frontend, agent, AG-UI, A2UI. What's stubbed: Firestore, auth, GCS, Vertex Search, telemetry.

Already covered above. Move on when you want persistent data or real Firebase auth.

Tier 2 — Shared dev Firestore (5 min)

Available when the workshop is being run by the Aitana Labs team. The shared your-project-id Firebase project has Anonymous Auth enabled (since 2026-05-02) with Firestore rules scoping each attendee's writes to their own anonymous UID.

Operator note (Aitana Labs team only): the dev billing account already has an account-wide budget alert configured. If you fork this repo and want a project-scoped budget guard on your own shared tier, here's the canonical one-liner:

gcloud billing budgets create \
  --billing-account=YOUR-BILLING-ACCOUNT-ID \
  --display-name="<your-project> workshop cap" \
  --budget-amount=50 \
  --threshold-rule=percent=0.5 \
  --threshold-rule=percent=0.9 \
  --filter-projects=projects/<your-project>

# In frontend/.env.local — paste from workshop materials:
NEXT_PUBLIC_FIREBASE_API_KEY=<from workshop>
NEXT_PUBLIC_FIREBASE_AUTH_DOMAIN=your-project-id.firebaseapp.com
NEXT_PUBLIC_FIREBASE_PROJECT_ID=your-project-id
NEXT_PUBLIC_FIREBASE_APP_ID=<from workshop>

# Do NOT set NEXT_PUBLIC_LOCAL_MODE (or set it to 0).
# Backend stays in LOCAL_MODE so no GCP creds are needed there.

Public Firebase web API keys are not secrets — security is enforced by Firestore rules and auth-domain restrictions (Firebase docs). The budget is capped at $50 and rules deny cross-user reads.

Tier 3 — Your own GCP project (30–60 min) {#graduating-from-local-mode}

For local development against your own Firestore, or when you want to deploy the platform yourself.

Step 1 — Create a GCP project

Enable these APIs in the Cloud Console:

  • Firestore (Native mode)
  • Identity Platform (or Firebase Auth)
  • Vertex AI
  • Cloud Storage

Step 2 — Application Default Credentials

gcloud auth application-default login
gcloud auth application-default set-quota-project YOUR_PROJECT_ID

The backend reads ADC for Firestore, Firebase Admin, and Vertex AI access.

Step 3 — Firebase web app config

In the Firebase Console, Project Settings → Your apps → Add app → Web. Copy the config snippet. Paste into frontend/.env.local:

NEXT_PUBLIC_FIREBASE_API_KEY=...
NEXT_PUBLIC_FIREBASE_AUTH_DOMAIN=...
NEXT_PUBLIC_FIREBASE_PROJECT_ID=...
NEXT_PUBLIC_FIREBASE_STORAGE_BUCKET=...
NEXT_PUBLIC_FIREBASE_MESSAGING_SENDER_ID=...
NEXT_PUBLIC_FIREBASE_APP_ID=...

# Unset (or set to 0):
NEXT_PUBLIC_LOCAL_MODE=0

Step 4 — Backend env

In backend/.env:

GOOGLE_CLOUD_PROJECT=your-project-id
# Unset LOCAL_MODE
LOCAL_MODE=0

Step 5 — Optional persistence beyond local Firestore

# Persistent chat history via Vertex AI Agent Engine
AGENT_ENGINE_ID=projects/YOUR_PROJECT/locations/YOUR_LOC/reasoningEngines/YOUR_ID
AGENT_ENGINE_STAGING_BUCKET=gs://your-staging-bucket

# GCS-backed artifacts (uploaded docs, intermediate tool outputs)
ADK_ARTIFACT_BUCKET=gs://your-artifact-bucket

# Vertex AI Search (the ai_search tool returns "disabled" if unset)
VERTEX_AI_SEARCH_DATASTORE_ID=projects/.../dataStores/...

⚠️ Set both AGENT_ENGINE_ID and ADK_ARTIFACT_BUCKET together, or neither. Mixing cloud sessions with in-memory artifacts strands sessions across backend restarts — the session's docs_loaded list survives but the artifacts behind it don't, and the document injector then loads nothing. The backend emits a WARNING at boot when only one is set.

Step 6 — Restart

make dev

The yellow LOCAL_MODE banner disappears. Sign in with Google through Firebase. Data persists in Firestore.

Troubleshooting

"Backend refuses to start: LOCAL_MODE=1 is set together with K_SERVICE..." You're trying to run LOCAL_MODE in a deployed context (Cloud Run / App Engine / GKE). This is a deliberate safety refusal — the auth-bypass stub must never be active in production. Unset LOCAL_MODE for cloud deployments.

Banner appears but chat returns 401 The frontend is sending the stub token but the backend isn't recognising it. Check that LOCAL_MODE=1 is set in backend/.env and the backend has been restarted since.

Sign-in screen appears in LOCAL_MODE Frontend isn't reading NEXT_PUBLIC_LOCAL_MODE. Make sure it's in frontend/.env.local (Next.js only reads .env.local, not .env) and that you've restarted npm run dev since.

"Search disabled in LOCAL_MODE" in chat replies Expected — Vertex AI Search isn't running locally. Set VERTEX_AI_SEARCH_DATASTORE_ID in cloud mode if you need it.

Architecture overview

For workshop attendees who want to read the code:

[Frontend :3456]
   ├─ Next.js 15 + React 19 + Tailwind
   ├─ AG-UI HttpAgent (streams SSE from backend)
   ├─ A2UI renderer for declarative UI from agents
   └─ MCP Apps iframe sandbox for protocol-prefixed widgets
       │
       └── /api/proxy/* → [Backend :1956]
                              ├─ FastAPI + Google ADK
                              ├─ Skills layer (3 demo skills in LOCAL_MODE)
                              ├─ ADK agent factory (Gemini / Claude / OpenAI)
                              ├─ MCP toolset for remote tools
                              └─ Auth: Firebase JWT or LOCAL_MODE stub

Read docs/design/v6.1.0/SEQUENCE.md for the build sequence and which sprints shipped which protocol layer.

Next steps

  • Browse the demo skills' code at backend/db/local_fixture.py
  • Read the protocol stack talk: docs/talks/ai-ui-protocol-stack.md
  • Build your own skill: open the chat UI's + Create a new skill button
  • Contribute: see CONTRIBUTING.md