Python Agentic Chat

examples/python-agentic-chat demonstrates an assistant-ui frontend calling a FastAPI backend that runs OpenAI Agents SDK through the vel4ai Python SDK.

Files To Read

path	Role
`examples/python-agentic-chat/app/assistant.tsx`	assistant-ui runtime, Hosted Connect handoff, and state rendering
`examples/python-agentic-chat/backend/app.py`	FastAPI endpoints and OpenAI Agents streaming conversion
`examples/python-agentic-chat/backend/state.py`	AssistantTransport state helpers
`packages/sdk-python/src/vel4ai/integrations/openai_agents.py`	Per-model-call OpenAI Agents adapter

Local Setup

Install Python packages from the monorepo checkout:

python -m pip install -e packages/sdk-python[agents] -e examples/python-agentic-chat/backend

Start the backend:

python -m uvicorn backend.app:app --reload --port 8100 --app-dir examples/python-agentic-chat

Start the frontend:

pnpm --filter @vela/example-python-agentic-chat dev

Production serves this demo at https://demo.vel4.ai/agent. The browser calls same-origin /agent/api/* routes, and the Core Worker gateway rewrites those requests to the FastAPI backend’s internal /api/agent/* routes with x-vela-demo-proxy-token.

For local direct backend calls, set:

NEXT_PUBLIC_AGENT_API_URL=http://localhost:8100/api/agent

Hosted Connect opens on https://app.vel4.ai by default. If a local or preview Vela API returns a /connect?session=... URL on another origin, the backend normalizes that URL to VELA_CONNECT_PUBLIC_BASE_URL.

For production deploys, use Fly.io for the backend:

fly apps create vela-python-agentic-chat-api
fly secrets set VELA_APP_ID="$VELA_AGENT_DEMO_APP_ID" VELA_APP_TOKEN="$VELA_AGENT_DEMO_APP_TOKEN" VELA_AGENT_DEMO_PROXY_TOKEN=... -a vela-python-agentic-chat-api
pnpm python-agentic-chat-api:deploy
pnpm python-agentic-chat:deploy

Set VELA_AGENT_MODEL to a model that the connected Wallet source can execute. If authorization returns capability_requirement_unmet, pick a model that is both allowed by the delegation and available in the tenant model inventory.

State Shape

The AssistantTransport state is a JSON object with:

field	Purpose
`messages`	User and assistant transcript
`steps`	Tool calls and agent progress
`activeAgent`	Current OpenAI Agents SDK agent name
`readiness`	Optional readiness snapshot
`error`	User-visible error or non-allow reason

OpenAI Agents stream events are mapped into assistant text deltas, tool step updates, active-agent changes, and error messages.

Demo Tools

The demo keeps tools local and deterministic. This lets the example show the agent loop, Vela authorization, and assistant-ui state updates without adding external API, database, or RAG setup.

tool	When it should be used
`lookup_project_policy`	Vela policy, audit, or Local LLM notes
`inspect_delegation_readiness`	Hosted Connect / delegation readiness before a model call
`estimate_agent_run_cost`	Expected model-call count, demo cost, and per-call authorization
`compare_execution_sources`	Vela Credits vs BYOK vs Local LLM tradeoffs
`draft_revoke_test_plan`	Delegation revoke and deny behavior checks

Example prompts:

Vela の監査はどのステップで記録される？policy tool も使って。
この delegation で agent を動かす前の readiness を確認して。
OpenAI BYOK と Vela Credits と Local LLM の違いを比較して。
この agent run の概算コストと model call ごとの authorization を説明して。
delegation を revoke した直後に何を確認すべきかテスト手順を書いて。

Acceptance

Create or reuse a Hosted Connect delegation.
Ask an agentic prompt that triggers the demo function tool.
Confirm assistant-ui shows a tool step and final response.
Confirm Vela audit shows authorization / execution entries for each OpenAI Agents model call.
Revoke the delegation and confirm the next model call is denied before the provider call starts.