Custom Demo Agent

A customizable deep agent for customer demos: analyze data, build dashboards or HTML assets, and use connected tools. Each customer gets their own branding, instructions and sample questions.

Watch the walkthrough.

If you're a LangChain employee, you can access a hosted version here.

Prerequisites

• uv and Python 3.13 (uv installs the pinned Python version).
• Node 22.12+ and npm for the frontend.
• Set LANGSMITH_API_KEY and ANTHROPIC_API_KEY in the root .env (for the default model).

Run locally

uv sync --group dev
cp .env.example .env                   # fill in LangSmith and model-provider keys
uv run python scripts/preflight.py     # checks connectivity; makes real API calls
./run.sh                              # backend :2024, frontend :3000

Open http://127.0.0.1:3000. In Settings, choose a workspace, then + New to enter a use case. Setup creates the branded assistant, sample files, skills and questions.

Setup time: A new use case takes about 30 seconds to setup. Sandbox prewarming runs in the background; the first turn can be slow while the VM and analysis packages start.

Optionally enable demo traffic to populate 200 sample traces for Monitoring, Insights and Engine.

What this demos

Platform	Features
Deep Agents	Skills, sandbox/code execution, dynamic subagents, MCP tools and Apps, human approval, generative UI (streamed dashboards and HTML assets)
LangSmith	Context Hub prompts/skills, tracing, Monitoring, Insights, Engine (issue detection), Evals and live prompt fixes

Architecture

A basic create_deep_agent with custom tools and middleware, plus a React frontend:

• Setup resolves the customer scenario and prepares data, skills, prompts and evals.
• Runtime applies the assistant's model/tools, reads its prompt fresh and runs the agent.
• Sandbox owns working files; Context Hub stores prompts and skills.
• Frontend streams tool activity, dashboards and HTML assets from the same conversation.

Assistants:

**A use-case-specific configuration of the shared demo agent. **

Its context controls execution; metadata holds branding and quick actions.

// Research: dashboards, web search and a connected operations system.
{
  "enabled_tools": ["push_widget", "web_search"],
  "mcp_servers": [{ "id": "ops", "label": "Operations", "url": "https://your-server.example/mcp" }]
}

Additional ways to run

Mode	How
Local	./run.sh; override ports with PORT and SPA_PORT
Deployed	The configured LangSmith GitHub integration deploys the backend; Vercel builds the frontend. Configure VITE_LG_URL and VITE_LG_API_KEY in Vercel; provider keys belong on the backend deployment.
Local frontend, deployed backend	Set VITE_LG_URL and VITE_LG_API_KEY in frontend/.env.local, then npm --prefix frontend run dev -- --port 3000. An existing lgUrl localStorage override takes precedence.
Voice	Click the composer microphone. Requires GEMINI_API_KEY on the backend; no per-assistant enable flag.
Connected systems / MCP Apps	Add a server in Settings → MCP servers, then press Test. See below.

Implementation details: AGENTS.md. Development and checks: CLAUDE.md. More demos: voice, MCP Apps, release evals (these score the planted bug firing, opposite to presenter evals).