GUIDES

A LangGraph Agent with Tools, Guardrails, and Memory

April 15, 2026

A LangGraph agent with local function tools, guardrails, memory, human-in-the-loop, and optional MCP server connections.

Features

Local tools: filesystem, network, web search, math, bash
Optional MCP server connections (math, perf)
Configurable LLM providers: Ollama (default), OpenAI, Google Gemini, Anthropic, Microsoft Azure, Meta, DeepSeek, Mistral, xAI, OpenRouter
Conversation memory via LangGraph checkpointing (SQLite)
Guardrails: input validation, LLM-as-judge, output PII redaction, bash command denylist
Human-in-the-loop tool call confirmation
Graceful error handling for LLM connection failures
HTTP API via FastAPI with streaming support
Playground UI at /playground
React frontend under frontend/
LangSmith tracing support

Getting Started

uv sync
./run-cli-agent.sh

For a different LLM provider:

export GEMINI_API_KEY=your-api-key
./run-cli-agent.sh --provider gemini

HTTP API

./run-server-api.sh

# Single response
curl -X POST http://localhost:8000/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "what is 2 + 2?", "thread_id": "session-1"}'

# Streaming response
curl -X POST http://localhost:8000/chat/stream \
  -H "Content-Type: application/json" \
  -d '{"message": "list files in /tmp", "thread_id": "session-1"}'

Use the same thread_id across requests to maintain conversation history.

React Frontend

./frontend/run-frontend.sh   # http://localhost:5173

Requires server_api.py running on port 8000. Vite proxies /chat to the backend automatically.

MCP Servers

./mcp_servers/example/run-mcp-server.sh
./run-cli-agent.sh

Environment Variables

Variable	Default	Purpose
`LLM_PROVIDER`	`ollama`	LLM backend
`LLM_MODEL`	(provider default)	Override model name
`OPENAI_API_KEY`	—	Required for `openai`; default model: `gpt-4.1-nano`
`GEMINI_API_KEY`	—	Required for `gemini`; default model: `gemini-2.5-flash`
`ANTHROPIC_API_KEY`	—	Required for `anthropic`; default model: `claude-sonnet-4-6`
`AZURE_OPENAI_API_KEY`	—	Required for `azure`; default model: `gpt-4o`
`AZURE_OPENAI_ENDPOINT`	—	Required for `azure`; your Azure OpenAI endpoint URL
`AZURE_OPENAI_API_VERSION`	`2025-01-01-preview`	Azure OpenAI API version
`LLAMA_API_KEY`	—	Required for `meta`; default model: `Llama-4-Scout-17B-16E-Instruct`
`DEEPSEEK_API_KEY`	—	Required for `deepseek`; default model: `deepseek-chat`
`MISTRAL_API_KEY`	—	Required for `mistral`; default model: `mistral-large-latest`
`XAI_API_KEY`	—	Required for `xai`; default model: `grok-3-mini`
`OPENROUTER_API_KEY`	—	Required for `openrouter`; default model: `openrouter/auto`
`GUARDRAILS_ENABLED`	`true`	Enable input/output guardrails
`HUMAN_IN_LOOP`	`false`	Prompt user to confirm tool calls before exec
`LANGCHAIN_API_KEY`	—	Enables LangSmith tracing when set

Source

Full source code available on GitHub: jecklgamis/langgraph-example