RapidAgent

Plug-and-play agentic framework with multi-backend support, native multi-LLM support, Azure OpenAI + Azure AD auth, and built-in security guardrails (PII/PHI redaction, prompt injection prevention, tool authorization).

Installation

pip install rapidagent                  # Core (Direct backend — all Python 3.10+)
pip install rapidagent[langgraph]       # + LangGraph backend
pip install rapidagent[azure]           # + Azure AD auth
pip install rapidagent[anthropic]       # + Anthropic Claude SDK
pip install rapidagent[google]          # + Google Gemini SDK
pip install rapidagent[litellm]         # + LiteLLM (100+ providers — see note below)
pip install rapidagent[cli]             # + CLI commands
pip install rapidagent[all]             # Everything

Quickstart

from rapidagent import RapidAgent, tool

@tool
def get_weather(city: str) -> str:
    """Get current weather."""
    return f"Weather in {city}: 72F, sunny"

agent = RapidAgent(model="openai:gpt-4o-mini", tools=[get_weather])
result = agent.invoke("Weather in SF?")
print(result["content"])

Or from YAML:

name: my-agent
model: openai:gpt-4o-mini
system_prompt: "You are helpful."
tools:
  - name: current_time
    description: Get the current time
    function: rapidagent.tools.builtins:current_time

agent = RapidAgent.from_config("agent.yaml")

---

LLM Provider Support

RapidAgent supports 40+ LLM providers natively using their official SDKs. No single point of compromise.

Native providers (official SDKs — recommended)

Provider	Model String	SDK	Auth Env Var
OpenAI	openai:gpt-4o-mini	openai (core)	OPENAI_API_KEY
Azure OpenAI	(dict config)	openai (core)	AZURE_OPENAI_ENDPOINT
Anthropic Claude	anthropic:claude-3-5-sonnet-latest	anthropic	ANTHROPIC_API_KEY
Google Gemini	google:gemini-2.0-flash	google-genai	GEMINI_API_KEY

OpenAI-compatible providers (no extra SDKs)

Uses the same openai SDK (already installed) with a custom base_url. Zero additional dependencies, zero additional risk.

Provider	Config
Groq	provider: openai_compatible, endpoint: https://api.groq.com/openai/v1
Together AI	provider: openai_compatible, endpoint: https://api.together.xyz/v1
Ollama (local)	provider: openai_compatible, endpoint: http://localhost:11434/v1
OpenRouter	provider: openai_compatible, endpoint: https://openrouter.ai/api/v1
DeepSeek	provider: openai_compatible, endpoint: https://api.deepseek.com
Perplexity	provider: openai_compatible, endpoint: https://api.perplexity.ai
Fireworks	provider: openai_compatible, endpoint: https://api.fireworks.ai/inference/v1
+ 30 more	Any OpenAI-compatible endpoint

model:
  provider: openai_compatible
  model: llama-3.3-70b-versatile
  endpoint: https://api.groq.com/openai/v1

agent = RapidAgent(model="anthropic:claude-3-5-sonnet-latest")
agent = RapidAgent(model="google:gemini-2.0-flash")

LiteLLM (explicit opt-in — for providers not covered above)

LiteLLM can route through 100+ providers via a single package. Has had past security incidents (supply-chain attacks exfiltrating credentials). Use only if your provider isn't covered by the native providers above.

# Explicit opt-in — only works if you set this
agent = RapidAgent(model="litellm:bedrock/anthropic.claude-v2")

Installing rapidagent[litellm] shows a runtime warning every time it's used.

---

Multi-Backend Architecture

Python	Default Backend	Features
3.10 – 3.13	LangGraph	Full durability, checkpointing, HITL, multi-agent, LangSmith.
3.14+	Direct	Pure Python ReAct loop. Tools, streaming, memory, checkpointing included.

agent = RapidAgent(model="openai:gpt-4o-mini", backend="direct")

---

Security Guardrails

Six built-in guardrails + audit trail. Enable via YAML or code.

guardrails:
  input_filter:
    enabled: true
    block_patterns:
      - "ignore all instructions"
  tool_auth:
    enabled: true
    allowlist: [get_weather, current_time]
    denylist: [rm.*, exec.*]
  pii:
    enabled: true
    patterns:
      email: redact
      ssn: block
  phi:
    enabled: false           # HIPAA — opt-in
    include_pii: true
  output_filter:
    enabled: true
  audit_log:
    enabled: true
    path: ./audit/audit.jsonl

What each guardrail does

Guardrail	Hook Points	Default
InputGuardrail	Input → LLM	Blocks prompt injection, jailbreaks
ToolAuthGuardrail	Before tool exec	Denylist: exec, eval, rm, subprocess, drop table, shutdown
PIIScanner	All I/O points	Redacts email, phone, API keys. Blocks 9-digit SSNs
PHIScanner	Same as PII	14 HIPAA patterns: MRN, Medicare IDs, VINs, IPs, health dates. Opt-in
OutputGuardrail	Response → user	Blocks hate speech, violent content
RateLimiter	Input	Sliding window (100 req / 60s)

Every decision logged to JSONL audit file. Supports custom sinks (Datadog, Splunk).

agent = RapidAgent(
    model="openai:gpt-4o-mini",
    guardrails={
        "input_filter": {"block_patterns": ["BADWORD"]},
        "phi": {"enabled": True},
    },
)

---

CLI

rapidagent doctor              # Diagnose environment
rapidagent init my-project     # Interactive setup with prereq checks
rapidagent new my-project      # Quick scaffold
rapidagent run agent.yaml      # Interactive chat
rapidagent ask agent.yaml "Hi" # Single question
rapidagent list                # Available blueprints

---

Features

Feature	Description
Multi-backend	LangGraph (3.10-3.13) or Direct (3.14+) — auto-selected
40+ LLM providers	OpenAI, Anthropic, Google, Groq, Together, Ollama, OpenRouter, DeepSeek, ...
Azure OpenAI + AD	API key, service principal, or DefaultAzureCredential
Security guardrails	PII/PHI redaction, prompt injection, tool auth, rate limiter, audit trail
Crash recovery	SQLite checkpointing after every LLM + tool step
Structured output	Pass any Pydantic model for JSON output
Token/cost tracking	Token counts and cost per invocation
Retry with backoff	Exponential backoff on API errors (rate limits, timeouts)
@tool decorator	Auto-generates LLM tool schemas from type hints
Multi-agent	Supervisor/worker, swarm, and sequential patterns
CLI	Scaffold, init, doctor, run, chat
128 tests	Completely covered

---

Configuration

name: customer-support
model:
  provider: azure_openai
  deployment: gpt-4o
  endpoint: https://my-resource.openai.azure.com
  auth_type: default_credential
backend: direct
system_prompt: "You are a helpful support agent."
max_iterations: 10

tools:
  - name: lookup_order
    description: Look up an order
    function: tools.orders:lookup_order

guardrails:
  input_filter:
    enabled: true
  tool_auth:
    enabled: true
  pii:
    enabled: true
  audit_log:
    path: ./audit/audit.jsonl

human_in_the_loop:
  enabled: true
  require_approval_for:
    - process_refund

---

License

MIT