Invoke is a lightweight framework that connects LLMs with real-world APIs using natural language and structured tool calls.

See the full demo here.

pip install invoke-agent

from langchain_openai import ChatOpenAI
from invoke_agent.agent import InvokeAgent

# Use GPT-4.1 for best results
llm = ChatOpenAI(model="gpt-4.1")  
# Pass built-in aliases, file paths, or URLs; or omit to auto-load agents_map.yaml
invoke = InvokeAgent(llm, agents=["google-calendar", "./custom/weather.json"])

while True:
    user_input = input("📝 You: ").strip()
    if user_input.lower() in ["exit", "quit"]:
        break
    response = invoke.chat(user_input)
    print("\n🤖", response)

Youtube tutorial

For full control you can integrate with LangChain directly:

from invoke_agent.core import api_executor
from invoke_agent.context import build_context
from langchain.agents import AgentExecutor, create_tool_calling_agent
from langchain_core.prompts import ChatPromptTemplate

# Define your LLM
llm = ChatOpenAI(model="gpt-4.1")

# Build a prompt template
prompt = ChatPromptTemplate.from_messages([
    ("placeholder", "{chat_history}"),
    ("human", "{input}"),
    ("placeholder", "{agent_scratchpad}")
])

# Create the agent
agent = create_tool_calling_agent(llm=llm, tools=[api_executor], prompt=prompt)
executor = AgentExecutor(agent=agent, tools=[api_executor], verbose=True)

# Invoke with context containing your integrations
result = executor.invoke({
    "input": "What's the weather in Paris?",
    "chat_history": build_context(agents=["open-meteo"])
})
print(result["output"])

• 🌐 Access any HTTP API using natural language.
• 🔑 Automatic OAuth and API key management.
• 🧩 Flexible integrations via JSON, or TXT definitions.
• 🤖 Works with any LangChain-compatible LLM (we recommend GPT-4.1).

You can specify integrations via the agents parameter:

• Built-in aliases: e.g. "google-calendar", "open-meteo"
• File paths or URLs: direct references to .json or .txt definitions
• Explicit mappings: { "my-calendar": "./calendar.json" }

# All valid:
invoke = InvokeAgent(llm, agents=[
  "google-calendar",
  "https://example.com/my_agents.txt",
  {"custom-weather": "./agents/weather.json"}
])

# Or omit to load './agents_map.yaml' if present:
invoke = InvokeAgent(llm)

• agents.json is a structured schema defining tool names, URLs, methods, parameters, headers, and auth.
• agents.txt is the Markdown-rendered version produced by render_agents_txt(), used in the system prompt.

{
  "agent": "gmail",
  "label": "Gmail API",
  "base_url": "https://www.googleapis.com",
  "auth": {"type": "oauth", "code": "i"},
  "endpoints": [ /* ... */ ]
}

• None: no auth.
• api_key: api_key (Locally-managed API key)
• oauth: oauth::i (Invoke-managed OAuth flow)

Override per-endpoint using auth_code. When specified, it overrides the top-level auth.

• Auto YAML: omit agents to load agents_map.yaml if available.
• Explicit list: pass aliases, file paths, or mappings.
• Custom context: override system prompt via context argument.

Here’s how the pieces fit together:

      ┌─────────┐        ┌─────────────┐
      │ auth.py │───────▶︎│   core.py   │
      └─────────┘        └──────┬──────┘
                                │
                                ▼
      ┌───────────┐      ┌─────────────┐
      │context.py │─────▶︎│   agent.py  │
      └───────────┘      └─────────────┘         
           ▲                                      
           │                                      
      ┌────────────┐     ┌─────────────┐ 
      │compile.py  │◀︎────┤ agents.json │
      └────────────┘     └─────────────┘

Each file plays a specific role:

Jump right in with example notebooks to run your first agent with OpenAI, Claude, or Mistral.

Want to integrate a new API? Head to agents_json.md, add your agents.json and follow the prompts.

Need OAuth? See auth.md or override io.get_oauth_code().

from invoke_agent.auth import set_current_user

# Set namespace before running queries:
set_current_user('current_user_id')

All OAuthManager calls now use credentials under that user_id namespace, and will never prompt interactively.

⚠️ BETA SOFTWARE — NOT FOR PRODUCTION

This project is in active development and provided as-is, for testing and evaluation purposes only. Use at your own risk. The author is not liable for any bugs, breakage, data loss, security issues, or cosmic anomalies that may arise from using this code.