agent-runtime-sdk

Single-agent runtime SDK for exposing MCP tools as an A2A-compatible agent service.

Install

pip install agent-runtime-sdk

Install optional Langfuse/smolagents observability support:

pip install "agent-runtime-sdk[observability]"

Basic Usage

Create an agent project:

my-agent/
├── agent_app/
│   ├── agent.yaml
│   └── plugin.py
└── main.py

Example main.py:

from pathlib import Path

from agent_runtime.standalone import build_single_agent_app, run_single_agent_app


app = build_single_agent_app(
    Path(__file__).resolve().parent / "agent_app" / "agent.yaml"
)


if __name__ == "__main__":
    run_single_agent_app(app)

Example SDK import:

from agent_runtime import AgentBuilder

Minimal Configuration

Example agent_app/agent.yaml:

agent:
  agent_id: example-agent
  name: Example Agent
  description: Example A2A agent backed by MCP tools.

runtime:
  public_base_url: ${AGENT_PUBLIC_BASE_URL:-http://127.0.0.1:10020}
  model:
    provider: openai_compatible
    api_base: ${MODEL_SOURCE_API_BASE}
    model_id: ${MODEL_SOURCE_MODEL_ID}
    api_key_env: MODEL_SOURCE_API_KEY

mcps:
  - name: example
    url: ${MCP_EXAMPLE_URL}
    transport: streamable-http

Run:

export MODEL_SOURCE_API_BASE="https://your-model-api.example.com"
export MODEL_SOURCE_MODEL_ID="your-model-id"
export MODEL_SOURCE_API_KEY="your-api-key"
export MCP_EXAMPLE_URL="http://127.0.0.1:8000/mcp"
export AGENT_PUBLIC_BASE_URL="http://127.0.0.1:10020"

python main.py

Langfuse

After installing agent-runtime-sdk[observability], configure:

export MCP_AGENT_LANGFUSE_ENABLED=true
export LANGFUSE_BASE_URL="http://your-langfuse-host:3000"
export LANGFUSE_PUBLIC_KEY="pk-lf-..."
export LANGFUSE_SECRET_KEY="sk-lf-..."

Set MCP_AGENT_LANGFUSE_ENABLED=false to disable instrumentation.