Agent Framework AppKit

Utility app kit for Microsoft Agent Framework: configuration, DI, agent/client factories, resilience, and Skills helpers without adding a second framework.

Why this package exists

Microsoft Agent Framework already provides the core agent runtime. This package stays intentionally small and provides reusable app-level helpers around native MAF extension points:

Install

pip install agent-framework-appkit

For local development:

git clone https://github.com/Dongbumlee/agent-framework-appkit.git
cd agent-framework-appkit
uv sync --locked --dev
uv run pytest -q
uv run ruff check .
uv run python -m build

Quick start

from agent_framework_appkit.agents import AgentClientRegistry, build_agent
from agent_framework_appkit.config import AgentFrameworkSettings

settings = AgentFrameworkSettings(
    env={
        "AZURE_OPENAI_ENDPOINT": "https://example.openai.azure.com/",
        "AZURE_OPENAI_MODEL": "gpt-4o",
    }
)

registry = AgentClientRegistry(credential_factory=lambda service: get_token_provider())
registry.initialize(settings)

client = registry.get_client("default")
agent = build_agent(
    client=client,
    name="Assistant",
    instructions="You are a helpful assistant.",
    temperature=0.2,
)

get_token_provider() is intentionally not provided by this package. Bring your own credential factory from Azure Identity, your host app, or your platform.

Common use cases

1. Discover model endpoints from environment variables

from agent_framework_appkit.config import AgentFrameworkSettings

settings = AgentFrameworkSettings(
    custom_service_prefixes={
        "east": "AZURE_OPENAI_EASTUS",
        "west": "AZURE_OPENAI_WESTUS",
    }
)

for service_id in settings.get_available_services():
    service = settings.get_service_config(service_id)
    print(service_id, service.chat_deployment_name)

Supported model aliases include:

• {PREFIX}_CHAT_DEPLOYMENT_NAME
• {PREFIX}_CHAT_MODEL
• {PREFIX}_CHAT_COMPLETION_MODEL
• {PREFIX}_MODEL

Supported endpoint aliases include:

• {PREFIX}_ENDPOINT
• {PREFIX}_BASE_URL
• {PREFIX}_BASEURL

2. Use lightweight dependency injection

from agent_framework_appkit.di import AppContext

class Repository:
    def list_items(self) -> list[str]:
        return ["alpha", "beta"]

class Service:
    def __init__(self, repo: Repository) -> None:
        self.repo = repo

ctx = AppContext()
ctx.add_singleton(Repository)
ctx.add_transient(Service, lambda c: Service(c.get_service(Repository)))

service = ctx.get_service(Service)
print(service.repo.list_items())

3. Build agents with native MAF options

from agent_framework_appkit.agents import build_agent

agent = build_agent(
    client=chat_client,
    name="ResearchAssistant",
    instructions="Answer with citations when possible.",
    tools=[search_docs],
    temperature=0.1,
    max_tokens=1200,
)

build_agent returns a native agent_framework.Agent.

4. Add retry and compaction helpers

from agent_framework_appkit.agents import ClientType, create_client
from agent_framework_appkit.resilience import RateLimitRetryConfig

client = create_client(
    ClientType.OpenAIChatClient,
    azure_endpoint="https://example.openai.azure.com/",
    model="gpt-4o",
    credential=token_provider,
    api_version="2024-10-01-preview",
    enable_rate_limit_retry=True,
    retry_config=RateLimitRetryConfig(max_retries=3),
)

Compaction can be controlled with environment variables:

MAF_COMPACTION_ENABLED=true
MAF_COMPACTION_MAX_MESSAGES=80
MAF_COMPACTION_COMPACT_TO=40

5. Attach file-based Skills

from agent_framework_appkit.agents import build_agent

agent = build_agent(
    client=chat_client,
    instructions="Use available skills when helpful.",
    skill_paths=["./skills"],
    require_script_approval=True,
)

You can also configure skills through environment variables:

MAF_SKILL_PATHS=./skills,./more-skills
MAF_SKILL_REQUIRE_APPROVAL=true
MAF_SKILL_SCRIPT_EXTENSIONS=.py,.ps1

Examples

See the examples/ directory:

Design guardrails

• This is not a second agent framework.
• No custom dotenv loading; load .env before constructing settings if needed.
• No template-specific credential utilities; pass credentials or credential factories explicitly.
• No declarative agent/orchestration layer; use native MAF APIs for those surfaces.
• Keep helpers thin, typed, and easy to remove if MAF adds native equivalents.

Development

uv sync --locked --dev
uv run pytest -q
uv run ruff check .
uv run python -m build

CI runs the same commands on pull requests and pushes to main.

Publishing to PyPI

Publishing uses PyPI trusted publishing through GitHub Actions. No PyPI API token is stored in GitHub secrets.

GitHub Actions workflow:

• Workflow file: .github/workflows/ci.yml
• Publish trigger: GitHub release published
• GitHub environment: pypi
• PyPI project: agent-framework-appkit

Configure the trusted publisher in PyPI with these values:

Release flow:

1. Update version in pyproject.toml.

2. Run local validation:

   uv sync --locked --dev
   uv run pytest -q
   uv run ruff check .
   uv run python -m build
   

3. Commit and push the version change.

4. Create and publish a GitHub release, for example:

   gh release create v0.1.0 --title "v0.1.0" --notes "Initial release"
   

5. The publish job builds the distributions and publishes them to PyPI using the GitHub OIDC identity.

License

MIT