Takes the donkey work out of AI development.
An independent, community-maintained Python SDK for consuming MuleSoft Agent Fabric capabilities: governed model and MCP tool access, Exchange publishing, user identity, and approvals from your own agent framework and IDE, without adopting Mule.
DDK connects your Python AI application to MuleSoft Agent Fabric while you keep your own framework and IDE, without deploying or running the Mule runtime. The gateway enforces policy; DDK makes its decisions visible and actionable in your code.
With donkey-kit, you can:
- 🔗 Use governed models and MCP tools: call models through the gateway, discover governed MCP servers in Exchange, and bind their tools to native framework objects.
- 📦 Publish from code and CI: scan agents and MCP servers, publish them to Exchange, and verify live tools and agent cards against what you published.
- 🔐 Carry user identity and pause for approval: forward on-behalf-of user tokens, gate tools with human-in-the-loop approvals, and handle the gateway's CIBA step-up decisions as typed outcomes.
- 🤝 Serve A2A agents: make your agent callable through the official
a2a-sdk, with local simulated ingress and CLI tooling to expose it behind a gateway. - 🚦 React to typed refusals and pace usage: catch structured governance exceptions, register refusal handlers and custom policy types, and use gateway-reported budget windows to slow down before hitting a limit.
- 🔭 Trace and test governance: OpenTelemetry GenAI spans, correlation IDs, cost attribution, a local gateway simulator, and a pytest conformance plugin help you observe calls and test refusal paths before production.
Framework support is deliberately differentiated. LangGraph and the OpenAI Agents SDK are conformance-tested adapters. Google ADK, Strands, Microsoft Agent Framework, Anthropic, CrewAI, and LlamaIndex provide native factories and connection_kwargs(), with framework-specific limitations documented in the support matrix.
DDK consumes platform capabilities; it does not replace policy enforcement or ship a provisioning control plane. See the feature overview and verification ledger for shipped features, evidence, and remaining verification-gated surfaces.
| Repo | What it is |
|---|---|
| donkey-development-kit | The Python SDK and CLI: governed model and MCP tool access, Exchange publication, user identity, approvals, A2A serving, typed refusals, budget pacing, telemetry, and local testing. |
| donkey-development-kit-demos | Runnable scenario and presentation demos, companion to the SDK. |
Alpha, published on PyPI. See the release notes for what's new. Requires Python 3.10 or newer.
pip install "donkey-kit[llm,langgraph]" # base + raw client + one frameworkAdd mcp for governed tool binding, a2a for agent serving, or cli,local for command-line tooling and the simulator. The all extra includes these plus LangGraph and OpenTelemetry, not every framework; add your framework's extra separately.
Start with the quickstart, choose your framework, and explore the demos. Registry, discovery, and publication operations need an Anypoint connected app with the appropriate scopes.
Unofficial & independent — not affiliated with, endorsed by, or supported by Salesforce or MuleSoft. “Agent Fabric”, “MuleSoft”, “Anypoint”, and “Omni Gateway” are Salesforce trademarks. Licensed under Apache-2.0.