Extensions Integrations

---
name: extensions-integrations
description: "Install, choose, configure, and troubleshoot AutoGen `autogen_ext`
  optional integrations for model clients, tools/workbenches, code executors,
  memory/cache backends, helper agents, and gRPC worker runtime surfaces."
disable-model-invocation: true
metadata:
  disco-role: operating
license: NOASSERTION
---

# AutoGen Extensions Integrations

Use this sub-skill when maintaining existing AutoGen Python applications that depend on `autogen_ext` optional integrations. AutoGen is in maintenance mode; for brand-new greenfield agent systems, consider Microsoft Agent Framework first. For existing AutoGen systems, this sub-skill helps future agents make safe, precise decisions about extension packages, provider configuration, and service-backed integrations.

## Start Here

1. Identify the exact integration surface before installing extras: model client, tool/workbench, code executor, memory/cache backend, helper agent, or gRPC runtime.
2. Install the narrow extra for that surface instead of a broad collection of optional dependencies. See `references/optional-extras.md`.
3. From this sub-skill directory, inspect the current environment without starting services or making provider calls:

   ```bash
   python scripts/inspect_extensions.py --json
   ```

4. Configure credentials and services explicitly; do not infer that an importable package means provider access, Docker, Redis, ChromaDB, Jupyter, Playwright, MCP, or Azure resources are usable.
5. Route team/agent orchestration questions to `agentchat-workflows`, low-level runtime patterns to `core-runtime`, and Studio/Magentic-One CLI/AG Bench/`pyautogen` compatibility to `tools-studio-bench`.

## Integration Map

| Need | Primary reference | Public imports to check |
| --- | --- | --- |
| OpenAI, Azure OpenAI, Anthropic, Azure AI Foundry/GitHub Models, Ollama, llama.cpp, Semantic Kernel, replay clients | `references/model-clients.md` | `autogen_ext.models.openai`, `autogen_ext.models.azure`, `autogen_ext.models.anthropic`, `autogen_ext.models.ollama`, `autogen_ext.models.llama_cpp`, `autogen_ext.models.semantic_kernel`, `autogen_ext.models.replay` |
| MCP tools/workbenches, HTTP tools, LangChain adapters, Azure AI Search, GraphRAG, Semantic Kernel tool adapter | `references/tools-workbenches.md` | `autogen_ext.tools.mcp`, `autogen_ext.tools.http`, `autogen_ext.tools.langchain`, `autogen_ext.tools.azure`, `autogen_ext.tools.graphrag`, `autogen_ext.tools.semantic_kernel` |
| Local, Docker, Jupyter, Docker-Jupyter, Azure Container Apps Dynamic Sessions code execution | `references/code-executors-and-runtimes.md` | `autogen_ext.code_executors.local`, `.docker`, `.jupyter`, `.docker_jupyter`, `.azure` |
| Disk/Redis cache and ChromaDB, Redis, mem0, canvas, task-centric memory | `references/optional-extras.md`, `references/troubleshooting.md` | `autogen_ext.cache_store.diskcache`, `.redis`, `autogen_ext.memory.chromadb`, `.redis`, `.mem0`, `.canvas` |
| Web/file/video surfer and Magentic-One helper agent surfaces inside `autogen_ext` | `references/optional-extras.md`, `references/troubleshooting.md` | `autogen_ext.agents.web_surfer`, `.file_surfer`, `.video_surfer`, `.magentic_one` |
| gRPC worker runtime extension package surfaces | `references/code-executors-and-runtimes.md` | `autogen_ext.runtimes.grpc` |

## Safety Defaults

- Treat `LocalCommandLineCodeExecutor` as trusted-code only; prefer Docker or a controlled remote executor when code is generated by an LLM.
- Do not start Docker containers, Jupyter kernels, MCP servers, browsers, Redis/ChromaDB/mem0 services, or Azure resources just to diagnose imports.
- For MCP stdio servers, review `command`, `args`, `env`, working directory, and filesystem roots before execution; MCP servers are code with the privileges of the process that starts them.
- Never print real API keys, tokens, endpoint secrets, connection strings, Redis passwords, or serialized model configs containing credentials.
- If an extension error is actually about AgentChat team composition, streaming, termination, or state, switch to `agentchat-workflows`; if it is about routed agents, topics, subscriptions, or distributed runtime design, switch to `core-runtime`.

## References

- `references/optional-extras.md` — narrow extra selection, dependency boundaries, and safe environment inspection.
- `references/model-clients.md` — provider client setup, required kwargs/env vars, model metadata, and no-credential diagnostics.
- `references/tools-workbenches.md` — MCP, HTTP, LangChain, Azure AI Search, GraphRAG, and Semantic Kernel tool integration.
- `references/code-executors-and-runtimes.md` — local/Docker/Jupyter/Azure executors, gRPC runtime extras, and service-dependent checks.
- `references/troubleshooting.md` — symptom-to-cause matrix for optional dependencies, credentials, service lifecycles, and serialization.