Skip to content
The Agents Index
Sign inSubmit an agent

Microsoft Agent Framework

Open-source .NET, Python and Go SDK for building AI agents and graph-based multi-agent workflows.

Microsoft Agent Framework is an MIT-licensed SDK for building AI agents and graph-based multi-agent workflows in .NET, Python and Go. It is the successor to Semantic Kernel and AutoGen, combining their agent abstractions with session state, middleware and telemetry.

Best for
.NET or Python teams already invested in Azure or Microsoft Foundry who want one supported agent SDK, an optional managed hosting path, and a documented migration off AutoGen or Semantic Kernel.
Not for
Teams with no Azure or Microsoft affinity who want the largest third-party integration ecosystem and the longest production track record: LangChain or CrewAI carry more years of battle-testing today.

What we checked

Pages last read
Facts checked
26 factsFrom 23 sources
AI answers
Named in 5 of 150Our October check

How we checkEditorial policy

Visit the site CompareAlternatives

Microsoft Agent Framework in our research

All research

Study

AI Agent Pricing Index, October 2026

Pricing

Free only

17 of the 149 agents we checked are free with no paid plan.

Price checked 19 September 2026.

What we checked on Microsoft Agent Framework

YesPartlyNo

What it doesPartly

Writes code

Agents you build can write and run code through CodeAct or a provider-hosted code interpreter, but both are gated: the documented CodeAct connector is in preview and code interpreter depends on the model provider you choose.

CodeAct lets an agent solve a task by writing code and executing it through an execute_code tool. Instead of asking the model to emit one tool call…

learn.microsoft.com/en-us/agent-framewo…Checked

What it doesPartly

Whole-codebase work

The Harness targets long multi-step work with context compaction, file memory and todo tracking, and shell tools default to working-directory confinement; the vendor documents no repository indexing or repo-wide migration.

A harness wraps a chat client with the scaffolding an agent needs to work through long, multi-step tasks - planning / execution modes, a todo list to…

learn.microsoft.com/en-us/agent-framewo…

What it doesPartly

Ships it

The beta agent-framework-tools package gives an agent local or Docker shell execution, so it can run any command you approve; no deploy, release or pull-request step is documented.

Shell execution can modify files, launch processes, access credentials, and communicate with external systems. Use the least-privileged execution…

learn.microsoft.com/en-us/agent-framewo…

What it doesNo

Done-for-you service

Microsoft ships a package you install and run yourself; nobody builds or operates the agent for you, and support is the public GitHub repository, discussions and office hours.

Microsoft Agent Framework is an open-source SDK for building AI agents that can reason, use tools, and interact with users and other agents. It…

learn.microsoft.com/en-us/agent-framewo…Checked

How much it does unattendedYes

Runs autonomously

The agent-mode provider ships an explicitly autonomous execute mode that works a plan to completion, and the Harness keeps an agent progressing through long multi-step tasks without a person each turn.

Execute is autonomous. The agent works through the plan, makes reasonable choices when details are ambiguous, and marks todos complete.

learn.microsoft.com/en-us/agent-framewo…Checked

How much it does unattendedYes

Multi-agent

Five built-in orchestration patterns coordinate several agents on one task, and a parent agent can additionally delegate concurrent work to named child agents in their own sessions.

Agent Framework provides several built-in multi-agent orchestration patterns:

learn.microsoft.com/en-us/agent-framewo…Checked

How much it does unattendedYes

Agent permissions

Approval is per tool and configurable in code: the local shell tool asks for every command by default, regex allow and deny lists pre-filter commands, and disabling approval requires an explicit unsafe acknowledgement.

LocalShellTool runs commands directly on the host. It defaults to a persistent shell, a 30-second timeout, 64-KiB output truncation,…

learn.microsoft.com/en-us/agent-framewo…Checked

Which models it runs onYes

Claude

Anthropic is a first-class model provider with its own client package, reachable through Anthropic directly or through Foundry, Bedrock and Vertex AI endpoints.

Anthropic — Claude model inference through Anthropic and supported hosted endpoints.

learn.microsoft.com/en-us/agent-framewo…Checked

Which models it runs onYes

GPT

Both OpenAI directly and Azure OpenAI are listed model providers, covering Chat Completions and the Responses API, with Microsoft Foundry as a third route to the same models.

OpenAI — OpenAI Chat Completions and Responses APIs.

learn.microsoft.com/en-us/agent-framewo…Checked

Which models it runs onYes

Open models

Ollama, ONNX Runtime GenAI and Foundry Local are all listed as providers for running open-weight models locally, though the comparison table shows they support fewer hosted-tool features than the cloud providers.

Ollama — Local open-source model inference.

learn.microsoft.com/en-us/agent-framewo…Checked

Which models it runs onYes

Model choice

You choose the model: eleven providers are documented and, in .NET, any inference service implementing the standard chat-client interface can back an agent without framework changes.

Any inference service that provides a Microsoft.Extensions.AI.IChatClient implementation can back a ChatClientAgent.

learn.microsoft.com/en-us/agent-framewo…Checked

Where you use itNo

In your editor

Not an editor product. The documented client surfaces are OpenAI-compatible HTTP, A2A, AG-UI web apps, Telegram and MCP; no IDE or editor extension appears in the docs.

After choosing a host, select the client integration that fits your scenario:

learn.microsoft.com/en-us/agent-framewo…Checked

Where you use itPartly

On the command line

A terminal surface exists only as developer tooling: the DevUI command line launches agents from a directory, and the Harness console apps are explicitly samples rather than shipped components.

If you have agents and workflows organized in a directory structure, launch DevUI from the command line:

learn.microsoft.com/en-us/agent-framewo…Checked

Where you use itPartly

In a browser

DevUI gives a browser interface for running and debugging agents, but the vendor labels it a sample app unfit for production; shipping a real web surface means adding the preview AG-UI or ChatKit integrations.

DevUI is a sample app to help you visualize and debug your agents and workflows during development. It is not intended for production use.

learn.microsoft.com/en-us/agent-framewo…Checked

Whose machine it runs onYes

Self-hosted

Self-hosting is one of the two documented hosting models: the agent runs in your own web service, container or runtime and you keep routing, identity, storage and scaling.

Your application runs in your web service, container, runtime, or existing infrastructure.

learn.microsoft.com/en-us/agent-framewo…Checked

Whose machine it runs onYes

Open source

The whole framework is published under the MIT licence on GitHub, permitting commercial use, modification and redistribution without a fee or a contributor agreement.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to…

github.com/microsoft/agent-framework/bl…Checked

What it costs to runPartly

How it meters

The framework itself costs nothing, so all spend is model and tool consumption metered by whichever provider you configure; Microsoft publishes no calculator and states the usage cost is yours.

Third-Party Systems are Non-Microsoft Products under the Microsoft Product Terms and are governed by their own third-party license terms. You are…

learn.microsoft.com/en-us/agent-framewo…Checked

What it costs to runYes

Free tier

No sales call, no account and no licence fee: the packages install from PyPI and NuGet under an MIT grant that is explicitly free of charge.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software")

github.com/microsoft/agent-framework/bl…Checked

What it costs to runYes

API access

The product is itself a code library, and its hosting packages additionally expose an agent over OpenAI-compatible Responses and Chat Completions endpoints from your own server.

OpenAI-compatible endpoints for Responses and Chat Completions-compatible APIs.

learn.microsoft.com/en-us/agent-framewo…Checked

What it costs to runPartly

MCP server

Another agent can reach yours over MCP or A2A, but the Python hosting packages that expose those endpoints are prerelease, and you must build and secure the server yourself.

MCP tools for exposing an agent or workflow as a native MCP tool.

learn.microsoft.com/en-us/agent-framewo…Checked

What it costs to runYes

Bring your own key

You bring your own inference: the framework requires you to configure your chosen provider's keys and endpoints, and it never brokers or resells model access.

API Dependencies: Requires proper configuration of LLM provider keys and endpoints

github.com/microsoft/agent-framework/bl…Checked

Buying it for a teamPartly

Admin controls

Constraints are real but developer-set in code, not an admin console: per-tool approval, regex shell allow and deny lists, and FIDES policy middleware that blocks a sensitive tool before it runs.

ShellPolicy applies regular-expression allow and deny lists before execution. Deny rules take precedence.

learn.microsoft.com/en-us/agent-framewo…Checked

Buying it for a teamPartly

Audit log

Agent actions are traceable via OpenTelemetry, and the FIDES security provider keeps per-session audit records of policy decisions, but there is no administrative action log because there is no admin plane.

SecureAgentConfig stores labels, hidden variables, audit records, and pending approvals in the active AgentSession.

learn.microsoft.com/en-us/agent-framewo…Checked

What happens to your codePartly

Opt out of training

Microsoft makes no training claim for the framework and instead pushes the question to whichever model or tool provider you configure, telling you to review their retention and privacy terms first.

Provider Due Diligence: Before integrating any external service, you should review their privacy policies, security practices, data retention…

github.com/microsoft/agent-framework/bl…Checked

What happens to your codePartly

Data residency

Because you host it, residency follows your deployment and your chosen providers; Microsoft explicitly makes keeping data inside your compliance and geographic boundaries your responsibility rather than guaranteeing a region.

It is your responsibility to manage whether your data will flow outside of your organization's Azure compliance and geographic boundaries and any…

learn.microsoft.com/en-us/agent-framewo…Checked

What happens to your codeYes

Getting out

Nothing is locked in a vendor store: sessions and history go to a session store you implement over your own database or blob storage, and there are no credits to expire.

Your application controls routing, identity, authorization, request policy, storage, deployment, and scaling.

learn.microsoft.com/en-us/agent-framewo…Checked

Microsoft Agent Framework price and terms

Terms

Free plan
YesYesThe entire framework is free under the MIT licence with no paid tier, no seats and no usage cap
The terms in full

the only spend is inference and tooling billed by the model provider you configure.

For AI agents

MCP server
Partly
API
Yes

Developer docs

Where these prices come from

github.com/microsoft/agent-framework/blob/main/LICENSE, checked

AdvertisementReach the teams comparing AI agent frameworks here.One banner on The Agents Index's category pages and agent pages.Advertise here

Microsoft Agent Framework alternatives

All 9 alternatives

Head to headMicrosoft Agent Framework vs Google Agent Development Kit (ADK)Microsoft Agent Framework vs AutoGen

Questions about Microsoft Agent Framework

How much does Microsoft Agent Framework cost?
Microsoft Agent Framework is free to use. From github.com/microsoft/agent-framework/blob/main/LICENSE, checked 19 September 2026.
Does Microsoft Agent Framework have an API or an MCP server?
It has an API. It has a limited MCP server.
Is Microsoft Agent Framework open source, and can you host it yourself?
It is open source. You can host it yourself.
What have we checked on Microsoft Agent Framework?
26 facts, under these headings: What it does, how much it does unattended, which models it runs on, where you use it, whose machine it runs on, what it costs to run, buying it for a team, what happens to your code.
How do we check Microsoft Agent Framework?
We hold 26 facts on Microsoft Agent Framework, quoted from 23 pages. Facts and price last checked 19 September 2026.

How we check every listing

Sources and changes

Last changed

Sources23

Written about it elsewhere1

Repository30 August 2026

Licence
MIT
Stars
13,219
Forks
2,240
Last push
30 August 2026

Is this yours?

AI assistants like ChatGPT and Claude opened The Agents Index pages 2,746 times in September. Claiming is free. You get this page's reader numbers and the verified owner mark. You confirm it with an email at learn.microsoft.com.

Claim this listingGet Featured

Report an error if a fact is wrong.