Can someone help explain the major differences between this project and the mcp-agent project that it was forked from?

[8bithalfadder]

A quick glance tells me that both are frameworks for writing agents quickly that interface with MCP servers, but I'm having a hard time seeing what the reason for the fork is. I don't want to sound rude, I'm genuinely curious why a user would want to choose one project over the other at this point.

[evalstate]

Not rude at all, quick summary (and to be honest I've lost track of the upstream status so some of these things may be merged in).

fast-agent gives you:

• Access to all of the MCP Specification Capabilities including Prompts, Resources and Sampling.
• A simpler, type-safe decorator style syntax for creating Agents and Workflows. In addition, Workflows in fast-agent are "just agents", whereas the upstream has a complex inheritance hierarchy internally. This drastically reduces the amount of code in the framework.
• fast-agent does not use instructor for structured outputs. mcp-agent makes 2 separate calls on generating a structured output, and we have observed major issues with differences in results with instructor (including made up agent names). Further, fast-agent uses the beta OpenAI feature to generate structured outputs.
• fast-agent uses MCP Types internally. this gives you the ability to easily transfer/serialize state/share context between agents, or to disk/over the network. It also makes it seamless working with the MCP libraries. This is a huge win if you are doing MCP development.
• Multi-modal support - support for PDF and Vision models / Tool Calls. I added the Audio type to the MCP Spec, and this [I think] will be the first framework to natively support Audio/Images with OpenAI (and possibly Gemini models). [edited to add - and utilities/api to make this easy too]
• ~>80% automated test coverage, which speeds up development, defect diagnosis and fixing and regressions. Also, massive improvements to internal type safety and User API definition.
• You can bring up an interactive shell to warm-up/discuss with agents. Further you can issue a ***SAVE_HISTORY command to save the conversation to a file, for later reload (or editing).
• In-built Passthrough and Playback (replay) LLMs to aid with testing and agent development. The Passthrough additionally lets you call talls with a ***CALL_TOOL command.
• Built in prompt-server to allow quick and easy deployment of complex multi-part/multi-resource prompts and resources.
• Coming soon, a documentation site where all of the above is made accessible (https://fast-agent.ai) - feel free to take a look (it's being completed as we speak).
• Various other improvements which I know have been merged upstream (progress display, v1 chat message display, JSONL logging with secrets revocation etc.)

As of 0.2.0 it's no longer a goal to maintain API compatibility with mcp-agent. Let me know if you have any questions?