An open source AI agent framework developed by llmindset.co.uk in the UK. Since its release in March 2025 it has held a clear position: an agent framework that speaks MCP (Model Context Protocol) natively. Agents are defined with a few Python decorators, and a CLI/TUI is bundled so you can talk to them right away. Standard workflow patterns — chain, parallel, router, evaluator-optimizer and orchestrator — ship with the framework, so multi-agent coordination does not have to be written from scratch. It is licensed under Apache 2.0, so the framework itself costs nothing to use.
Key Features
- A full MCP implementation: Beyond tool calls, it covers a wide slice of the MCP specification — Elicitation (the agent asking the user follow-up questions), Sampling (LLM calls initiated by the server), Roots and resource references. It also supports OAuth 2.1 with keyring storage for SSE/HTTP servers, plus transport diagnostics for Streamable HTTP, which makes MCP connection problems easier to isolate
- Agents defined with decorators: Decorators such as
@fast.agentand@fast.parallellet you declare agents and workflows. This is the default style when embedding agents in a Python application - Portable agent cards: Agent definitions can be written out as Markdown files (Agent Cards) and loaded from the TUI or the command line, which suits sharing definitions across a team or switching between definitions per task. Starter card packs for analyst, developer and researcher roles are included
- Five workflow patterns: Chain (sequential execution), Parallel (fan-out / fan-in), Router (LLM-based routing), Evaluator-Optimizer (iterating generation and evaluation to raise quality) and Orchestrator (task decomposition and planning). Agents-as-Tools, where one agent is handed to another as a tool, is also supported
- Develop interactively from the CLI/TUI: Talk to an agent directly in the terminal and watch intermediate steps and tool calls as you verify behavior. Agent skill management via the
/skillscommand, batch processing and structured outputs are supported as well - Broad model support: Anthropic, OpenAI and Google are supported natively, alongside providers such as Azure, Bedrock, DeepSeek and xAI, and local models through llama.cpp or Ollama. The same agent definition can be pointed at different models for comparison
Pricing
| Plan | Price | Highlights |
|---|---|---|
| Open source | Free (Apache 2.0) | All features. Install from PyPI and run it in your own environment |
The framework is free, but running it incurs the API costs of whichever model you use (local models aside).
Pricing is as of August 2026. Check the official site for the latest information.
Pros & Cons
✅ Pros
- The MCP implementation is broad, so it works as a real client for validating MCP servers including Elicitation and Sampling
- Agent definitions are short — a minimal setup runs in a few lines — keeping prototypes close to production
- Workflow patterns ship with the framework, so the boilerplate of multi-agent setups does not have to be written by hand
- The CLI/TUI lets you debug interactively and see how tool calls behave
- Apache 2.0 open source, with little concern about lock-in
⚠️ Cons
- It assumes you write Python; this is not a no-code agent builder
- No hosting or GUI admin console is provided, so operations must be arranged yourself
- Development continues in the 0.x line, and major updates can require migration work (the v0.10 release in August 2026 also came with a migration guide)
- Built by a small team, so support is more limited than with frameworks from large vendors
Comparison with Similar Services
| Criteria | fast-agent | LangGraph | CrewAI | OpenAI Agents SDK |
|---|---|---|---|---|
| Provider | llmindset.co.uk | LangChain | CrewAI | OpenAI |
| Core idea | MCP-native + workflow patterns | State graphs for control | Role-based agent collaboration | Handoffs and guardrails |
| MCP support | Wide slice of the spec | Supported (as an extension) | Supported | Supported |
| CLI/TUI | Built in | None (library-first) | CLI available | None (library-first) |
| License | Apache 2.0 | MIT | MIT | MIT |
Who Is It For
- Developers who want to validate their own MCP server from a real agent, including Elicitation and Sampling
- People who want multi-agent coordination — parallel execution, routing, evaluation loops — expressed concisely within a framework
- Terminal-centric users who prefer confirming agent behavior in a CLI/TUI rather than a GUI
- Anyone comparing output quality and cost across models while keeping the same agent definition
- Teams looking for an open source agent foundation they can run entirely inside their own environment
Summary
fast-agent is an easy framework to use as a testbed for MCP-centered agent development. Short decorator-based definitions, built-in workflow patterns and an interactive CLI/TUI cover everything from validating an MCP server to prototyping a multi-agent setup. It is not for people who want to build agents without code, but if you can write Python and intend to use MCP seriously, it is worth trying. The documentation and GitHub repository are thorough, so running the samples in the official guide is a good starting point.