fast-agent is a flexible way to interact with LLMs, excellent for use as a Coding Agent, Development Toolkit, Evaluation or Workflow platform.

To start an interactive session with shell support, install uv and run

uvx fast-agent-mcp@latest -x

To start coding with Hugging Face inference providers or use your OpenAI Codex plan:

# Code with Hugging Face Inference Providers
uvx fast-agent-mcp@latest --pack hf-dev

# Code with Codex (agents optimized for OpenAI)
uvx fast-agent-mcp@latest --pack codex

Enter a shell with !, or run shell commands e.g. ! cd web && npm run build.

Manage skills with the /skills command, and connect to MCP Servers with /connect. The default fast-agent registry contains skills to let you set up LSP, Agent and Tool Hooks, Compaction strategies, Automation and more.

# /connect supports stdio or streamable http (with OAuth)

# Start a STDIO server
/connect @modelcontextprotocol/server-everything

# Connect to a Streamable HTTP Server
/connect https://huggingface.co/mcp

It's recommended to install fast-agent to set up the shell aliases and other tooling.

# Install fast-agent
uv tool install -U fast-agent-mcp

# Run fast-agent with opus, shell support and subagent/smart mode
fast-agent --model opus -x --smart

Use local models with the generic provider, or automatically create the correct configuration for llama.cpp:

fast-agent model llamacpp

Any fast-agent setup or program can be used with any ACP client - the simplest way is to use fast-agent-acp:

# Run fast-agent inside Toad
toad acp "fast-agent-acp -x --model sonnet"

fast-agent enables you to create and interact with sophisticated multimodal Agents and Workflows in minutes. It is the first framework with complete, end-to-end tested MCP Feature support including Sampling and Elicitations.

fast-agent is CLI-first, with an optional prompt_toolkit-powered interactive terminal prompt (TUI-style input, completions, and in-terminal menus); responses can stream live to the terminal via rich without relying on full-screen curses UIs or external GUI overlays.

The simple declarative syntax lets you concentrate on composing your Prompts and MCP Servers to build effective agents.

Model support is comprehensive with native support for Anthropic, OpenAI and Google providers as well as Azure, Ollama, Deepseek and dozens of others via TensorZero. Structured Outputs, PDF and Vision support is simple to use and well tested. Passthrough and Playback LLMs enable rapid development and test of Python glue-code for your applications.

Recent features include:

• Agent Skills (SKILL.md)
• MCP-UI Support |
• OpenAI Apps SDK (Skybridge)
• Shell Mode
• Advanced MCP Transport Diagnsotics
• MCP Elicitations

fast-agent is the only tool that allows you to inspect Streamable HTTP Transport usage - a critical feature for ensuring reliable, compliant deployments. OAuth is supported with KeyRing storage for secrets. Use the fast-agent auth command to manage.

git clone --recurse-submodules https://github.com/evalstate/fast-agent.git

git submodule update --init --recursive

Prompts and configurations that define your Agent Applications are stored in simple files, with minimal boilerplate, enabling simple management and version control.

Chat with individual Agents and Components before, during and after workflow execution to tune and diagnose your application. Agents can request human input to get additional context for task completion.

Simple model selection makes testing Model <-> MCP Server interaction painless. You can read more about the motivation behind this project here

Start by installing the uv package manager for Python. Then:

uv pip install fast-agent-mcp          # install fast-agent!
fast-agent go                          # start an interactive session
fast-agent go --url https://hf.co/mcp  # with a remote MCP
fast-agent go --model=generic.qwen2.5  # use ollama qwen 2.5
fast-agent go --pack analyst --model haiku  # install/reuse a card pack and launch it
fast-agent scaffold                    # create an example agent and config files
uv run agent.py                        # run your first agent
uv run agent.py --model='o3-mini?reasoning=low'    # specify a model
uv run agent.py --transport http --port 8001  # expose as MCP server (server mode implied)
fast-agent quickstart workflow  # create "building effective agents" examples

--server remains available for backward compatibility but is deprecated; --transport now automatically switches an agent into server mode.

For packaged starter agents, use fast-agent go --pack <name> --model <model>. This installs the pack into the selected fast-agent environment if needed, then starts go normally. --model is a fallback for cards without an explicit model setting; a model declared directly in an AgentCard still wins.

Other quickstart examples include a Researcher Agent (with Evaluator-Optimizer workflow) and Data Analysis Agent (similar to the ChatGPT experience), demonstrating MCP Roots support.

Defining an agent is as simple as:

@fast.agent(
  instruction="Given an object, respond only with an estimate of its size."
)

We can then send messages to the Agent:

async with fast.run() as agent:
  moon_size = await agent("the moon")
  print(moon_size)

Or start an interactive chat with the Agent:

async with fast.run() as agent:
  await agent.interactive()

Here is the complete sizer.py Agent application, with boilerplate code:

import asyncio
from fast_agent import FastAgent

# Create the application
fast = FastAgent("Agent Example")

@fast.agent(
  instruction="Given an object, respond only with an estimate of its size."
)
async def main():
  async with fast.run() as agent:
    await agent.interactive()

if __name__ == "__main__":
    asyncio.run(main())

The Agent can then be run with uv run sizer.py.

Specify a model with the --model switch - for example uv run sizer.py --model sonnet.

Model strings also accept query overrides. For example:

• uv run sizer.py --model "gpt-5?reasoning=low"
• uv run sizer.py --model "claude-sonnet-4-6?web_search=on"
• uv run sizer.py --model "claude-sonnet-4-5?context=1m"

For Anthropic models, ?context=1m is only needed for earlier Sonnet 4 / Sonnet 4.5 models that still require the explicit 1M context opt-in. Claude Sonnet 4.6 and Claude Opus 4.6 already use their long context window by default, so ?context=1m is accepted for backward compatibility but is unnecessary there.

To generate examples use fast-agent quickstart workflow. This example can be run with uv run workflow/chaining.py. fast-agent looks for configuration files in the current directory before checking parent directories recursively.

Agents can be chained to build a workflow, using MCP Servers defined in the fastagent.config.yaml file:

@fast.agent(
    "url_fetcher",
    "Given a URL, provide a complete and comprehensive summary",
    servers=["fetch"], # Name of an MCP Server defined in fastagent.config.yaml
)
@fast.agent(
    "social_media",
    """
    Write a 280 character social media post for any given text.
    Respond only with the post, never use hashtags.
    """,
)
@fast.chain(
    name="post_writer",
    sequence=["url_fetcher", "social_media"],
)
async def main():
    async with fast.run() as agent:
        # using chain workflow
        await agent.post_writer("http://llmindset.co.uk")

All Agents and Workflows respond to .send("message") or .prompt() to begin a chat session.

Saved as social.py we can now run this workflow from the command line with:

uv run workflow/chaining.py --agent post_writer --message "<url>"

Add the --quiet switch to disable progress and message display and return only the final response - useful for simple automations.

MAKER (“Massively decomposed Agentic processes with K-voting Error Reduction”) wraps a worker agent and samples it repeatedly until a response achieves a k-vote margin over all alternatives (“first-to-ahead-by-k” voting). This is useful for long chains of simple steps where rare errors would otherwise compound.

• Reference: Solving a Million-Step LLM Task with Zero Errors
• Credit: Lucid Programmer (PR author)

@fast.agent(
  name="classifier",
  instruction="Reply with only: A, B, or C.",
)
@fast.maker(
  name="reliable_classifier",
  worker="classifier",
  k=3,
  max_samples=25,
  match_strategy="normalized",
  red_flag_max_length=16,
)
async def main():
  async with fast.run() as agent:
    await agent.reliable_classifier.send("Classify: ...")

The Agents As Tools workflow takes a complex task, breaks it into subtasks, and calls other agents as tools based on the main agent instruction.

This pattern is inspired by the OpenAI Agents SDK Agents as tools feature.

With child agents exposed as tools, you can implement routing, parallelization, and orchestrator-workers decomposition directly in the instruction (and combine them). Multiple tool calls per turn are supported and executed in parallel.

Common usage patterns may combine:

• Routing: choose the right specialist tool(s) based on the user prompt.
• Parallelization: fan out over independent items/projects, then aggregate.
• Orchestrator-workers: break a task into scoped subtasks (often via a simple JSON plan), then coordinate execution.

@fast.agent(
    name="NY-Project-Manager",
    instruction="Return NY time + timezone, plus a one-line project status.",
    servers=["time"],
)
@fast.agent(
    name="London-Project-Manager",
    instruction="Return London time + timezone, plus a one-line news update.",
    servers=["time"],
)
@fast.agent(
    name="PMO-orchestrator",
    instruction=(
        "Get reports. Always use one tool call per project/news. "  # parallelization
        "Responsibilities: NY projects: [OpenAI, Fast-Agent, Anthropic]. London news: [Economics, Art, Culture]. "  # routing
        "Aggregate results and add a one-line PMO summary."
    ),
    default=True,
    agents=["NY-Project-Manager", "London-Project-Manager"],  # orchestrator-workers
)
async def main() -> None:
    async with fast.run() as agent:
        await agent("Get PMO report. Projects: all. News: Art, Culture")

Extended example and all params sample is available in the repository as examples/workflows/agents_as_tools_extended.py.

For SSE and HTTP MCP servers, OAuth is enabled by default with minimal configuration. A local callback server is used to capture the authorization code, with a paste-URL fallback if the port is unavailable.

• Minimal per-server settings in fastagent.config.yaml:

mcp:
  servers:
    myserver:
      transport: http # or sse
      url: http://localhost:8001/mcp # or /sse for SSE servers
      auth:
        oauth: true # default: true
        redirect_port: 3030 # default: 3030
        redirect_path: /callback # default: /callback
        # scope: "user"       # optional; if omitted, server defaults are used

• The OAuth client uses PKCE and in-memory token storage (no tokens written to disk).
• Token persistence: by default, tokens are stored securely in your OS keychain via keyring. If a keychain is unavailable (e.g., headless container), in-memory storage is used for the session.
• To force in-memory only per server, set:

mcp:
  servers:
    myserver:
      transport: http
      url: http://localhost:8001/mcp
      auth:
        oauth: true
        persist: memory

• To disable OAuth for a specific server , set auth.oauth: false for that server.

The MCP ping utility can be enabled by either peer (client or server). See the Ping overview.

Client-side pinging is configured per server (default: 30s interval, 3 missed pings):

mcp:
  servers:
    myserver:
      ping_interval_seconds: 30 # optional; <=0 disables
      max_missed_pings: 3 # optional; consecutive timeouts before marking failed

The chain workflow offers a more declarative approach to calling Agents in sequence:

@fast.chain(
  "post_writer",
   sequence=["url_fetcher","social_media"]
)

# we can them prompt it directly:
async with fast.run() as agent:
  await agent.post_writer()

This starts an interactive session, which produces a short social media post for a given URL. If a chain is prompted it returns to a chat with last Agent in the chain. You can switch the agent to prompt by typing @agent-name.

Chains can be incorporated in other workflows, or contain other workflow elements (including other Chains). You can set an instruction to precisely describe it's capabilities to other workflow steps if needed.

Agents can request Human Input to assist with a task or get additional context:

@fast.agent(
    instruction="An AI agent that assists with basic tasks. Request Human Input when needed.",
    human_input=True,
)

await agent("print the next number in the sequence")

In the example human_input.py, the Agent will prompt the User for additional information to complete the task.

The Parallel Workflow sends the same message to multiple Agents simultaneously (fan-out), then uses the fan-in Agent to process the combined content.

@fast.agent("translate_fr", "Translate the text to French")
@fast.agent("translate_de", "Translate the text to German")
@fast.agent("translate_es", "Translate the text to Spanish")

@fast.parallel(
  name="translate",
  fan_out=["translate_fr","translate_de","translate_es"]
)

@fast.chain(
  "post_writer",
   sequence=["url_fetcher","social_media","translate"]
)

If you don't specify a fan-in agent, the parallel returns the combined Agent results verbatim.

parallel is also useful to ensemble ideas from different LLMs.

When using parallel in other workflows, specify an instruction to describe its operation.

Evaluator-Optimizers combine 2 agents: one to generate content (the generator), and the other to judge that content and provide actionable feedback (the evaluator). Messages are sent to the generator first, then the pair run in a loop until either the evaluator is satisfied with the quality, or the maximum number of refinements is reached. The final result from the Generator is returned.

If the Generator has use_history off, the previous iteration is returned when asking for improvements - otherwise conversational context is used.

@fast.evaluator_optimizer(
  name="researcher",
  generator="web_searcher",
  evaluator="quality_assurance",
  min_rating="EXCELLENT",
  max_refinements=3
)

async with fast.run() as agent:
  await agent.researcher.send("produce a report on how to make the perfect espresso")

When used in a workflow, it returns the last generator message as the result.

See the evaluator.py workflow example, or fast-agent quickstart researcher for a more complete example.

Routers use an LLM to assess a message, and route it to the most appropriate Agent. The routing prompt is automatically generated based on the Agent instructions and available Servers.

@fast.router(
  name="route",
  agents=["agent1","agent2","agent3"]
)

Look at the router.py workflow for an example.

Given a complex task, the Orchestrator uses an LLM to generate a plan to divide the task amongst the available Agents. The planning and aggregation prompts are generated by the Orchestrator, which benefits from using more capable models. Plans can either be built once at the beginning (plan_type="full") or iteratively (plan_type="iterative").

@fast.orchestrator(
  name="orchestrate",
  agents=["task1","task2","task3"]
)

See the orchestrator.py or agent_build.py workflow example.

All definitions allow omitting the name and instructions arguments for brevity:

@fast.agent("You are a helpful agent")          # Create an agent with a default name.
@fast.agent("greeter","Respond cheerfully!")    # Create an agent with the name "greeter"

moon_size = await agent("the moon")             # Call the default (first defined agent) with a message

result = await agent.greeter("Good morning!")   # Send a message to an agent by name using dot notation
result = await agent.greeter.send("Hello!")     # You can call 'send' explicitly

await agent.greeter()                           # If no message is specified, a chat session will open
await agent.greeter.prompt()                    # that can be made more explicit
await agent.greeter.prompt(default_prompt="OK") # and supports setting a default prompt

agent["greeter"].send("Good Evening!")          # Dictionary access is supported if preferred

@fast.agent(
  name="agent",                          # name of the agent
  instruction="You are a helpful Agent", # base instruction for the agent
  servers=["filesystem"],                # list of MCP Servers for the agent
  model="o3-mini?reasoning=high",        # specify a model for the agent
  use_history=True,                      # agent maintains chat history
  request_params=RequestParams(temperature= 0.7), # additional parameters for the LLM (or RequestParams())
  human_input=True,                      # agent can request human input
)

@fast.chain(
  name="chain",                          # name of the chain
  sequence=["agent1", "agent2", ...],    # list of agents in execution order
  instruction="instruction",             # instruction to describe the chain for other workflows
  cumulative=False,                      # whether to accumulate messages through the chain
  continue_with_final=True,              # open chat with agent at end of chain after prompting
)

@fast.parallel(
  name="parallel",                       # name of the parallel workflow
  fan_out=["agent1", "agent2"],          # list of agents to run in parallel
  fan_in="aggregator",                   # name of agent that combines results (optional)
  instruction="instruction",             # instruction to describe the parallel for other workflows
  include_request=True,                  # include original request in fan-in message
)

@fast.evaluator_optimizer(
  name="researcher",                     # name of the workflow
  generator="web_searcher",              # name of the content generator agent
  evaluator="quality_assurance",         # name of the evaluator agent
  min_rating="GOOD",                     # minimum acceptable quality (EXCELLENT, GOOD, FAIR, POOR)
  max_refinements=3,                     # maximum number of refinement iterations
)

@fast.router(
  name="route",                          # name of the router
  agents=["agent1", "agent2", "agent3"], # list of agent names router can delegate to
  model="o3-mini?reasoning=high",        # specify routing model
  use_history=False,                     # router maintains conversation history
  human_input=False,                     # whether router can request human input
)

@fast.orchestrator(
  name="orchestrator",                   # name of the orchestrator
  instruction="instruction",             # base instruction for the orchestrator
  agents=["agent1", "agent2"],           # list of agent names this orchestrator can use
  model="o3-mini?reasoning=high",        # specify orchestrator planning model
  use_history=False,                     # orchestrator doesn't maintain chat history (no effect).
  human_input=False,                     # whether orchestrator can request human input
  plan_type="full",                      # planning approach: "full" or "iterative"
  plan_iterations=5,                     # maximum number of full plan attempts, or iterations
)

@fast.maker(
  name="maker",                           # name of the workflow
  worker="worker_agent",                  # worker agent name
  k=3,                                    # voting margin (first-to-ahead-by-k)
  max_samples=50,                         # maximum number of samples
  match_strategy="exact",                 # exact|normalized|structured
  red_flag_max_length=256,                # flag unusually long outputs
  instruction="instruction",              # optional instruction override
)

@fast.agent(
  name="orchestrator",                    # orchestrator agent name
  instruction="instruction",              # orchestrator instruction (routing/decomposition/aggregation)
  agents=["agent1", "agent2"],            # exposed as tools: agent__agent1, agent__agent2
  max_parallel=128,                       # cap parallel child tool calls (OpenAI limit is 128)
  child_timeout_sec=600,                  # per-child timeout (seconds)
  max_display_instances=20,               # collapse progress display after top-N instances
)

Register Python functions as tools directly in code — no MCP server or external file needed. Both sync and async functions are supported. The function name and docstring are used as the tool name and description by default, or you can override them with name= and description=.

Per-agent tools (@agent.tool) — scope a tool to a specific agent:

@fast.agent(name="writer", instruction="You write things.")
async def writer(): ...

@writer.tool
def translate(text: str, language: str) -> str:
    """Translate text to the given language."""
    return f"[{language}] {text}"

@writer.tool(name="summarize", description="Produce a one-line summary")
def summarize(text: str) -> str:
    return f"Summary: {text[:80]}..."

Global tools (@fast.tool) — available to all agents that don't declare their own tools:

@fast.tool
def get_weather(city: str) -> str:
    """Return the current weather for a city."""
    return f"Sunny in {city}"

@fast.agent(name="assistant", instruction="You are helpful.")
# assistant gets get_weather (global @fast.tool)

Agents with @agent.tool or function_tools= only see their own tools — globals are not injected. Use function_tools=[] to explicitly opt out of globals with no tools.

Add Resources to prompts using either the inbuilt prompt-server or MCP Types directly. Convenience class are made available to do so simply, for example:

  summary: str =  await agent.with_resource(
      "Summarise this PDF please",
      "mcp_server",
      "resource://fast-agent/sample.pdf",
  )

LLM APIs have restrictions on the content types that can be returned as Tool Calls/Function results via their Chat Completions API's:

• OpenAI supports Text
• Anthropic supports Text and Image
• Google supports Text, Image, PDF, and Video (e.g., video/mp4).

  Note: Inline video data is limited to 20MB. For larger files, use the File API. YouTube URLs are supported directly.

For MCP Tool Results, ImageResources and EmbeddedResources are converted to User Messages and added to the conversation.

MCP Prompts are supported with apply_prompt(name,arguments), which always returns an Assistant Message. If the last message from the MCP Server is a 'User' message, it is sent to the LLM for processing. Prompts applied to the Agent's Context are retained - meaning that with use_history=False, Agents can act as finely tuned responders.

Prompts can also be applied interactively through the interactive interface by using the /prompt command.

Sampling LLMs are configured per Client/Server pair. Specify the model name in fastagent.config.yaml as follows:

mcp:
  servers:
    sampling_resource:
      command: "uv"
      args: ["run", "sampling_resource_server.py"]
      sampling:
        model: "haiku"

The documentation site is included as a submodule in docs/. To work with the docs locally:

# Install docs dependencies (first time only)
uv run scripts/docs.py install

# Generate reference docs from source code
uv run scripts/docs.py generate

# Run the dev server (http://127.0.0.1:8000)
uv run scripts/docs.py serve

# Or generate and serve in one command
uv run scripts/docs.py all

The generator extracts configuration field descriptions, model aliases, and API references directly from the source code to keep documentation in sync.

fast-agent builds on the mcp-agent project by Sarmad Qadri.

Contributions and PRs are welcome - feel free to raise issues to discuss. Full guidelines for contributing and roadmap coming very soon. Get in touch!