cagent

🤖 cagent 🤖

A powerful, easy-to-use, customizable multi-agent runtime that orchestrates AI agents with specialized capabilities and tools, and the interactions between agents.

✨ What is cagent? ✨

cagent lets you create and run intelligent AI agents, where each agent has specialized knowledge, tools and capabilities.

Think of it as allowing you to quickly build, share and run a team of virtual experts that collaborate to solve complex problems for you.

And it's dead easy to use!

⚠️ Note: cagent is in active development, breaking changes are to be expected ⚠️

Your First Agent

Example basic_agent.yaml:

Creating agents with cagent is straightforward. They are described in a short .yaml file, like this one:

agents:
  root:
    model: openai/gpt-5-mini
    description: A helpful AI assistant
    instruction: |
      You are a knowledgeable assistant that helps users with various tasks.
      Be helpful, accurate, and concise in your responses.

Run it in a terminal with cagent run basic_agent.yaml.

Many more examples can be found here!

Improving an agent with MCP tools

cagent supports MCP servers, enabling agents to use a wide variety of external tools and services.

It supports three transport types: stdio, http and sse.

Giving an agent access to tools via MCP is a quick way to greatly improve its capabilities, the quality of its results and its general usefulness.

Get started quickly with the Docker MCP Toolkit and catalog

Here, we're giving the same basic agent from the example above access to a containerized duckduckgo mcp server and its tools by using Docker's MCP Gateway:

agents:
  root:
    model: openai/gpt-5-mini
    description: A helpful AI assistant
    instruction: |
      You are a knowledgeable assistant that helps users with various tasks.
      Be helpful, accurate, and concise in your responses.
    toolsets:
      - type: mcp
        ref: docker:duckduckgo # stdio transport

When using a containerized server via the Docker MCP gateway, you can configure any required settings/secrets/authentication using the Docker MCP Toolkit in Docker Desktop.

Aside from the containerized MCP servers the Docker MCP Gateway provides, any standard MCP server can be used with cagent!

Here's an example similar to the above but adding read_file and write_file tools from the rust-mcp-filesystem MCP server:

agents:
  root:
    model: openai/gpt-5-mini
    description: A helpful AI assistant
    instruction: |
      You are a knowledgeable assistant that helps users with various tasks.
      Be helpful, accurate, and concise in your responses. Write your search results to disk.
    toolsets:
      - type: mcp
        ref: docker:duckduckgo
      - type: mcp
        command: rust-mcp-filesystem # installed with `cargo install rust-mcp-filesystem`
        args: ["--allow-write", "."]
        tools: ["read_file", "write_file"] # Optional: specific tools only
        env:
          - "RUST_LOG=debug"

See the USAGE docs for more detailed information and examples

Exposing agents as MCP tools

cagent can expose agents as MCP tools via the cagent mcp command, allowing other MCP clients to use your agents.

Each agent in your configuration becomes an MCP tool with its description.

# Start MCP server with local file
cagent mcp ./examples/dev-team.yaml

# Or use an OCI artifact
cagent mcp agentcatalog/pirate

This exposes each agent as a tool (e.g., root, designer, awesome_engineer) that MCP clients can call:

{
  "method": "tools/call",
  "params": {
    "name": "designer",
    "arguments": {
      "message": "Design a login page"
    }
  }
}

See MCP Mode documentation for detailed instructions on exposing your agents through MCP with Claude Desktop, Claude Code, and other MCP clients.

🎯 Key Features

• 🏗️ Multi-agent architecture - Create specialized agents for different domains.
• 🔧 Rich tool ecosystem - Agents can use external tools and APIs via the MCP protocol.
• 🔄 Smart delegation - Agents can automatically route tasks to the most suitable specialist.
• 📝 YAML configuration - Declarative model and agent configuration.
• 💭 Advanced reasoning - Built-in "think", "todo" and "memory" tools for complex problem-solving.
• 🔍 RAG (Retrieval-Augmented Generation) - Pluggable retrieval strategies (chunked_embeddings, BM25, more to come..) with hybrid retrieval, fusion, and result reranking support.
• 🌐 Multiple AI providers - Support for OpenAI, Anthropic, Gemini, xAI, Mistral, Nebius and Docker Model Runner.

🚀 Quick Start 🚀

Installation

Using Homebrew

Install cagent with a single command using homebrew!

$ brew install cagent

Using binary releases

Prebuilt binaries for Windows, macOS and Linux can be found on the release page of the project's GitHub repository

Once you've downloaded the appropriate binary for your platform, you may need to give it executable permissions. On macOS and Linux, this is done with the following command:

# linux amd64 build example
chmod +x /path/to/downloads/cagent-linux-amd64

You can then rename the binary to cagent and configure your PATH to be able to find it (configuration varies by platform).

Set your API keys

Based on the models you configure your agents to use, you will need to set the corresponding provider API key accordingly, all these keys are optional, you will likely need at least one of these, though:

# For OpenAI models
export OPENAI_API_KEY=your_api_key_here

# For Anthropic models
export ANTHROPIC_API_KEY=your_api_key_here

# For Gemini models
export GOOGLE_API_KEY=your_api_key_here

# For xAI models
export XAI_API_KEY=your_api_key_here

# For Nebius models
export NEBIUS_API_KEY=your_api_key_here

# For Mistral models
export MISTRAL_API_KEY=your_api_key_here

Run Agents!

# Run an agent!
cagent run ./examples/pirate.yaml

# or specify a different starting agent from the config, useful for agent teams
cagent run ./examples/pirate.yaml -a root

# or run directly from an image reference here I'm pulling the pirate agent from the creek repository
cagent run creek/pirate

Multi-agent team example

agents:
  root:
    model: claude
    description: "Main coordinator agent that delegates tasks and manages workflow"
    instruction: |
      You are the root coordinator agent. Your job is to:
      1. Understand user requests and break them down into manageable tasks
      2. Delegate appropriate tasks to your helper agent
      3. Coordinate responses and ensure tasks are completed properly
      4. Provide final responses to the user
      When you receive a request, analyze what needs to be done and decide whether to:
      - Handle it yourself if it's simple
      - Delegate to the helper agent if it requires specific assistance
      - Break complex requests into multiple sub-tasks
    sub_agents: ["helper"]

  helper:
    model: claude
    description: "Assistant agent that helps with various tasks as directed by the root agent"
    instruction: |
      You are a helpful assistant agent. Your role is to:
      1. Complete specific tasks assigned by the root agent
      2. Provide detailed and accurate responses
      3. Ask for clarification if tasks are unclear
      4. Report back to the root agent with your results

      Focus on being thorough and helpful in whatever task you're given.

models:
  claude:
    provider: anthropic
    model: claude-sonnet-4-0
    max_tokens: 64000

You'll find a curated list of agents examples, spread into 3 categories, Basic, Advanced and multi-agents in the /examples/ directory.

DMR (Docker Model Runner) provider options

When using the dmr provider, you can use the provider_opts key for DMR runtime-specific (e.g. llama.cpp/vllm) options and speculative decoding:

models:
  local-qwen:
    provider: dmr
    model: ai/qwen3
    max_tokens: 8192
    provider_opts:
      # general flags passed to the underlying model runtime
      runtime_flags: ["--ngl=33", "--repeat-penalty=1.2", ...] # or comma/space-separated string
      # speculative decoding for faster inference
      speculative_draft_model: ai/qwen3:1B
      speculative_num_tokens: 5
      speculative_acceptance_rate: 0.8

The default base_url cagent will use for DMR providers is http://localhost:12434/engines/llama.cpp/v1. DMR itself might need to be enabled via Docker Desktop's settings on macOS and Windows, and via the command-line on Docker CE on Linux.

See the DMR Provider documentation for more details on runtime flags and speculative decoding options.

RAG (Retrieval-Augmented Generation)

Give your agents access to your documents with cagent's modular RAG system. It supports multiple retrieval strategies that can be used individually or combined for hybrid search.

Quick RAG Example

models:
  embedder:
    provider: openai
    model: text-embedding-3-small

rag:
  my_knowledge_base:
    docs: [./documents, ./pdfs]
    strategies:
      - type: chunked-embeddings
        model: embedder
        threshold: 0.5
        chunking:
          size: 1000
          overlap: 100
    results:
      limit: 5

agents:
  root:
    model: openai/gpt-4o
    instruction: |
      You are an assistant with access to an internal knowledge base.
      Use the knowledge base to gather context before answering user questions
    rag: [my_knowledge_base]

Hybrid Retrieval (Chunked-Embeddings + BM25)

Combine semantic search (chunked-embeddings) with keyword search (BM25) for best results:

rag:
  hybrid_search:
    docs: [./shared_docs]
    
    strategies:
      - type: chunked-embeddings
        model: embedder
        threshold: 0.5
        limit: 20
        chunking:
          size: 1000
          overlap: 100
      
      - type: bm25
        k1: 1.5
        b: 0.75
        threshold: 0.3
        limit: 15
        chunking:
          size: 1000
          overlap: 100
    
    results:
      fusion:
        strategy: rrf  # Reciprocal Rank Fusion
        k: 60
      deduplicate: true
      limit: 5

agents:
  root:
    model: openai/gpt-4o
    rag: [hybrid_search]

Features:

• Multiple strategies: Vector (semantic), BM25 (keyword), or both
• Parallel execution: Strategies run concurrently for fast results
• Pluggable fusion: RRF, weighted, or max score combining
• Result reranking: Re-score results with specialized models for improved relevance
• Per-strategy configuration: Different thresholds, limits, and documents
• Auto file watching: Reindex automatically on file changes

Result Reranking

Improve search quality by re-scoring retrieved results with a reranking model:

rag:
  knowledge_base:
    docs: [./documents]
    strategies:
      - type: chunked-embeddings
        model: openai/text-embedding-3-small
        limit: 20  # Retrieve more candidates for reranking
    
    results:
      reranking:
        model: openai/gpt-4.1-mini   # Any chat model or DMR reranker
        top_k: 10                   # Only rerank top 10 (optional)
        threshold: 0.3              # Filter low-scoring results (optional)
        criteria: |                 # Domain-specific relevance guidance (optional, not used with DMR reranking specific models)
          Prioritize recent documentation and practical examples.
          Documents from official sources are more relevant.
      limit: 5  # Final top results after reranking

Supported providers: DMR (native /rerank endpoint), OpenAI, Anthropic, Gemini (via structured outputs)
Note: Temperature defaults to 0.0 for more deterministic scoring when not explicitly set.

See the RAG documentation in USAGE.md for complete details, examples, and debugging guides.

Quickly generate agents and agent teams with cagent new

Using the command cagent new you can quickly generate agents or multi-agent teams using a single prompt!
cagent has a built-in agent dedicated to this task.

To use the feature, you must have an Anthropic, OpenAI or Google API key available in your environment or specify a local model to run with DMR (Docker Model Runner).

You can choose what provider and model gets used by passing the --model provider/modelname flag to cagent new

If --model is unspecified, cagent new will automatically choose between these three providers in order based on the first api key it finds in your environment.

export ANTHROPIC_API_KEY=your_api_key_here  # first choice. default model claude-sonnet-4-0
export OPENAI_API_KEY=your_api_key_here     # if anthropic key not set. default model gpt-5-mini
export GOOGLE_API_KEY=your_api_key_here     # if anthropic and openai keys are not set. default model gemini-2.5-flash

--max-tokens can be specified to override the context limit used.
When using DMR, the default is 16k to limit memory usage. With all other providers the default is 64k

--max-iterations can be specified to override how many times the agent is allowed to loop when doing tool calling etc. When using DMR, the default is set to 20 (small local models have the highest chance of getting confused and looping endlessly). For all other providers, the default is 0 (unlimited).

Example of provider, model, context size and max iterations overriding:

# Use GPT-5 via OpenAI
cagent new --model openai/gpt-5

# Use a local model (ai/gemma3-qat:12B) via DMR
cagent new --model dmr/ai/gemma3-qat:12B

# Override the max_tokens used during generation, default is 64k, 16k when using the dmr provider
cagent new --model openai/gpt-5-mini --max-tokens 32000

# Override max_iterations to limit how much the model can loop autonomously when tool calling
cagent new --model dmr/ai/gemma3n:2B-F16 --max-iterations 15

$ cagent new

------- Welcome to cagent! -------
(Ctrl+C to stop the agent and exit)

What should your agent/agent team do? (describe its purpose):

> I need an agent team that connects to <some-service> and does...

Pushing and pulling agents from Docker Hub

cagent push

Agent configurations can be packaged and shared to Docker Hub using the cagent push command

cagent push ./<agent-file>.yaml namespace/reponame

cagent will automatically build an OCI image and push it to the desired repository using your Docker credentials

cagent pull

Pulling agents from Docker Hub is also just one cagent pull command away.

cagent pull creek/pirate

cagent will pull the image, extract the .yaml file and place it in your working directory for ease of use.

cagent run creek.yaml will run your newly pulled agent

Usage

More details on the usage and configuration of cagent can be found in USAGE.md

Telemetry

We track anonymous usage data to improve the tool. See TELEMETRY.md for details.

Contributing

Want to hack on cagent, or help us fix bugs and build out some features? 🔧

Read the information on how to build from source and contribute to the project in CONTRIBUTING.md

DogFooding: using cagent to code on cagent

A smart way to improve cagent's codebase and feature set is to do it with the help of a cagent agent!

We have one that we use and that you should use too:

cd cagent
cagent run ./golang_developer.yaml

This agent is an expert Golang developer specializing in the cagent multi-agent AI system architecture.

Ask it anything about cagent. It can be questions about the current code or about improvements to the code. It can also fix issues and implement new features!

Share your feedback

We’d love to hear your thoughts on this project. You can find us on Slack