A2A MCP Server

An MCP server that implements an A2A Client for the A2A Protocol. The server can be used to connect and send messages to A2A Servers (remote agents).

The server needs to be initialised with one or more Agent Card URLs, each of which can have custom headers for authentication, configuration, etc.

All agents (name and description) can be viewed with the get_agents tool, an agent's skills (name and description) can be viewed with the get_agent tool, Messages can be sent to the agents with the send_message tool, long-running Tasks can be polled with get_task, and Artifacts can be viewed with view_text_artifact and view_data_artifact tools.

✨ Features

• 6 MCP tools — get_agents, get_agent, send_message, get_task, view_text_artifact, and view_data_artifact for communicating with A2A agents
• Simple message sending — send messages to any A2A agent by ID; Agent Card fetching, headers, and non-blocking streaming are handled automatically
• Multi-turn conversations — continue conversations across multiple messages using context_id
• Long-running task support — if send_message times out, use get_task to monitor the task until it reaches a terminal state
• Automatic artifact minimization — large text and data artifacts are automatically minimized for LLM context windows, with dedicated tools for detailed navigation
• Task and file storage — tasks and file artifacts are saved locally to ~/.a2a-mcp/ by default
• Custom headers and authentication — configure per-agent custom headers for API keys and other credentials
• Configurable timeouts and limits — customisable timeouts, polling intervals, and character limits via environment variables

📋 Requirements

To run the server you need to install uv if you haven't already.

MacOS/Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

🚀 Quick Start

1. Download Claude for Desktop
2. Add the below to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "a2a": {
      "command": "uvx",
      "args": ["a2anet-mcp"],
      "env": {
        "A2A_MCP_AGENT_CARDS": "{\"tweet-search\": {\"url\": \"https://example.com/.well-known/agent-card.json\"}}"
      }
    }
  }
}

Tip: If you don't have an Agent Card URL, see: A2A Net Demo

⚙️ Configuration

All configuration is via environment variables prefixed with A2A_MCP_.

A2A_MCP_AGENT_CARDS (required)

A JSON object mapping agent IDs to their configuration. Each agent must have a url key with the full path to the Agent Card. It can optionally have a custom_headers key with an object in the form {"header": "value"}:

export A2A_MCP_AGENT_CARDS='{
  "tweet-search": {
    "url": "https://example.com/.well-known/agent-card.json",
    "custom_headers": {"X-API-Key": "your-key"}
  }
}'

Optional settings

Env Var	Default	Description
A2A_MCP_TASK_STORE	true	Enable task persistence via JSONTaskStore
A2A_MCP_FILE_STORE	true	Enable file artifact storage via LocalFileStore
A2A_MCP_SEND_MESSAGE_CHARACTER_LIMIT	50000	Character limit for artifact minimization in send_message
A2A_MCP_MINIMIZED_OBJECT_STRING_LENGTH	5000	Max string length when minimizing objects
A2A_MCP_VIEW_ARTIFACT_CHARACTER_LIMIT	50000	Character limit for view_text_artifact / view_data_artifact
A2A_MCP_AGENT_CARD_TIMEOUT	15	Timeout in seconds for fetching agent cards
A2A_MCP_SEND_MESSAGE_TIMEOUT	60	Timeout in seconds for send_message
A2A_MCP_GET_TASK_TIMEOUT	60	Timeout in seconds for get_task
A2A_MCP_GET_TASK_POLL_INTERVAL	5	Interval in seconds between get_task polls

🛠️ Tools

get_agents

Get all agent names and descriptions.

get_agent

Get an agent's name, description, and skill names and descriptions.

Parameter	Required	Description
agent_id	Yes	Agent ID

send_message

Send a message to an agent.

Parameter	Required	Description
agent_id	Yes	Agent ID from get_agents
message	Yes	Your message or request
context_id	No	Continue an existing conversation
task_id	No	Task ID for input_required flows
timeout	No	Override HTTP timeout (seconds)

get_task

Get the current state of a task. Monitors until terminal/actionable state or timeout. On timeout, returns the current task state (which may still be non-terminal).

Parameter	Required	Description
agent_id	Yes	Agent ID that owns the task
task_id	Yes	Task ID from a previous send_message
timeout	No	Override monitoring timeout (seconds)
poll_interval	No	Override interval between polls (seconds)

view_text_artifact

View text content from an artifact with optional line or character range selection.

Parameter	Required	Description
agent_id	Yes	Agent ID that produced the artifact
task_id	Yes	Task ID containing the artifact
artifact_id	Yes	Artifact to view
line_start	No	Starting line number (1-based, inclusive)
line_end	No	Ending line number (1-based, inclusive)
character_start	No	Starting character index (0-based)
character_end	No	Ending character index (0-based)

view_data_artifact

View structured data from an artifact with optional filtering.

Parameter	Required	Description
agent_id	Yes	Agent ID that produced the artifact
task_id	Yes	Task ID containing the artifact
artifact_id	Yes	Artifact to view
json_path	No	Dot-separated path to extract specific fields
rows	No	Row selection (index, list, range string, or "all")
columns	No	Column selection (name, list, or "all")

📖 Examples

List agents

get_agents({})

{
  "tweet-search": {
    "name": "Tweet Search",
    "description": "Find and analyze tweets by keyword, URL, author, list, or thread. Filter by language, media type, engagement, date range, or location. Get a clean table of tweets with authors, links, media, and counts; then refine the table and generate new columns with AI."
  }
}

Get agent details

get_agent({
  "agent_id": "tweet-search"
})

{
  "name": "Tweet Search",
  "description": "Find and analyze tweets by keyword, URL, author, list, or thread. Filter by language, media type, engagement, date range, or location. Get a clean table of tweets with authors, links, media, and counts; then refine the table and generate new columns with AI.",
  "skills": [
    {
      "name": "Search Tweets",
      "description": "Search X by keywords, URLs, handles, or conversation IDs. Filter by engagement (retweets/favorites/replies), dates, language, location, media type (images/videos/quotes), user verification status, and author/reply/mention relationships. Sort by Top or Latest. Return 1-10,000 results."
    },
    ...,
    {
      "name": "Generate Table",
      "description": "Generate a new table from any table with AI. Explain what table you want to generate from, what columns you want to keep, and what new columns you want to generate."
    }
  ]
}

Send a message

send_message({
  "agent_id": "tweet-search",
  "message": "Find tweets about AI from today (January 12, 2026)"
})

{
  "id": "tsk-123",
  "context_id": "ctx-123",
  "kind": "task",
  "status": {
    "state": "completed",
    "message": {
      "context_id": "ctx-123",
      "kind": "message",
      "parts": [
        {
          "kind": "text",
          "text": "I found 10 tweets about \"AI\" posted on January 12, 2026. The search parameters used were:\n\n- Search Terms: AI\n- Start Date: 2026-01-12\n- End Date: 2026-01-13\n- Maximum Items: 10\n\nWould you like to see more tweets, or do you want a summary or analysis of these results?"
        }
      ]
    }
  },
  "artifacts": [
    {
      "artifact_id": "art-123",
      "description": "Tweets about AI posted on January 12, 2026.",
      "name": "AI Tweets from January 12, 2026",
      "parts": [
        {
          "kind": "data",
          "data": {
            "records": {
              "_total_rows": 10,
              "_columns": [
                {
                  "count": 1,
                  "unique_count": 1,
                  "types": [
                    {
                      "name": "int",
                      "count": 1,
                      "percentage": 100.0,
                      "sample_value": 213,
                      "minimum": 213,
                      "maximum": 213,
                      "average": 213
                    }
                  ],
                  "name": "quote.author.mediaCount"
                },
                ...,
                {
                  "count": 10,
                  "unique_count": 1,
                  "types": [
                    {
                      "name": "bool",
                      "count": 10,
                      "percentage": 100.0,
                      "sample_value": false
                    }
                  ],
                  "name": "isPinned"
                }
              ]
            },
            "_tip": "Data was minimized. Call view_data_artifact() to navigate to specific data."
          }
        }
      ]
    }
  ]
}

Handle a long-running task

If the remote agent takes longer than A2A_MCP_SEND_MESSAGE_TIMEOUT (default: 60 seconds), send_message returns the task in its current state:

send_message({
  "agent_id": "tweet-search",
  "message": "Find tweets about AI from today (January 12, 2026)"
})

{
  "id": "tsk-123",
  "context_id": "ctx-123",
  "kind": "task",
  "status": {
    "state": "working",
    "message": null
  },
  "artifacts": []
}

Use get_task to check progress:

get_task({
  "agent_id": "tweet-search",
  "task_id": "tsk-123"
})

When complete, the response matches the format shown in Send a message. If still working, call get_task again to continue monitoring.

Multi-turn conversation

Use context_id to continue a conversation:

send_message({
  "agent_id": "tweet-search",
  "message": "Can you summarize each of the 10 tweets in the table in 3-5 words each? Just give me a simple list with the author name and summary.",
  "context_id": "ctx-123"
})

{
  "id": "tsk-456",
  "context_id": "ctx-123",
  "kind": "task",
  "status": {
    "state": "completed",
    "message": {
      "context_id": "ctx-123",
      "kind": "message",
      "parts": [
        {
          "kind": "text",
          "text": "Here is a simple list of each tweet's author and a 3-5 word summary:\n\n1. alienofeth – Real-time STT intent detection\n2. UnderdogEth_ – AI ownership discussion thread\n3. Count_Down_000 – Learning new vocabulary word\n4. ThaJonseBoy – AI and market predictions\n5. Evelyn852422353 – AI model comparison debate\n6. SyrilTchouta – Language learning with AI\n7. cx. – AI in marketing insights\n8. Halosznn_ – Graphic design course shared\n9. xmaquina – AI smarter models discussion\n10. Flagm8_ – AI and business strategy\n\nLet me know if you want more details or a different format!"
        }
      ]
    }
  },
  "artifacts": [
    {
      "artifact_id": "art-456",
      "description": "A simple list of each tweet's author and a 3-5 word summary of the tweet content.",
      "name": "AI Tweet Summaries 3-5 Words",
      "parts": [
        {
          "kind": "data",
          "data": {
            "records": [
              {
                "author.userName": "ai_q2_",
                "summary": "Possibly understand"
              },
              ...,
              {
                "author.userName": "CallStackTech",
                "summary": "Real-time STT intent detection"
              }
            ]
          }
        }
      ]
    }
  ]
}

View data artifact

view_data_artifact({
  "agent_id": "tweet-search",
  "task_id": "tsk-123",
  "artifact_id": "art-123",
  "json_path": "records",
  "rows": "all",
  "columns": ["author.userName", "text"]
})

{
  "artifact_id": "art-123",
  "description": "Tweets about AI posted on January 12, 2026.",
  "name": "AI Tweets from January 12, 2026",
  "parts": [
    {
      "kind": "data",
      "data": [
        {
          "author.userName": "ai_q2_",
          "text": "@nyank_x わかるかもしれない"
        },
        ...,
        {
          "author.userName": "CallStackTech",
          "text": "Just built a real-time STT pipeline that detects intent faster than you can say \"Hello!\" 🎤✨ Discover how I used Deepgram to achieve su...\n\n🔗 https://t.co/dgbvdlATZ0\n\n#VoiceAI #AI #BuildInPublic"
        }
      ]
    }
  ]
}

💾 Data Storage

Tasks and file artifacts are persisted locally at ~/.a2a-mcp/:

• Tasks: ~/.a2a-mcp/tasks/
• Files: ~/.a2a-mcp/files/

Both can be disabled via environment variables (A2A_MCP_TASK_STORE=false, A2A_MCP_FILE_STORE=false).

🔧 Development

Claude Desktop Setup

For local development:

1. Clone the repository: git clone https://github.com/a2anet/a2a-mcp.git
2. Download Claude for Desktop.
3. Add to the below to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "a2a": {
      "command": "uv",
      "args": ["--directory", "/path/to/a2a-mcp", "run", "a2anet-mcp"],
      "env": {
        "A2A_MCP_AGENT_CARDS": "{\"tweet-search\": {\"url\": \"https://example.com/.well-known/agent-card.json\"}}"
      }
    }
  }
}

📄 License

a2anet is distributed under the terms of the Apache-2.0 license.

🤝 Join the A2A Net Community

A2A Net is a site to find and share AI agents and open-source community. Join to share your A2A agents, ask questions, stay up-to-date with the latest A2A news, be the first to hear about open-source releases, tutorials, and more!

• 🌍 Site: A2A Net
• 🤖 Discord: Join the Discord