Agent-to-Agent Bridge

MCP server for agent-to-agent communication -- capability discovery, task delegation, and result aggregation across MCP agents.

MCP connects agents to tools, but not to each other. This server adds a standardized agent-to-agent layer within the MCP protocol -- register agents, discover capabilities, delegate tasks, and broadcast work to multiple agents. Inspired by Google's A2A protocol.

Install

npx a2a-bridge-mcp

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "a2a-bridge": {
      "command": "npx",
      "args": ["a2a-bridge-mcp"]
    }
  }
}

From source

git clone https://github.com/mdfifty50-boop/a2a-bridge-mcp.git
cd a2a-bridge-mcp
npm install
node src/index.js

Tools

register_agent

Register an agent with capabilities for discovery by other agents.

Param	Type	Default	Description
agent_id	string	required	Unique agent identifier
capabilities	string[]	required	Capability strings (e.g. ["code-review", "summarization"])
description	string	""	Human-readable description
input_schema	object	{}	JSON schema for accepted input
output_schema	object	{}	JSON schema for produced output
endpoint	string	""	Optional endpoint or transport hint

discover_agents

Find agents matching a capability need with fuzzy text similarity scoring.

Param	Type	Default	Description
capability_needed	string	required	Capability to search for
min_score	number	0.3	Minimum similarity score (0-1)

Returns scored matches sorted by relevance.

delegate_task

Delegate a task to a specific registered agent. Creates a tracked task with status lifecycle.

Param	Type	Default	Description
from_agent	string	required	Delegating agent ID
to_agent	string	required	Target agent ID
task_description	string	required	What the target should do
input_data	object	{}	Input data for the task
timeout_ms	number	30000	Timeout in milliseconds

Returns a task_id for tracking.

get_task_result

Get status and result of a delegated task. Also used by executing agents to submit results.

Param	Type	Description
task_id	string	Task ID from delegate_task or broadcast_task
submit_result	object	(Optional) Submit completion result
submit_error	string	(Optional) Submit failure error

Status lifecycle: pending -> running -> completed / failed.

broadcast_task

Send a task to ALL agents matching a capability. Creates individual tracked tasks for each match.

Param	Type	Default	Description
from_agent	string	required	Broadcasting agent ID
capability_needed	string	required	Capability to match
task_description	string	required	What matched agents should do
input_data	object	{}	Input data
min_score	number	0.3	Minimum match score
timeout_ms	number	30000	Timeout per agent

Returns list of task IDs for aggregation via get_task_result.

get_agent_card

Get an agent's full capability card with schemas, stats, and task history.

Param	Type	Description
agent_id	string	Agent identifier

Returns capabilities, input/output schemas, success rate, and task counts. Inspired by Google A2A agent cards.

list_agents

List all registered agents, optionally filtered by capability.

Param	Type	Default	Description
filter	string	""	Capability keyword to filter by (empty = all)

Resources

URI	Description
a2a://agents	All registered agents with capabilities and stats

Usage Pattern

1. register_agent    -- each agent registers at startup
2. discover_agents   -- find who can handle a task
3. delegate_task     -- send work to a specific agent
   OR broadcast_task -- send work to all matching agents
4. get_task_result   -- poll for completion or submit results
5. get_agent_card    -- inspect an agent's full profile
6. list_agents       -- overview of the agent network

Multi-agent workflow example

Agent A (orchestrator):
  1. register_agent(agent_id="orchestrator", capabilities=["planning", "coordination"])
  2. discover_agents(capability_needed="code review")
     -> finds Agent B (score: 0.95)
  3. delegate_task(from="orchestrator", to="agent-b", task="Review PR #42")
     -> task_id: "task_1234"
  4. get_task_result(task_id="task_1234")
     -> status: "completed", result: { approved: true, comments: [...] }

Agent B (worker):
  1. register_agent(agent_id="agent-b", capabilities=["code-review", "linting"])
  2. (receives task via external notification or polling)
  3. get_task_result(task_id="task_1234", submit_result={ approved: true })

License

MIT