Standard A2A protocol bridge for OpenClaw — agent card, JSON-RPC endpoint, and remote agent calling
The A2A Bridge server defines only ONE tool (a2a_call_remote) with a basic description and input schema. However, critical gaps in parameter descriptions, lack of output schema documentation, missing error handling guidance, and absent tool annotations significantly limit quality. The tool name is clear and action-oriented, but parameter documentation is minimal. The server is designed for agent-to-agent JSON-RPC bridging, not as a full MCP server, it lacks prompts, resources, and most standard MCP features. The single tool's implementation is functional but falls short of production-grade standards for LLM-driven tool selection and error recovery.
Call a remote A2A agent by name. Fetches the agent's card from .well-known/agent-card.json, then sends a message via JSON-RPC to the agent's URL.
Output schema is not documented. Callers (LLMs) do not know what fields to expect in the response, forcing them to guess the structure or make additional calls to discover available fields.
Parameter descriptions are generic and lack constraints. 'agentName' description says it 'must be configured in remoteAgents' but does not specify the format, valid values, or how to discover available agent names. 'message' lacks guidance on expected length, format, or content type.
Error handling does not guide recovery. The tool can fail (network timeout, missing agent card, JSON-RPC error) but error responses do not suggest alternative actions (e.g., 'Agent not reachable. Check if the agent URL is correct.' or 'Agent returned an error. Try simplifying your message.').
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | <=2025-11-25 | v2 |
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). The risk is marked READ_ONLY in metadata, but there is no structured annotation in the tool definition itself. This limits agent reasoning about side effects and retry safety.
No pagination or result limits documented. If a remote agent returns a large response, the tool does not cap or paginate results, risking token exhaustion in the LLM context.
Tool description lacks 'when to use' guidance. Callers do not understand when to invoke a2a_call_remote versus calling local tools or other agents. The description does not clarify whether this bridges to external A2A agents only or internal OpenClaw agents.