Protocol Crosswalk (MCP ⇄ A2A ⇄ OpenAI ⇄ Anthropic ⇄ Google)¶

This page maps the canonical tool model to the dominant agent/tool protocols as of 2026-02-02. It exists to keep adapter work grounded in the latest specs and to clarify how we normalize divergent schemas.

Canonical Model (toolfoundation/adapter)¶

The stack uses CanonicalTool as the hub for adapter conversions. It captures: - Tool identity (namespace, name, version) - Display metadata (display name, summary, tags, category) - Input/Output schemas (JSON Schema superset) - Execution traits (deterministic, idempotent, streaming, input/output modes) - Security schemes + requirements - Source metadata for round-tripping

Adapter Hub-and-Spoke

Protocol Snapshot (Key Spec Facts)¶

MCP 2025-11-25: JSON-RPC 2.0 base protocol; servers expose resources, prompts, tools; clients may offer sampling, roots, elicitation; explicit user consent is required for data access and tool execution.
A2A (Agent2Agent): layered model with canonical data types, abstract operations, and protocol bindings (JSON-RPC, gRPC, HTTP/REST). The AgentCard advertises capabilities, security schemes, and skills.
OpenAI Agents SDK: tools can be hosted, local runtime, function tools, or agents-as-tools; function tools derive JSON schema from function signatures.
Anthropic tool use: tools are defined by name, description, input_schema; tool calls return tool_use and tool_result content blocks; MCP tools can be used by renaming inputSchema to input_schema.
Google Gemini function calling: tool definitions are function declarations with name, description, and parameters defined using a subset of the OpenAPI schema.

Crosswalk Table (Concept Mapping)¶

Concept	MCP	A2A	OpenAI Agents	Anthropic	Google Gemini
Tool identity	`name` (+ optional annotations)	`AgentSkill` + Task routing	Tool `name`	Tool `name`	Function `name`
Description	`description`	Skill description	Tool description / docstring	`description`	`description`
Input schema	`inputSchema` (JSON Schema)	Skill/task inputs (proto → binding)	JSON Schema from function signature	`input_schema` (JSON Schema)	`parameters` (OpenAPI subset)
Output schema	`outputSchema` (optional)	Task/Artifact outputs	Tool output (string/structured)	`tool_result` payload	Function response payload
Discovery	`search_tools`, `describe_tool`	AgentCard + task listing	Tool lists within Agents SDK	Tools array in request	Tools in request
Streaming	MCP streaming responses	Task/Artifact streaming (SSE/stream)	Model streaming events	Streaming tool use	Streaming responses
Security	User consent + transport auth	Security schemes + requirements	API key / tool scopes	API key + tool permissions	API key / IAM

Adapter Guidance¶

Use CanonicalTool as the source of truth and preserve vendor-specific fields in SourceMeta for round-trip fidelity.
Normalize schemas to JSON Schema internally and emit feature-loss warnings for protocols that do not support advanced schema features.
Treat A2A AgentCard as a tool provider descriptor and map A2A skills to canonical tools.
Use CanonicalProvider to model provider-level capabilities, security schemes, and default I/O modes.
Map OpenAI/Anthropic/Gemini function tools to canonical tools with InputSchema and OutputSchema, and surface tool categories as tags or backend hints.

Sources (URLs)¶

https://modelcontextprotocol.io/specification/2025-11-25
https://a2a-protocol.org/latest/specification/
https://agent2agent.info/specification/
https://openai.github.io/openai-agents-python/tools/
https://platform.openai.com/docs/guides/agents-sdk
https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview
https://ai.google.dev/gemini-api/docs/function-calling
https://docs.cloud.google.com/vertex-ai/generative-ai/docs/multimodal/function-calling