Interop matrix

For: application developers.

Use this matrix to check which protocol features were exercised with each client's own SDK. Every agent and flow is available as an MCP tool and an A2A agent. The rows report test results; they do not compare products or establish whether an unlisted client works.

make interop starts a dev server, a node and two agents: one answers and one asks a person a question. Each client runs with an operator's tenant API key, another operator's key and a key scoped to one agent. For the OAuth check, a person signs in through a fake identity provider and approves the client in the console; that check uses no API key.

Last run: 2026-10-21, on orchestrator-zero 5af15e15e (commit 5af15e15e, built 2026-10-20T23:05:25Z, go1.27.2 linux/amd64).

pass is a feature that did what the protocol and this server's documentation say; n/a is a feature the client does not offer (its protocol revision or its SDK has none), as its note says; fail did not work, and its note says how. An empty cell is a feature that client's driver does not run.

MCP clients

• handshake: server/discover for MCP 2026-07-28, initialize for 2025-11-25 and before.
• input required (MRTR): the agent asks a person a question; the client's elicitation handler answers it and the SDK retries the call with the answer.
• Tasks extension: tasks/get, tasks/update with the answer, and tasks/cancel.
• conversation: a call that sends back _meta["ai.oz0/conversation"] is the next turn; another key of the same tenant that sends it is told it is unknown.
• scoped key: a key scoped to one agent lists and calls only that agent.
• OAuth: no key at all: the client finds the tenant's authorization server from the 401, registers with a loopback redirect address, sends the person to the console, where they sign in through the identity provider and approve the client, redeems the code with PKCE, and then lists and calls the person's agents as them.
• input required (MRTR), n/a: 2025-11-25 has no input_required results; the question waits in the inbox (Python MCP SDK (mcp 1.x) 1.30.0, Python MCP SDK (mcp), legacy mode 2.3.0, TypeScript MCP SDK (@modelcontextprotocol/client), legacy mode 2.3.1, TypeScript MCP SDK (@modelcontextprotocol/sdk) 1.32.1).
• Tasks extension, n/a: the SDK has no client for the Tasks extension: it refuses a task as tools/call's result (TypeScript MCP SDK (@modelcontextprotocol/client) 2.3.1).
• Tasks extension, n/a: the Tasks extension needs 2026-07-28 (Python MCP SDK (mcp 1.x) 1.30.0, Python MCP SDK (mcp), legacy mode 2.3.0, TypeScript MCP SDK (@modelcontextprotocol/client), legacy mode 2.3.1, TypeScript MCP SDK (@modelcontextprotocol/sdk) 1.32.1).
• subscriptions/listen, n/a: subscriptions/listen is 2026-07-28's (Python MCP SDK (mcp 1.x) 1.30.0, Python MCP SDK (mcp), legacy mode 2.3.0, TypeScript MCP SDK (@modelcontextprotocol/client), legacy mode 2.3.1, TypeScript MCP SDK (@modelcontextprotocol/sdk) 1.32.1).

A2A clients

• signed card: the card's JWS (ES256 over its RFC 8785 form) verified with the key at its jku.
• contexts: a message with a task's contextId is the next turn; another key of the same tenant that sends it is told it is unknown.
• artifact hash: metadata["ai.oz0/sha256"] recomputed from the parts the client received.
• The server speaks A2A 1.0 only, and refuses a client of 0.3 (no A2A-Version: 1.0 header) with -32009.
• signed card, n/a: ADK resolves the card by its URL and does not check its signature (Google ADK (google-adk), RemoteA2aAgent 2.11.0).

Our A2A client

Outside A2A agents run through the A2A client of the verified plugin a2a-remote. make interop also runs it against a sample server built with each A2A SDK, on loopback, as a job's session would call an outside agent.

• card and interface: the card read, and its JSON-RPC interface on the card's own origin chosen.
• signed card: the card's signature (the SDK's own signer) verified with the key at its jku.
• SendMessage + GetTask: a task sent with returnImmediately and read until it answers.
• input required: the server's question answered, as a person answers it, and the task completed.
• CancelTask: a running task cancelled when the job is.

Agent frameworks as harnesses

The harness templates run an agent written for one of these frameworks as one session of a job. make interop-frameworks installs each template's example plugin on a dev server with two nodes and runs it against a fake model API behind the gateway, with the framework's own pinned release. Last run: 2026-10-21, on orchestrator-zero 5af15e15e (commit 5af15e15e, built 2026-10-20T23:05:25Z, go1.27.2 linux/amd64).

• turn: a job ran as one session; every model call went through the gateway, in the framework's own API, and was metered under the job.
• tool call asked and answered: each of the framework's tool calls was asked as <harness>.<tool>; a read ran, a write waited for a person's approval and ran once approved.
• quota resume: a quota refused the session's second model call; once it was raised, the session was started again with its session ID and went on from its conversation.
• budget refusal: the gateway's refusal stopped the session's call, and the job waited for a budget decision instead of failing.
• lost node: the node was killed under a session; the restart waited for a person, and a fresh session ran from scratch on the other node.
• Each is governed and metered, and durable only as one step: a lost node starts the session again from scratch, and that restart asks a person first unless lint proves the agent's tools only read.

Not tested

Product names belong to their owners and are used here only to say which client or framework was run; a row says what we ran, not that the owner endorses it.