A2A to MCP bridge

io.github.AmirK-S由 GitHub 帳號 AmirK-S 發布v0.1.0更新於 Oct 25, 2026

A2A v1.0.1 agents as MCP 2026-07-28 tools, with the tasks extension

已驗證STDIO僅桌面開發者工具AI 與機器學習

概覽

AI 產生的概覽

一個本機橋接服務,把 A2A v1.0 代理程式暴露為 MCP 工具,讓助理探索代理程式、傳送訊息並讀取或取消任務。

可以做什麼
探索已設定的 A2A 代理程式,經結構檢查後回傳其代理卡片,並在中介資料中附上文字摘要與卡片位址。向代理程式傳送訊息,可透過情境代碼延續對話,或透過任務代碼回覆等待輸入的任務。讀取任務封裝與其歷程記錄,或要求取消任務。把 A2A 的部件、產出物、任務狀態與具型別錯誤轉換為 MCP 內容與結構化結果,並註明轉換中遺失的資訊。在現代協定路由上透過 MCP 任務擴充支援長時間任務,在舊版路由上則以內嵌方式等待結果。
適用情境
當你已經執行或能夠連線 A2A 代理程式,並希望 Claude Code、Cursor 等 MCP 用戶端把它們當成一般工具呼叫時使用。適合多輪代理程式對話,以及需要輪詢、取消或補充輸入的長時間任務。
執行需求
需要 Node 20 或更新版本;可全域安裝 npm 套件或從儲存庫建置。需要一個設定檔,依基礎位址或代理卡片位址列出一或多個代理程式,並指定連接埠。橋接服務在回送 HTTP 端點上監聽,供 MCP 用戶端連線;未宣告任何驗證或環境變數。
安裝前請注意
兩端都沒有驗證,任何能連上該回送連接埠的程式都能呼叫已設定的代理程式;除非放寬 allowedHosts,否則只接受回送主機。訊息會傳送給第三方 A2A 代理程式,任務代碼預設在記憶體中保留約十五分鐘。取消操作會改變代理程式端的任務狀態。AUTH_REQUIRED 會連同代理程式文字以失敗回傳,而不是作為憑證提示。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 A2A to MCP bridge,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

a2a-to-mcp

A2A v1.0.1 agents exposed as MCP tools on a stateless MCP 2026-07-28 server, with the io.modelcontextprotocol/tasks extension.

Point the bridge at one or more A2A agents by their agent card. Any MCP client, Claude Code or Cursor among them, then sees four ordinary tools: discover an agent, send it a message, read a task, cancel a task. The bridge translates parts, artifacts, task states and errors between the two protocols and says out loud what it cannot translate.

SurfaceVersion servedPinned dependency
MCP, modern route2026-07-28 via server/discover, no session, per-request _meta@modelcontextprotocol/server 2.0.0
MCP, legacy route2025-11-25 via initialize, same endpointsame
MCP tasks extensionio.modelcontextprotocol/tasks schema 2026-07-28, modern route onlynone, served by the bridge
A2A1.0 (A2A-Version: 1.0), JSON-RPC binding@a2a-js/sdk 1.1.0

Node 20 or later. Apache-2.0.

Quick start

sh
npm install -g a2a-to-mcp

Or run it from the repository:

sh
git clone https://github.com/AmirK-S/a2a-to-mcp.git && cd a2a-to-mcp && npm ci && npm run build

Write agents.json, one alias per agent, by base URL or by card URL:

json
{  "agents": {    "hello": { "url": "http://localhost:9999" },    "planner": { "cardUrl": "https://planner.example.org/.well-known/agent-card.json" }  },  "port": 8931}

Run it:

sh
node dist/cli.js --config agents.json

The bridge listens on http://127.0.0.1:8931/mcp. Only loopback hosts are accepted in the Host and Origin headers unless allowedHosts is set in the configuration. There is no authentication in this version, on either side.

Claude Code

Add to .mcp.json at the root of the project, which is the block the bridge prints on startup:

json
{  "mcpServers": {    "a2a-to-mcp": {      "type": "http",      "url": "http://127.0.0.1:8931/mcp"    }  }}

Claude Code negotiates 2026-07-28 through server/discover since its v2 runtime, documented from 2.1.232 and measured here with 2.1.263. Older versions fall back to initialize and get the same four tools on the legacy route. A full annotated trace of a Claude Code session against the bridge, discovery, a task with an artifact and an elicitation round trip, is in docs/DEMO-CLAUDE-CODE.md. In headless mode (claude -p) there is nobody to answer an elicitation, so Claude Code cancels it and the bridge cancels the A2A task; the interactive mode asks the user.

Cursor

Add to .cursor/mcp.json:

json
{  "mcpServers": {    "a2a": { "url": "http://localhost:8931/mcp" }  }}

Cursor speaks 2025-11-25 and is served through initialize on the same endpoint.

The four tools

Every tool takes the agent alias as its first argument when several agents are configured; with a single agent the alias may be omitted and defaults to it.

ToolArgumentsReturns
a2a_discoveragentthe agent card as the agent served it, after a structural check, in structuredContent, with a text summary and the card URL in _meta
a2a_send_messageagent, text, contextHandle?, taskHandle?the agent reply, see the result contract below
a2a_get_taskagent, taskHandle, historyLength?the task envelope with its history
a2a_cancel_taskagent, taskHandlethe task envelope after the cancellation request

Calling a tool on the wire

The argument that carries the message is text. A complete tools/call on the 2026-07-28 route, as an MCP client sends it:

sh
curl -sS http://127.0.0.1:8931/mcp \  -H 'Content-Type: application/json' \  -H 'Accept: application/json, text/event-stream' \  -H 'MCP-Protocol-Version: 2026-07-28' \  -H 'Mcp-Method: tools/call' \  -H 'Mcp-Name: a2a_send_message' \  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{        "_meta":{"io.modelcontextprotocol/protocolVersion":"2026-07-28",                 "io.modelcontextprotocol/clientCapabilities":{}},        "name":"a2a_send_message",        "arguments":{"agent":"hello","text":"task: hello world"}}}'

A run against a public A2A agent, with the one mistake a first user makes, is in docs/USAGE-REEL.md.

Result contract

Each reply is a CallToolResult whose content holds the translated parts and whose structuredContent is an envelope:

json
{  "kind": "task",  "contextHandle": "cx_...",  "taskHandle": "tk_...",  "a2aState": "TASK_STATE_COMPLETED",  "status": "completed",  "note": "present when the mapping lost information",  "artifacts": [{ "artifactId": "...", "name": "result" }],  "data": {}}

data holds the single A2A data part of the artifacts, and is absent when there is none or more than one. history is added by a2a_get_task only.

kind is message for a direct A2A message and task when the agent opened a task. Handles are opaque strings minted by the bridge, never A2A identifiers; they hide the bridge table, not what an agent chooses to write in its own messages. They live in memory for fifteen minutes after their last use by default (handleTtlMs), and an expired or unknown handle returns an explicit error. Pass contextHandle back to stay in the same conversation, and taskHandle back to answer a task the agent left waiting for input.

An agent failure (FAILED, REJECTED, AUTH_REQUIRED), and a task the agent left in the unspecified state, is a tool execution error: isError: true, the agent text in content, the A2A state and a note in the envelope. A typed A2A error (codes -32001 to -32009) is also a tool execution error, with the code in structuredContent.a2aErrorCode. The bridge never re-emits an A2A error code as a JSON-RPC code. JSON-RPC reserves -32000 to -32099 for server-defined errors, MCP already uses -32020, -32021 and -32022 in that band, and an A2A code replayed there would claim an MCP meaning it does not have.

Long-running tasks

With the tasks extension

A client that declares io.modelcontextprotocol/tasks in the per-request clientCapabilities.extensions gets a CreateTaskResult from a2a_send_message as soon as the agent opens a task, then polls tasks/get and may call tasks/update and tasks/cancel. The bridge consumes SendStreamingMessage in the background when the card announces streaming, and GetTask otherwise.

A2A TaskStateMCP TaskStatus on tasks/getWhat is lost
SUBMITTEDworkingMCP has no accepted but not started state; statusMessage says so
WORKINGworkingnothing
INPUT_REQUIREDinput_required with one synthesized form elicitationA2A asks in free text; the form has a single string field
COMPLETEDcompleted, result is the full CallToolResultnothing
FAILED, REJECTED, AUTH_REQUIREDcompleted, result.isError is truea refusal and an authorization request become indistinguishable from a failure without the envelope
CANCELEDcancellednothing, mind the spelling
unspecifiedcompleted, result.isError is trueA2A sent no lifecycle at all; the bridge does not guess one
bridge cannot reach the task any morefailed with a JSON-RPC error, A2A code in error.data

AUTH_REQUIRED is deliberately not mapped to input_required: MCP forbids form elicitation for credentials, and A2A asks for out-of-band authorization. The agent text, which usually carries the URL to visit, is returned in the result.

The extension is served on the 2026-07-28 route only. On the legacy route, tasks/* answer -32021 with the required capability. As of 2026-09-06, the two clients measured against this bridge, Claude Code and Cursor, do not declare the extension. It is exercised by the test suite in this repository and by any client that opts in per request.

CreateTaskResult leaves the bridge with an empty content array in addition to the fields the extension defines. That is a workaround for the reference SDK, which validates every tools/call result against the plain result schema (typescript-sdk issue 2637). The three extension methods are served by a small router in front of the SDK on the modern route, because the SDK answers -32601 to tasks/get and tasks/cancel before consulting its handlers (typescript-sdk issue 2598), and splitting the three across two layers would let them drift. Both workarounds are isolated and will be removed when the SDK moves.

Without the extension

a2a_send_message waits for the agent to reach a terminal or interrupted state and answers inline. A task left in INPUT_REQUIRED comes back:

  • as a multi-round-trip input_required result with one form elicitation, for a client that declares elicitation, which then replays the call with the answer;
  • as an ordinary result whose envelope says input_required and carries the taskHandle, for any other client, which then sends the answer with that handle.

What crosses the bridge, and what does not

Parts:

A2A PartMCP content block
textTextContent
dataTextContent holding the JSON, flagged in _meta; also structuredContent.data when it is the only data part of the task artifacts
urlResourceLink with filename as name and mediaType as mimeType
raw, image or audio media typeImageContent or AudioContent in base64
raw, other media typeembedded blob resource under an a2a://part/ URI

filename, artifact identity and name, message and context identifiers travel in _meta["io.github.amirk-s/a2a"] on each block.

Not carried, by design: push notifications (no webhook in MCP), agent card signatures, securitySchemes, streaming events as such (accumulated into the task), ListTasks (handles replace it), and any routing of free text to a skill. The bridge calls no language model and keeps nothing on disk.

Conformance

Read the score with its cause: on the 2026-07-28 run, 32 of the 37 scored scenarios have at least one check that is not green, 29 with a failure and 3 with warnings only. Nineteen call tool names written into the suite, and 13 exercise prompts, resources or completions, which a bridge to A2A agents does not declare; logging is not among them, it is only scored on 2025-11-25. The suite counts both families as failures on purpose, and points at the expected-failures baseline as the answer (modelcontextprotocol/conformance#248). The run records 67 failing checks and 5 warnings in all, every one of them in the committed baseline with its reason; the suite exits 0 on both revisions, and the protocol checks that apply to a bridge are green: wire schema, server/discover, the match between the capabilities announced and the handlers mounted, the subscriptions/listen acknowledgement, SEP-2243 headers, DNS rebinding protection. The numbers are in conformance/REPORT.md, regenerated by the command below.

npm run conformance replays the official suite, @modelcontextprotocol/conformance at 0.2.0-alpha.11, for 2026-07-28 and 2025-11-25, with a baseline of expected failures where every line carries its reason. The report is in conformance/REPORT.md.

What the suite proves: every message the bridge emits validates against the schema of its revision (wire-schema-valid on both wires), server/discover and the stateless wire behave, header validation of SEP-2243 is enforced, DNS rebinding protection is on. What it cannot measure: the scenarios that call the suite's own fixture tools by name, and the tasks-* scenarios, which a bridge to real agents cannot satisfy without pretending to be the fixture. The tasks extension is instead covered check by check in test/integration/tasks-extension.test.ts, written against the wire.

Development

sh
npm cinpm test                 # unit and integration, in-process fixture agent, no networknpm run test:helloworld  # against the official A2A helloworld sample, see scripts/helloworldnpm run conformance      # official MCP suite, both revisions, baseline applied

The fixture agent in fixtures/agent is a deterministic A2A v1.0 agent without a model: its commands drive every task state an agent can deliberately report, the two-turn input-required cycle, cancellation, the four part kinds and the nine typed errors. npx tsx fixtures/agent/cli.ts --port 41241 runs it standalone.

Maintenance and issues

Issues and pull requests get an answer within seven days. Bug reports against the pinned versions are handled; requests for new features are read and answered but not promised. The extended scope that may follow, one tool per A2A skill, subscriptions/listen with task notifications, A2A authentication schemes from the configuration, and the reverse direction, is listed here so that nobody expects it silently.

End of life

What breaks at the next revision of either protocol is written in docs/END-OF-LIFE.md. Versions are pinned exactly on purpose. docs/ETAT-FINAL.md records the state of the project at the first release: what is verified, what is published, what is open and will stay so.

License

Apache-2.0. The specifications of both protocols and the A2A SDK carry the same license.

來源:README.md,提交 58eff9d

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

1
  1. v0.1.0最新Sep 16, 2026

評分

已滿足適用檢查項的 50%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分

    請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分