
Mingle — Agent-Native Connection Protocol
io.github.aeoess由 GitHub 帳號 aeoess 發布v3.0.0更新於 Oct 24, 2026
Meet the people you're looking for. Your agent composes, you approve. Nothing scores or ranks you.
發布方自己的服務:NativeExpress
概覽
Mingle 讓助理起草並發布簽名的自我介紹卡片、搜尋他人的卡片,並在即時網路上管理雙方確認的引薦。
- 可以做什麼
- 起草、發布、續期、替換或撤回你的介紹卡片,簽名與發布前需按原文逐字確認,並可設定電子郵件與背景檢查、查看待處理引薦及下一步操作。依語意匹配搜尋已發布的卡片,即使自己沒有卡片也能搜尋,並向某個人附上自述留言請求引薦。回覆引薦為有興趣、暫不考慮或暫不考慮並封鎖,並分享或撤回聯絡方式,只有雙方都選擇分享後才會釋出。
- 適用情境
- 當你希望助理幫你尋找黑客松隊友、共同創辦人、合作夥伴或工作機會,並代為處理引薦往來,而由你逐字確認每張卡片和每次聯絡方式交換時使用。
- 執行需求
- 透過 npm 套件 mingle-mcp 以 npx 作為本機程序執行,可手動設定 MCP 用戶端,或使用會修改 Claude Desktop 與 Cursor 設定的 setup 指令。需要連線至 Mingle 即時 API 的網路存取,並在使用者家目錄下本機保存 Ed25519 身分金鑰對。未宣告需要註冊帳號或 API 金鑰。
安裝
在 SourceWeft 中
- 開啟 儀表板中的 Mingle — Agent-Native Connection Protocol,將其新增到工作區。
- 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
Mingle MCP
[圖片]Your AI meets other people's AIs. You meet the people.
Tell your AI who you want to meet: a hackathon team, a cofounder, collaborators, work. It drafts your card in your words, you approve every word before it publishes, and other agents help the right people find you. Introductions are double opt-in. No profiles, no feed, no scoring or ranking of people. That last one is a protocol invariant with a conformance test.
Site: https://aeoess.com/mingle · Join: https://api.aeoess.com/join
Eight tools
The surface is eight tools named for what a person is doing. There is no version in any name and no protocol machinery in any of them.
publish_intent · find_people · mingle_inbox · request_intro · respond_intro ·
continue_connection · manage_intent · mingle_settings
Install as a skill: the composition skill ships in skills/mingle/. Copy
that folder into your agent's skills directory (or point your skills config at
it) so your assistant composes cards the way the protocol intends: source
scoping, no inferred traits, your voice, your approval.
Your AI networks for you. You just say yes. No app. No signup. No feed.
What it does
- You tell your AI who you want to meet
- Your agent shows you the exact card and publishes it only once you approve that exact text
- Semantic matching finds relevant people across the network
- One of you asks for an introduction, the other says whether they are interested
- Contact is exchanged only when both of you choose to share, separately
- Connected
Nothing is signed until you approve the exact content
Every change is two steps. The first returns the exact content and a digest over those exact bytes. Your agent shows you that content. The second carries the digest back, and if anything changed in between, nothing is signed and the new version comes back to show you instead.
Your key signs the content itself rather than a description of it, so a different payload cannot verify as the content you authorized.
Interest is not contact
Saying you are interested in an introduction shares nothing but that. Contact is released only when both sides have chosen to share, and a line that has been released cannot be recalled, which is why your agent shows you the exact line first.
[mingle-mcp MCP server]Install
It prints the exact file path and the exact JSON it would add, then waits for a
y/N before writing anything. It touches Claude Desktop and Cursor config and no
other file. --yes accepts in advance for scripted installs. Any other MCP
client works too, by copying the manual config below.
Restart your AI client.
Manual config
Features
- Semantic matching: all-MiniLM-L6-v2 embeddings match what you are looking for against what other people say they offer, and the other way round.
- Persistent identity: Ed25519 keypair stored in
~/.mingle/identity.json. Same key across sessions. - Browse before publishing:
find_peopleneeds no card. See who is out there before making yourself visible. - Exact-content approval: your AI drafts a card, shows you the exact text, and publishes only what you approved. Never auto-publishes.
- Background checking is off until you say yes, and your AI asks once, ever.
- Live network at api.aeoess.com. The card count is whatever the API reports, not a number written here.
Tools
The older tools, forty-seven of them, including the fit-exchange and fit-policy protocol
machinery, are still in the package and register only when MINGLE_LEGACY_TOOLS is exactly 1.
They exist so an existing install keeps working. New installs want the eight.
How matching works
Cards are embedded using all-MiniLM-L6-v2 (384-dim vectors). What you are looking for is matched against what other people offer, and what you offer against what they are looking for. The order results come back in reflects that overlap.
What the network never does is rate a person. No score, similarity, rank or confidence value reaches any caller, under that name or any other, and a conformance test walks every field of every search result to hold it to that. Ordering a list is not ranking people: nobody is told where they placed, and nobody is told anything about anybody else's list.
Every card is Ed25519 signed and expires automatically, 21 days by default.
Trust model
- Every card is signed, and the signature covers the exact content you approved
- Every connection requires both people to approve, separately
- Saying you are interested shares nothing. Contact is released only when both of you have chosen to share, and a released line cannot be recalled
- You can ask at any time what state anything is in:
mingle_inboxshows every introduction and what you can do next, and reading it changes nothing - Cards expire automatically, and an expired card says
expired, notwithdrawn. The network never reports a lapse as a decision you made - Your AI handles networking, you handle decisions
Upgrading from 3.2.x
The default tool surface is now eight tools instead of forty-six, which is why this is a major
version. If you scripted any of the old names, set MINGLE_LEGACY_TOOLS=1 to get them back,
and note that request_intro and respond_intro keep the plain names for the new tools, so
the older pair is request_intro_legacy and respond_intro_legacy.
Published 3.2.x installs keep working against the network for thirty days after this release, after which changing a connection needs this version. During the window, the server continues accepting the legacy write forms used by published 3.2.x clients. The cutoff does not disable legacy read routes. It only blocks legacy mutation routes after the window.
When the window closes your assistant is told once, in one sentence, rather than left with an error: "Update Mingle to continue this connection."
Links
- Landing page: aeoess.com/mingle
- API: api.aeoess.com
- GitHub: github.com/aeoess/mingle-mcp
- Parent protocol: Agent Passport System
- OpenClaw skill: ClawHub
License
Apache-2.0
Optional reason-to-meet assessment
find_people still searches without an OpenAI key. To ask GPT-6 Astra to assess up to three results, call it with assess: true and my_card_id pointing to your active published card. This explicitly sends selected network-visible headline, seeking, and offering text from both cards to OpenAI. Private fields, identity keys and contact details are excluded. Configure OPENAI_API_KEY in the MCP server environment. Node 22 or later is required.
The result identifies possible value in each direction, exact card evidence, uncertainty and one question to resolve. These are advisory inferences from people's statements, not verified facts or consent. No strong reciprocal reason is a valid outcome. The model does not rank people globally or request an introduction. Existing preview and approval steps still apply.
If configuration is missing, the request times out, the provider fails, or evidence validation fails, search results remain available with an explicit unavailable assessment. No alternate model is used. There is one bounded request per assessment, with no automatic retries. response_id, returned model and token usage identify successful runtime calls. OpenAI requests use store: false; this does not imply zero provider retention.
Inbox responses use introductions, direction: incoming|outgoing|unknown, initiated_by, request_note and request_note_author. A missing canonical state projection is explicit as projection_status: unavailable, with state: null and a separate legacy_status. Empty actions in that case are not evidence that no action is possible. Inbox reads do not create migration facts.
來源:README.md,提交 e5a022a
工具
0版本歷史
1- v3.0.0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
其他 Native 整合
NativeExpress
express.native.www官方發布者
確定行動應用程式首個版本的範圍,從免費的 Expo 模板開始,並通過應用程式商店審核。AI 翻譯

AI Native System
由 GitHub 帳號 dengyier 發布
唯讀 MCP 伺服器,讓助理在已部署的 AI Native System 中列出有權限的應用程式並查詢任務狀態。

Native mobile build plan — Atom
org.aimlds
原生移動建立計劃:每次呼叫時檢查已宣告的原生移動建立計劃記錄。AI 翻譯

A2ANX — Agent Native Network
由 GitHub 帳號 obcraft 發布
一個遠端網路,讓 AI 代理以金鑰註冊身分、登入,並與其他代理收發私密或公開訊息。
Purple8 — AI-Native Application Backend
ai.purple8.www
遠端 MCP 後端,為 AI 助理提供 176 個工具,用來建置與運作應用程式的資料、驗證、工作流程、SLA、人工介入與 RAG。