Mingle — Agent Native Connection Protocol

Mingle MCP

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

1. You tell your AI who you want to meet
2. Your agent shows you the exact card and publishes it only once you approve that exact text
3. Semantic matching finds relevant people across the network
4. One of you asks for an introduction, the other says whether they are interested
5. Contact is exchanged only when both of you choose to share, separately
6. 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.

Install

npx mingle-mcp setup

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.

{
  "mcpServers": {
    "mingle": {
      "command": "npx",
      "args": ["mingle-mcp"]
    }
  }
}

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_people needs 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_inbox shows every introduction and what you can do next, and reading it changes nothing
• Cards expire automatically, and an expired card says expired, not withdrawn. 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.