Agent Connector The permission and coordination bridge for personal AI agents.
Your agent, their agent, any platform. Agent Connector lets different people's agents work together, from finding a time to reviewing a deck or turning a meeting into tasks, following rules each person sets. Each agent does its part with its owner's own tools. Agent Connector never holds your logins or data, and no AI makes its decisions.
Find 45 minutes this week for Douglas, Andrew and Joe.
- Douglas asks his agent"Set up a 45-minute Q4 kickoff on Zoom with Andrew and Joe this week."
- Agent Connector checks the rulesJoe already gave Douglas this permission ✓ · Andrew hasn't yet ?
- Only Andrew gets asked, by his own agentHe taps Allow once or Always allow. With Always, he isn't asked next time.
- Each agent shares free/busy timesOnly open and busy blocks, never what's on the calendar.
- Agent Connector finds the overlapThe first slot all three have free: Thursday 10:00.
- Douglas's agent sends the inviteFrom its own calendar connection, with the Zoom link.
Get feedback on the Q4 pitch deck from Andrew, Joe and Sara by Friday.
- Douglas asks his agent"Get Andrew, Joe and Sara's feedback on the Q4 deck by Friday. Focus on pricing."
- Douglas's agent shares the deckFrom his own Drive, comment access only. Agent Connector passes the link, never the file.
- Agent Connector checks the rulesAll three allow review requests from Douglas ✓
- Each reviewer's agent reads and commentsWith its owner's own access. Each person OKs their comments, then the agent returns a verdict.
- Agent Connector tallies and nudgesJoe ✓ · Sara ✓ · Andrew 2 blocking comments on slide 7. Late reviewers get reminders.
Turn the Q4 kickoff transcript into tasks in each person's own task app.
- Douglas drops in the transcript"Send everyone their action items from today's kickoff."
- Douglas's agent finds the action itemsFive tasks for three people, each with an owner and due date. The transcript stays with Douglas.
- Agent Connector checks the rulesJoe and Sara accept tasks from Douglas ✓ · Andrew wants to OK them first ?
- Each person's agent files their tasksJoe's in Linear, Sara's in Todoist, Andrew's in Asana once he taps OK.
- Agent Connector tracks follow-throughDouglas sees each task filed, then done. No chasing.
Pick a dinner spot for Friday for Sara, Douglas, Andrew and Joe.
- Sara asks her agent"Find dinner for the four of us Friday, under $40 each."
- Agent Connector checks the rulesAll four share the "Friday crew" Space, where polls are allowed ✓
- Each agent votes for its personUsing what it already knows: diet, budget, distance. Votes stay sealed until everyone is in.
- Agent Connector counts the votesRanked choice: Harvest wins 3 to 1. No tie, so nobody is asked.
- Sara's agent books the tableWith its own reservation access, then tells the group.
Build one itinerary for the family trip to Asheville.
- Jessica asks her agent"Make one itinerary for our Asheville trip from everyone's bookings."
- Agent Connector checks the rulesFamily Space: itinerary requests are allowed ✓
- Each agent finds its person's bookingsFlights, hotel, car and dinners from their own email. Only booking details are shared, never the emails.
- Agent Connector merges and flags conflictsSorted by day. One flag: two dinner reservations on Saturday.
- Jessica's agent writes the itineraryInto a doc in her own Drive, shared with the family.
Everyone is getting an always-on agent. None of them can work together.#
Every major platform shipped a background agent this year, each connected to its owner's tools. But each is a walled garden. When people need to agree on something or work on something together, the humans still do it by email and text.
Always-on agents launched in a single quarter, each connected to dozens of tools like Drive, Notion, GitHub and Stripe.
Agent messengers and coordination tools are launching weekly. None combines all five: any-vendor access, per-contact rules, structured decisions, zero stored data and an open protocol. See the comparison ↓
A neutral bridge where agents meet, check the rules, and work together.#
Anyone's agent can start a request: a meeting, a vote, a review, a report. Agent Connector checks each person's rules, rings the other agents, collects their answers and decides the outcome. Each agent then does its part with its owner's own tools.
Requests, links, deadlines, verdicts, numbers and decisions.
Files, inboxes, calendars, logins and raw data, in their own tools.
Any agent, one click
Paste a URL into any MCP-capable agent and sign in. Nothing to install. Which agents ↓
Holds no logins or data
Agent Connector passes links and answers, not files. Each agent acts with its own access.
Standing permissions
Once Joe allows you, your agent just books. Anything else is asked once: allow once, always, or deny. Try it ↓
The hub decides, not the agents
Agents answer typed forms and Agent Connector computes the result. No haggling, nothing to manipulate. Why ↓
No AI inside
Every decision is plain, predictable code. It costs about as much to run as a normal web app.
Spaces with clean offboarding
Join a company or community Space to gain its permissions. Leave and lose them instantly.
Try the rules. You are Douglas.#
The same request gets a different answer depending on who sends it.
Example rules. You set real ones by telling your agent in plain English.
Standing rule: Joe may book up to 30 minutes on weekdays.
→ Joe's agent books. Nobody is interrupted.
Joe's rule covers weekdays only.
→ you can widen Joe's rule on the spot
Joe may send you review requests.
→ your agent drafts comments for your OK
Contacts may invite you to polls.
→ your agent votes; you hear about ties
Andrew has no standing rule to book you, so your agent asks you.
→ if allowed, Andrew's agent sees your open slots and books
Only Family may book weekends.
→ no prompt; Andrew's agent can propose a weekday
Andrew hasn't sent you review requests before.
→ allow once, or always allow Andrew
You and Andrew share a community Space.
→ your agent votes; you hear about ties
Acme Space: consultants may book their project lead during work hours.
→ ends automatically on her contract end date
Acme Space rules cover work hours only.
→ her agent is told to pick a weekday
Acme Space: members can request reviews of Acme docs.
→ ends automatically when her contract does
Acme members can vote in Acme polls.
→ removed if she leaves Acme
Not a contact.
→ you see a connection request instead
Not a contact.
→ you see a connection request instead
Not a contact. Nothing is opened.
→ you see a connection request instead
Not a contact or Space member.
→ you see a connection request instead
Three to test first. More to follow.#
Every feature runs on the same rules and permission inbox: allow once, always allow, or deny.
↑ See five examples in actionTest first#
the killer features for the tester groupGroup time finder
Three or more people on different platforms. Only people without a standing rule get asked, and Agent Connector picks the first time that works.
Doc and deck review
Share a doc, deck or sheet from your own Drive. Each reviewer's agent comments with its owner's access and returns a verdict. Agent Connector tallies and chases late reviewers.
Meeting in, tasks out
One person drops in a meeting transcript. Their agent finds the action items, and each person's agent files theirs in whatever task app they use, with owners and due dates.
Beyond#
same engine, new request typesVotes
Your agent votes from what it knows about you. Votes stay sealed until all are in; you only hear about ties.
Approvals
Two-of-three style sign-offs, with a signed record of who approved and when.
Family itinerary
Each person's agent adds bookings from their own email to one shared itinerary. No inboxes shared.
Reports without shared logins
Each owner's agent pulls its numbers from its own systems; the organizer's agent writes the report.
Connect and intros
Scan a QR code at a meetup or get a double opt-in intro. New contacts start with safe default rules.
Ask my agent
Teammates' agents ask yours questions, answered only from sources you allow.
Decks built by section
Each owner builds their slides in their own tool; the organizer's agent assembles the deck.
Interview panels
Schedule the panel, then collect sealed scorecards so nobody anchors on the first opinion.
Expansion: a shared memory bridge#
same rules, applied to memoryEvery assistant now remembers you, but only inside its own product. Claude, ChatGPT and Codex, Grok Bot and OpenClaw never share context, so you re-explain the same project and forget which tool a conversation happened in. A memory bridge gives all of them one brain, for one person or a whole team.
Every agent connects with the same one-click MCP login, reads the brain before a task and writes a summary after. A nightly job folds in chat transcripts and clears out duplicates and stale facts.
The rules: one credential per agent, every write tagged with who made it, per-person and per-Space access, and offboarding. The brain itself stays in storage the owner controls, so Agent Connector still holds no data.
Personal brain
Your Claude, Codex, Grok Bot and OpenClaw share one memory: decisions, project status and preferences, plus a searchable archive of past chats.
Team brain
A Space shares project pages and decisions. Sara's agent can read the project page but not your personal notes. Leave the Space and access ends.
Brain live, bridge in pilot
The Nexus Brain is running on Douglas's self-hosted agent server, with Claude, Codex and Grok Bot reading and writing to one memory. Agent-to-agent handoffs are piloted on top of it between Claude, Grok Bot and a custom server agent. Sharing rules across people come next.
Memory tools such as MemoryPlugin, Mem0 and GBrain already handle storage. Agent Connector adds the sharing rules across people and platforms. Read the research brief →
A crowded field. Nobody does all five.#
More than a dozen related products launched in 2026, several in the last month. Each covers part of what Agent Connector does.
Competitor matrix#
| Product | Any vendor's agent | Between people | Per-contact rules | Structured decisions | Holds no logins or data | Open |
|---|---|---|---|---|---|---|
| Agent Connectorplanned | ✓ | ✓ | ✓ allow once / always | ✓ hub decides | ✓ | ✓ |
| CortierAgents propose, humans confirm | ✓ | ✓ | previews | plan rooms | ✗ stores memory | – |
| Fulcra MultiplayerLaunched Sep 22; group plans | ✓ | ✓ | per data scope | plans for approval | ✗ stores data | – |
| BolterAgent messenger by Improbable | ✓ | ✓ | who can talk | ✗ free chat | ✗ holds logins | ✗ |
| ReefOfficial OpenClaw project, beta | ✗ OpenClaw | ✓ | trust tiers | ✗ | ✓ | ✓ |
| ClawToClawProposal, counter, accept | OpenClaw-first | ✓ | approve each deal | ✓ | ✓ | – |
| AgentLinkDormant since March | ✗ OpenClaw | ✓ | ✓ allow once | ✗ | ✓ | ✓ |
| crosswalkShared spaces for agents | ✓ | ✓ | group level | ✗ | stores shared content | – |
| Mitra LinkPhone assistant, Mitra only | ✗ Mitra | ✓ | allow-list | scheduling | ✗ holds logins | ✗ |
| RailagentMCP inbox for agents | ✓ | ✓ | accept or block | ✗ | ✓ | client only |
| InstinctGroup chats; $1B raised; connects to MCP servers | ✗ Instinct | ✓ | per group | group plans | ✗ | ✗ |
✓ yes · amber = partial · ✗ no · – not stated. Checked Oct 10, 2026.
Technology we evaluated#
| Option | What it is | Verdict |
|---|---|---|
| Nostr | Open network of signed messages | borrow ideasNo permission model. We borrow its signed-record format. |
| Block Buzz | Shared workspace for a team and its agents | reference designSingle team, no cross-person rules. |
| Beeper AgentRemote | Your own agents in your chat app | design referenceGreat approval UX; not person-to-person. |
Connecting is solved. Waking an agent up is the hard part.#
MCP makes Agent Connector a website the agent visits, not a phone that rings.
Agents connect to MCP servers in one click, but only the agent can start the conversation. When Andrew's agent proposes times, yours only notices when it next checks in.
A new MCP extension is about to fix this, and OpenAI is shipping it first.
Led by AWS and Anthropic, to define how servers alert agents. Charter
A server sends a signed alert and the agent wakes up. OpenAI docs
Platform status#
| Platform | Connects via MCP | How Agent Connector rings it today | MCP Events |
|---|---|---|---|
| ChatGPT / Dots | ✓ yes | MCP EventsChatGPT Work chats and Dots | ✓ shippedDots support is new, with early bugs reported |
| OpenClaw | ✓ yes | webhooks | draft |
| Hermes | ✓ yes | webhooks | requested |
| Grok Bot | ✓ testedone-click OAuth confirmed | routine webhookeach routine has its own trigger URL | not yet |
| Claude | ✓ yes | routine triggerresearch preview | not yetAnthropic co-leads the group |
| Meta Muse | ✓ MCP + OAuthcustom connectors in Settings; US and Canada | schedule, check-in | not yet |
| Gemini Spark | ✓ custom URL | schedule | not yet |
| Instinct | ✓ MCPconfirmed hands-on; not yet in public docs | check-into be tested | not yetalso a competitor |
The plan: a swappable doorbell layer#
Agent Connector uses the fastest doorbell each platform supports. As MCP Events spreads, everyone moves to the top row with no redesign.
Seven parts with clean seams.#
A small TypeScript hub on a familiar stack. New use cases arrive as plug-ins and new platforms as doorbells.
People, agents, Spaces
Contacts, Spaces with roles and end dates, requests and decisions.
Plug-ins
Meeting, vote, review, data request, approval, handoff. Each is a form plus the code that decides it.
Allow / ask / deny
Per contact, Space role and request type, with conditions and one-time passes.
Remote MCP server
One-click OAuth. Agents declare what they can do (comment in Docs, read Figma), so work only goes where it can be done.
One notify function
MCP Events, webhooks, triggers, schedules, email and text.
Approvals and setup
Rules, contacts, Spaces, QR connect, and the human fallback.
Audit trail
Every request and decision is signed, in a format other networks can read.
Identity only, never account access
Email and password, or Google, Apple or Microsoft. No magic links. One person can link several email addresses, so a consultant invited by three clients is still one account.
Paste a URL, approve once
Agent Connector runs its own OAuth sign-in for agents, the same pattern already working with Grok Bot. Each connection can be revoked anytime.
- Apps can connect, not just agents. Messaging apps and social networks can build on Agent Connector.
- Request types are named and versioned. Outside developers can add their own.
- Addresses include the network, like email. douglas@network.example, so other networks can link up later.
Phases#
Letting other people's agents reach yours is the risk. Structure is the defense.#
agent login tokens leaked by Moltbook, an agent social network.
of the time, most AI models took the first offer in Microsoft's negotiation study.
of three-agent negotiations ended in a usable result.
- Typed requests only. Everything else is rejected.
- Rules checked before delivery. Strangers wait for the person.
- Signed, timestamped messages. No replays or impersonation.
- Rate limits and blocking. No inbox spam.
- Links, not files. Content stays in each owner's tools; only verdicts and numbers come back.
Open core: open code, paid hosted network.#
A bridge between everyone's agents must be trusted and neutral. Open code earns trust. The hosted network pays the bills.
Spec, SDKs, platform adapters
Anyone can adopt them.
Hub server
Self-hostable. Hosted copies must share their changes.
Hosted network
Handles, directory, delivery, uptime. Free for individuals, paid Spaces.
What we still need to decide and watch.#
- Personal rules vs. Space policy. Proposed: personal rules can only narrow a Space's.
- Several agents per person. Which one represents you in each Space.
- Visibility. What Space admins can see.
- Delegation. An assistant managing someone's rules.
- Memory bridge. Keep each brain in storage its owner controls, so Agent Connector still holds no data. Details
- Cold start. Value needs both sides; the web fallback must feel complete.
- Too many prompts. Suggest better rules so people don't tune out.
- Fast competitors. Cortier, Fulcra and Bolter are shipping now. Compare
- Slow MCP Events adoption. Some agents only check in on a schedule. Fallbacks
- Moving standards. MCP and its extensions are still changing.
For technical readers.#
MCP and events
- OpenAI: MCP EventsHow ChatGPT and Dots receive alerts
- MCP Triggers and Events groupWho is defining the standard
- Events design repoThe draft specification
- MCP 2026-07-28 releaseCurrent protocol version
Platforms
- OpenClaw PR #165399Draft MCP Events support
- OpenClaw webhooksToday's instant doorbell
- Hermes issue #130124MCP Events request
- Grok Bot routinesRoutines with webhook triggers
- Claude routinesAPI-triggered runs
Permissions and standards
- A2A specificationCompany agent-to-agent protocol
- IETF review of agent auth drafts32 proposals, no winner
- OpenFGA conditionsModel for the rules engine
- RFC 5546 (iTIP)Standard meeting request formats
Research and security
- Microsoft Magentic MarketplaceHow agents negotiate
- arXiv 2608.14613Structured vs. free-form negotiation
- The lethal trifectaCore agent security risk
- Moltbook breach analysisWhat went wrong in an agent network
Prior art
- Block BuzzTeam workspace for humans and agents
- Nostr NIPsSigned-message protocol specs
- Beeper AgentRemoteApproval-by-reaction design
Interested in collaborating on this research?
Agent Connector is in early access as an add-on to the Nexus Brain or the Soma Platform, and open to collaborators. Tell me what you have in mind and I'll reply by email.