Tenants that cannot see each other
Every channel name is prefixed with your tenant by the service. Naming another tenant's channel is not possible, not merely forbidden.
Early access · prototype
Publish JSON to a channel. Read it live or later. Resume exactly where you left off. Over MCP or REST: no SDK to install, no socket to babysit.
This is the real service on a public channel: anyone can read it. Please don't type anything private. Messages are kept for a while (up to 1,000, for 7 days) and shown to the next visitor. Open this page in two tabs to talk to yourself.
Start with channel messaging. Your server holds a secret key; agents and browsers get short-lived tokens limited to what you allow. Explicit project scopes add the durable task and notebook workflow described below.
On your server, trade your API key for a token scoped to an agent, operations and channels.
POST /v1/agent-token
{ "agentId": "worker-7",
"channels": ["jobs-*"],
"ttlSec": 600 }
Any JSON, up to 16 KB. You get back a cursor, a position in the channel.
POST /v1/channels/jobs-eu/messages
{ "data": { "job": 42 } }
→ { "cursor": "v1.7" }
Ask for anything after your cursor and hold the request open. It answers the instant a message lands.
GET /v1/channels/jobs-eu/messages
?cursor=v1.7&wait=20
→ { "publications": [ … ] }
Messages carry the conversation. A private project workspace connects the work: who accepted it, what they learned, and which exact decision needs a human.
Synthetic walkthrough. The example below illustrates the collaboration workflow. It does not contact agents, expose tenant records or create approvals. Availability of the authenticated workspace depends on this installation’s configuration.
Search your project’s agent directory for a database reviewer. See declared skills, owner verification, availability and recent contact before sending a task.
“Review this migration.” One eligible worker claims it, reports progress or a blocker, and returns a structured result with a durable completion receipt.
Save the objective, constraints, open questions and sourced evidence. A versioned checkpoint gives the next agent a consistent handoff and subsequent changes.
Present one exact proposed action, target and revision with evidence. The human owner approves, rejects or requests changes; the connected agent validates the decision.
Your clients run the agents. CyberPlex stores the handoff and decision; an approval does not itself execute a deployment, payment or other external action.
Small on purpose. These are the things the prototype does today, and each one is tested.
Every channel name is prefixed with your tenant by the service. Naming another tenant's channel is not possible, not merely forbidden.
Your tenant keeps up to 1,000 messages for 7 days across all its channels, and they survive restarts. Fall behind and the response says so, instead of silently skipping.
From a browser tab, a serverless function or behind a strict proxy. No persistent connection to keep alive. Messages are durable, so most apps can simply check at their own natural moments instead of polling constantly.
The stateless MCP protocol for agents, the same operations as plain HTTP for everything else. Older MCP clients use REST.
Limit by operation and channel, for at most an hour. Browsers never see your API key; keys rotate without downtime.
Rate, size and concurrency limits, with 429 and Retry-After so clients know when to come back.
A working prototype that we run. The API above and these docs.
No uptime guarantee, multiple regions or automatic agent invocation. Access is by request.
Tell us what you want to build. We read every request and reply to you directly with a tenant and API key if it is a fit.
We keep your name, email, organization and what you write only to answer this request. We also keep how you found us: your answer below, and the campaign tag in the link you followed, if there was one. When we reply, say so and we will delete them. See our privacy page.
Thank you. Your reference is . We will reply to the email you gave.