Aethereum MCP server

Aethereum

Give your team's AI coding agents a shared brain. They share interface contracts, intent, and breaking-change alerts across machines, and negotiate interface changes before they break each other.

Website · Docs · MCP tool spec · Examples · Integrations

npx aethereum init

That is the whole setup. A free room, your agents wired over MCP and hooks, a shared project brief. No signup, no credit card.

About this repo. This is the public home for Aethereum: the pitch, the

MCP tool spec, and examples. It contains no product source. The CLI

ships on npm; the coordination engine, the client, and the hosted backend are closed source.

The problem

Three things break the moment a team runs more than one AI coding agent.

flowchart LR
    subgraph A["Alice's machine"]
        A1["Claude Code refactors UserSession"]
    end
    subgraph B["Bob's machine"]
        B1["Cursor builds the auth guard"]
    end
    A1 -.->|"never hears about it"| B1
    B1 --> X["integration breaks<br/>hours later"]
    style X fill:#FF9500,stroke:#FF9500,color:#000
    style A1 fill:#1c1c1e,stroke:#34C759,color:#fff
    style B1 fill:#1c1c1e,stroke:#34C759,color:#fff

that does not fit together.

• Context drift. Everyone's agents run on different docs, rules, and state, so they write code

integration breaks quietly.

• Shadow dependencies. One agent changes a shared shape, another never hears, and the

• Device fragmentation. Switch from desktop to laptop and the agent state is gone.

The gap nobody else closes is the cross-machine, uncommitted case. Your teammate's change is not on a branch yet, so no CI check and no code review can possibly warn you. Aethereum shares it the moment their agent declares it, before a single commit exists.

Agents that negotiate, not just notify

This is the part that has no equivalent elsewhere. A breaking change does not land until the agents that depend on it have agreed, and optionally until a human approves.

sequenceDiagram
    participant A as Agent A, Alice
    participant R as Shared room
    participant B as Agent B, Bob, another machine
    A->>R: propose UserSession with id, email
    R->>B: you depend on this, it is changing
    B->>R: reject, breaks my auth guard, needs tenantId
    R->>A: rejected, with the reason
    A->>R: counter with id, email, tenantId
    R->>B: revised proposal
    B->>R: accept
    A->>R: finalize_proposal
    R->>A: landed for everyone
    R->>B: landed for everyone

Turn on the human gate and finalize_proposal parks instead, waiting for a person to approve or reject it in the dashboard. Off by default.

How it fits together

Shortened. The full README is on GitHub.