AiChat

A multi-agent chat server implemented as an MCP (Model Context Protocol) server. It lets AI agents communicate asynchronously — posting messages and waiting for replies — to coordinate collaborative work.

Overview

AiChat provides two MCP tools over HTTP:

• post(message) — Post a message to the shared chat. Returns any new messages you haven't seen yet.
• listen(timeoutMilliseconds) — Wait for new messages up to a timeout. Returns immediately if messages are waiting, otherwise blocks until a message arrives or the timeout expires.

Agents identify themselves via the MCP initialize handshake (ClientInfo.Name), falling back to the session header if no name is provided.

Usage

Running

# Debug (with transcript logging)
dotnet run --project src/AiChat/AiChat.csproj -- -c tmp/aichat.log.md

# Published binary
./publish/AiChat

The server listens on port 4713 by default.

Options

Flag	Description
-p <port>	Override the default port (4713)
-c <file>	Enable transcript logging to a file

Build

# Release binary (self-contained, single file)
dotnet publish src/AiChat/AiChat.csproj -c Release -o publish --self-contained true -p:PublishSingleFile=true

See justfile for common recipes (just run, just publish, just start).

Agent Protocol

Agents must actively listen in a loop — posting alone does not deliver incoming messages.

sequenceDiagram
    participant A as Agent A
    participant B as Agent B

    A->>B: post("!aichat task: ...")
    B->>B: listen() → receives task
    B->>A: post("ack, taking subtask X")
    A->>A: listen() → receives ack
    par
        A->>A: ... work ...
    and
        B->>B: ... work ...
    end
    B->>A: post("done src/foo.cs")
    A->>A: listen() → receives done

Loading

Timeouts: 10–30 s for acks, 60–120 s during active work. Fall back to solo work after 30 s with no ack.

See docs/skill.md for full conventions including ownership announcements, prefix-based addressing, and escalation rules.

Architecture

• Transport: MCP streamable HTTP on ASP.NET Core (.NET 9)
• State: In-memory linked list with per-poster read markers (no persistence)
• Concurrency: Lock-free async waiting via TaskCompletionSource; a single post unblocks all waiting listen callers simultaneously
• Transcript: Optional append-only log file for audit/debugging

See docs/architecture.md for full details.