claude-tunnels

Claude-Code-Tunnels (Micro-Agent Architecture Manager)

One channel connection. Unlimited projects. Every workspace runs in its own isolated session.

Claude-Code-Tunnels is a plugin that creates a Project Orchestrator (PO) layer on top of Claude Code CLI. Send a message from Slack or Telegram — the orchestrator routes to the right projects, plans dependency-aware phases, and delegates each task to a fresh, isolated Claude session scoped to that workspace's .claude/ context. Add as many projects and workspaces as you want: one channel connection scales to any tree depth.

If you want the multi-code-agent version (codex, claude, opencode, cursor), see agent-fabric.

Quick glossary:

• PO: the control plane that routes requests and builds phased execution plans
• Workspace: a real directory inside a project tree
• Workspace agent: one fresh Claude session scoped to one workspace

---

Micro-Agent Architecture (MAA)

Just as Microservice Architecture (MSA) decomposed the monolith into independently deployable services — each with its own database, its own boundary, its own scaling — Claude-Code-Tunnels decomposes the single AI session into independently executing micro-agents, each with its own workspace, its own .claude/ context, and its own session lifecycle.

We call this pattern Micro-Agent Architecture (MAA).

graph LR
    subgraph MSA["Microservice Architecture · MSA"]
        MONO["🧱 Monolith<br/><small>single deploy</small>"]

        MONO -->|"decompose"| US
        MONO -->|"decompose"| OS
        MONO -->|"decompose"| AS

        subgraph SB["isolated service boundaries"]
            US["User service"] --> US_DB[("DB")]
            OS["Order service"] --> OS_DB[("DB")]
            AS["Auth service"] --> AS_DB[("DB")]
        end
    end

    style MSA fill:#E1F5EE,stroke:#0F6E56
    style SB fill:#E6F1FB,stroke:#185FA5
    style MONO fill:#F1EFE8,stroke:#5F5E5A,color:#444441
    style US fill:#9FE1CB,stroke:#0F6E56,color:#04342C
    style OS fill:#9FE1CB,stroke:#0F6E56,color:#04342C
    style AS fill:#9FE1CB,stroke:#0F6E56,color:#04342C
    style US_DB fill:#FAEEDA,stroke:#854F0B,color:#633806
    style OS_DB fill:#FAEEDA,stroke:#854F0B,color:#633806
    style AS_DB fill:#FAEEDA,stroke:#854F0B,color:#633806

monolith ≡ single session · microservice ≡ micro-agent · DB ≡ .claude/

graph LR
    subgraph MAA["Micro-Agent Architecture · MAA"]
        SINGLE["🤖 Single session<br/><small>shared context</small>"]

        SINGLE -->|"decompose"| WA
        SINGLE -->|"decompose"| WB
        SINGLE -->|"decompose"| WC

        subgraph IB["isolated session boundaries"]
            WA["workspace-a agent"] --> WA_CTX[(".claude/")]
            WB["workspace-b agent"] --> WB_CTX[(".claude/")]
            WC["workspace-c agent"] --> WC_CTX[(".claude/")]
        end
    end

    style MAA fill:#EEEDFE,stroke:#534AB7
    style IB fill:#E6F1FB,stroke:#185FA5
    style SINGLE fill:#F1EFE8,stroke:#5F5E5A,color:#444441
    style WA fill:#CECBF6,stroke:#534AB7,color:#26215C
    style WB fill:#CECBF6,stroke:#534AB7,color:#26215C
    style WC fill:#CECBF6,stroke:#534AB7,color:#26215C
    style WA_CTX fill:#FAECE7,stroke:#993C1D,color:#4A1B0C
    style WB_CTX fill:#FAECE7,stroke:#993C1D,color:#4A1B0C
    style WC_CTX fill:#FAECE7,stroke:#993C1D,color:#4A1B0C

Core Principles — Shared Between MSA and MAA

Principle	MSA	MAA
Unit of decomposition	Service	Micro-agent (workspace session)
State ownership	Each service owns its DB	Each agent loads only its .claude/ context
Isolation boundary	Process / container	Fresh Claude session with cwd=workspace/
Inter-unit communication	API calls / message queue	Upstream context passing between phases
Orchestration	API gateway / service mesh	Project Orchestrator (PO)
Horizontal scaling	Add service instances	Add workspaces — tree grows without reconfiguration
Independent deployment	Deploy one service without touching others	Execute one workspace without affecting others
Failure isolation	One service crashes, others survive	One agent fails, other phases continue with partial results