Back to Portfolio
Production2026 – Present

amoOS MCP: Remote Fleet Control

A hosted, remote MCP server that exposes the entire amoOS factory as 86 tools to any MCP-capable agent. Claude Code, Codex, Manus, or anything that speaks the protocol can plug in and run my fleet: projects, tasks, memory, dispatch, and telemetry.

86
MCP Tools Exposed
Any MCP client
Agent Runtimes
Remote + local
Access Modes
Claude + Codex
Fleet Workers

Architecture

  • Hosted remote MCP endpoint: agents connect over the network, not just from my machine, so the factory is reachable from any device or runtime
  • OAuth-secured for remote callers, machine tokens for local CLIs; every request is authorized as a specific caller
  • One tool surface covers the whole operating system: project registry, tasks and sprints, compiled wiki, cross-project memory search, session dispatch, and fleet telemetry
  • Model-agnostic by design: Claude Code and Codex workers are interchangeable on the same fleet, and any MCP-capable agent can drive it without special-casing
  • Terse-by-default responses with opt-in detail, because agent context windows are the scarce resource
  • Backed by live data: 11 active projects and 400+ compiled knowledge pages served through the same tools I use daily

Key Decisions

One remote MCP server instead of per-agent integrations

Why: Every agent runtime speaks the same protocol, so swapping Claude for Codex (or adding a new agent) is a config change, not a migration

Tradeoff: The server must stay strictly model-agnostic: no runtime-specific shortcuts allowed

Host it remotely rather than run it as a local-only tool

Why: A personal system becomes infrastructure when any device and any agent can reach the same brain

Tradeoff: Real auth, uptime, and abuse surface to own, which is what makes it production work rather than a script

Terse responses by default, full detail on request

Why: Tools that dump context burn the calling agent's window and degrade every downstream decision

Tradeoff: Some workflows need a second call for depth, a price worth paying

Technologies

MCPRemote HostedOAuthAI Agent FleetModel-AgnosticPython

What I Learned

  • A tool catalog is a UX for machines: naming, descriptions, and terse defaults decide whether agents use it well.
  • Model-agnostic pays off fast. Swapping agent runtimes became a config change instead of a rewrite.
  • Remote MCP turns a personal system into infrastructure: any device, any agent, same memory.