Tools

oh-my-openagent: An Agent OS for Multi-Model Orchestration

A 68k-star open-source Agent project built around multi-model and multi-harness orchestration, openly positioned against Claude Code. This piece walks through the three-layer architecture, a minimal config, and who should adopt it.

Font size

Spotted on GitHub Trending: code-yeongyu/oh-my-openagent. 68k stars, written in TypeScript, maintained by a developer who streams commits live on Discord. Plenty of projects build in public, but this one stands out — its founding premise is opposition to Claude Code.

The repo's tagline:

So this is my lil' open source agent project, oh-my-openagent. An AI harness for tokenmaxxers, parallel multi-model + multi-harness orchestration.

tokenmaxxer literally means someone squeezing every last drop of value out of their tokens. If your monthly model bill runs a few hundred dollars and you hold Kimi, GPT-5, and Gemini accounts side by side, this project is built for you.


What problem it solves

Single-model agents are getting more expensive and slower on complex codebases, while locking you to a single vendor. A single Claude Code session can burn through $20, and reading a long repo forces you into premium models like Opus.

oh-my-openagent positions itself as an "Agent OS" — wiring multiple models (Kimi, GPT-5, Gemini) and multiple harnesses (OpenCode, Codex, Pi) into an orchestrated workflow. Tasks that involve "reading" go to cheap, long-context Kimi; diff-writing goes to GPT-5's strong code generation; planning goes to Claude. Routing is driven by token cost and capability.

The maintainer goes further: he's publicly stated that Anthropic restricted certain OpenCode capabilities because of this project. Whether or not that claim holds, "multi-model neutrality + anti-vendor lock-in" is baked into the product's DNA.


How to use it

One-line install:

npx oh-my-opencode install

After install, create an agent.config.ts in your project root:

// oh-my-openagent Team Mode minimal config
import { defineConfig } from 'oh-my-opencode';

export default defineConfig({
  providers: {
    kimi:   { model: 'kimi-k3',     apiKey: process.env.KIMI_KEY },
    gpt:    { model: 'gpt-5.6-sol', apiKey: process.env.OPENAI_KEY },
    claude: { model: 'claude-sonnet', apiKey: process.env.ANTHROPIC_KEY },
  },
  mode: 'team',
  team: {
    planner:  'claude',
    readers:  ['kimi'],
    writers:  ['gpt'],
    fallback: ['claude'],
  },
  budgets: { perTaskUsd: 0.5 },
});

// Run:
//   npx oh-my-opencode run --task "Summarize this repo's architecture"

Four config blocks: providers, mode, team, and budgets.

For a lighter touch — just hooking into Codex without the full OmO stack — there's npx lazycodex-ai install. It's a Codex-only subset of OmO that ships only the harness layer you need.


How it works (architecture)

Three layers:

(1) Harness layer: adapters for OpenCode, Codex, and Pi, with a unified tool-calling protocol. OmO normalizes here, so upper layers don't care which engine runs underneath.

(2) Orchestrator layer: Team Mode splits a task into sub-agents and dispatches them to different models based on token cost and capability. team.readers, team.writers, and team.planner are the routing table.

(3) Provider layer: swap between Anthropic, OpenAI, Mooncake, Kimi, GLM with no business-code changes. Provider and routing are decoupled.

The repo's ROADMAP.md outlines three major refactors: a Multi-Harness Agent OS refactor, support for OpenCode / Codex / Pi, and a unified tool-calling protocol. The 1.0 release is still in active refactoring.


Strengths and weaknesses

Strengths:

  • Open source and multi-model neutral; 68k stars signal real community traction
  • The maintainer streams development on Discord, and commit messages are written for humans
  • Anti-vendor lock-in stance is explicit, with built-in adapters for non-US providers like Kimi, Moonshot, and Mooncake
  • Team Mode ships with sensible default combinations

Weaknesses:

  • Defaults to Discord streams and English docs, which hurts the install/upgrade experience for non-English users
  • Multi-model stitching naturally introduces prompt compatibility and latency overhead — half of the token savings can get eaten by the adapter layer
  • API is unstable during the refactor; check the ROADMAP before committing
  • Small maintenance team raises long-term sustainability questions

Who should use it

Good fit:

  • Independent developers with multiple model accounts looking to cut per-token costs
  • Small teams (under 5 people) running parallel sub-agents on large codebases
  • Users opposed to Claude Code dominance who want open, neutral tooling
  • Agent enthusiasts willing to read the ROADMAP and contribute early

Not a fit:

  • Anyone who just wants to fire up Claude Code and forget about configuration
  • Latency-sensitive production CI scenarios (sub-agent orchestration adds 1–3 seconds of overhead by itself)
  • Enterprise users who need stability over flexibility

The bottom line

oh-my-openagent isn't "yet another Claude Code wrapper." Its product philosophy is that agents should be orchestrable and not locked to a single vendor. The 68k stars come from that stance, not from a pretty UI.

If your monthly model bill already clears a few hundred dollars, read ROADMAP.md carefully before adopting. Anthropic loyalists can skip this one.

Source: https://github.com/code-yeongyu/oh-my-openagent