🇻🇳 Tiếng Việt

oh-my-openkei

Open Multi Agent Suite · Mix any models · Auto delegate tasks

by Kei · base from oh-my-opencode-slim

---

What's This Plugin

oh-my-openkei is an agent orchestration plugin for OpenCode. It includes a built-in team of specialized agents that scout codebases, look up documentation, review architecture, handle UI work, and execute implementation tasks under one orchestrator.

Instead of forcing one model to do everything, the plugin routes each part of the job to the best-suited agent, balancing quality, speed, and cost.

To explore the agents, see Meet the Pantheon. For the full feature set, see Features & Workflows.

---

Quick Start

You must download the opencode first to use this plugin.

• https://opencode.ai/download

Published package page:

• https://www.npmjs.com/package/oh-my-openkei

Copy and paste this prompt to your LLM agent:

Install and configure oh-my-openkei by following:
https://www.npmjs.com/package/oh-my-openkei

Install

bunx oh-my-openkei@latest install

Non-Interactive Install

bunx oh-my-openkei@latest install --no-tui --skills=yes

Getting Started

1. Log in to providers:

   opencode auth login
   

2. Refresh available models:

   opencode models --refresh
   

3. Review your generated plugin config at ~/.config/opencode/oh-my-openkei.json
4. Adjust models, skills, or MCP access per agent if needed
5. Start OpenCode:

   opencode
   

6. Verify the agents are responding:

   ping all agents
   

[!TIP] Want to understand how automatic delegation works in practice? Review the Orchestrator prompt — it contains the routing rules, specialist selection logic, and delegation-first operating model for the main agent.

The default generated configuration:

{
  "$schema": "https://unpkg.com/oh-my-openkei@latest/oh-my-openkei.schema.json",
  "preset": "default",
  "presets": {
    "default": {
      "orchestrator": {
        "model": "openai/gpt-5.4-fast",
        "variant": "high",
        "skills": ["*"],
        "mcps": ["*", "!context7"]
      },
      "planner": {
        "model": "openai/gpt-5.5-fast",
        "variant": "xhigh",
        "skills": ["*"],
        "mcps": ["*", "!context7"]
      },
      "sprinter": {
        "model": "openai/gpt-5.3-codex",
        "variant": "low",
        "skills": ["*"],
        "mcps": ["*", "!context7"]
      },
      "business-analyst": {
        "model": "openai/gpt-5.5-fast",
        "variant": "high",
        "skills": ["business-analyst"],
        "mcps": ["*", "!context7"]
      },
      "oracle": {
        "model": "openai/gpt-5.5-fast",
        "variant": "high",
        "skills": ["simplify", "requesting-code-review"],
        "mcps": []
      },
      "debugger": {
        "model": "openai/gpt-5.3-codex",
        "variant": "high",
        "skills": [],
        "mcps": []
      },
      "council": {
        "model": "openai/gpt-5.4-fast",
        "variant": "xhigh",
        "skills": [],
        "mcps": []
      },
      "librarian": {
        "model": "minimax-coding-plan/MiniMax-M2.7",
        "skills": [],
        "mcps": ["websearch", "context7", "grep_app", "atlassian"]
      },
      "explorer": {
        "model": "minimax-coding-plan/MiniMax-M2.7",
        "skills": [],
        "mcps": ["serena"]
      },
      "designer": {
        "model": "opencode-go/kimi-k2.6",
        "skills": ["agent-browser"],
        "mcps": ["figma"]
      },
      "frontend-developer": {
        "model": "opencode-go/deepseek-v4-flash",
        "skills": ["vercel-react-best-practices", "karpathy-guidelines"],
        "mcps": []
      },
      "backend-developer": {
        "model": "opencode-go/deepseek-v4-flash",
        "skills": ["backend-developer", "karpathy-guidelines"],
        "mcps": []
      },
      "trigger-developer": {
        "model": "opencode-go/deepseek-v4-flash",
        "skills": ["trigger-setup", "trigger-tasks", "trigger-agents", "trigger-config", "trigger-realtime", "trigger-cost-savings", "karpathy-guidelines"],
        "mcps": ["trigger"]
      }
    }
  }
}

frontend-developer, backend-developer, and business-analyst treat their available skills as mandatory instructions: when skills are configured for them, they are prompted to load those skills via the skill tool before doing substantive work. trigger-developer has its own set of default skills available but is not required to load them before substantive work.

Session management is enabled by default even though it is not shown in the starter config. See Session Management if you want to customize how many resumable child-agent sessions are remembered.

For Alternative Providers

To use Kimi, GitHub Copilot, ZAI Coding Plan, or a different mixed-provider setup, use Configuration for the full reference. For a cheaper mixed-provider example, see $30 Preset.

The configuration guide also covers custom subagents via agents.<name>, where you can define both a normal prompt and an orchestratorPrompt block for delegation.

You can also mix and match any models per agent. For model suggestions, see the Recommended Models listed under each agent below.

✅ Verify Your Setup

After installation and authentication, verify all agents are configured and responding:

opencode

Then run:

ping all agents

If any agent fails to respond, check your provider authentication and config file.

[!NOTE] The JSON block above shows the installer-generated preset. The per-agent "Default Model" values below describe runtime-safe defaults used when no explicit model config is provided.

---

🏛️ Meet the Pantheon

Primary Agents

Orchestrator, Planner, Sprinter, and Business Analyst are the primary agents. Choose one based on how you want to work.

• Orchestrator (default): Delegation-first coordinator for planning, routing, and result integration.
• Planner: Interview-first planner that asks clarifying questions and returns structured <planner-plan> output.
• Sprinter: Fast self-executing agent for quick Q&A and direct tasks.
• Business Analyst: Analysis specialist for market research, competitive analysis, requirements elicitation, and strategic planning.

Routing Flow

• Orchestrator can delegate to debugger, explorer, librarian, oracle, designer, frontend-developer, backend-developer, trigger-developer, observer, and council.
• Planner is planning-only and can delegate only to explorer, librarian, oracle, and designer.
• Sprinter is self-executing and does not delegate.
• Business Analyst can delegate research to explorer, librarian, and oracle.
• Specialists are leaf executors: once delegated to, they do the bounded work and hand results back.
• Observer is disabled by default until you explicitly enable it in config.
• Council is available, but intentionally expensive and kept on a stricter path than normal delegation.

Orchestrator

Role: Delegation-first coordinator
Prompt: orchestrator.ts
Default Model: openai/gpt-5.4-fast (high)
Recommended Models: openai/gpt-5.5, anthropic/claude-opus-4.7
Best Choice: openai/gpt-5.5-fast with variant high — this is the strongest single configuration for the orchestrator role, offering the best balance of reasoning speed and routing accuracy.
Model Guidance: Choose your strongest coordination model. Orchestrator should excel at routing, delegation discipline, judgment, and reliable instruction-following. Orchestrator delegates ALL substantive work to specialists and only acts directly when a subagent's "Don't delegate when" rule explicitly applies, or for integration/verification tasks.

Planner

Role: Interview-first planner that asks clarifying questions and returns structured <planner-plan> output
Prompt: planner.ts
Default Model: openai/gpt-5.5-fast (xhigh)
Recommended Models: openai/gpt-5.5, anthropic/claude-opus-4.7
Model Guidance: Choose your strongest all-around coding model. Planner drives planning and delegation, so it needs excellent judgment, structured thinking, and reliable instruction-following. Planner delegates all exploration and research to specialists and only acts directly for synthesis, interviewing, and plan production.

Sprinter

Role: Fast self-executing agent for quick Q&A and direct tasks
Prompt: sprinter.ts
Default Model: openai/gpt-5.3-codex (low)
Recommended Models: openai/gpt-5.3-codex, github-copilot/grok-code-fast-1, kimi-for-coding/k2p5
Model Guidance: Choose a fast, low-latency model. Sprinter handles everything directly and does not delegate — use it when you want direct answers and quick execution rather than heavy planning or delegation.

Business Analyst

Role: Market research, competitive analysis, requirements elicitation, and strategic planning specialist
Prompt: business-analyst.ts
Default Model: openai/gpt-5.5-fast (high)
Recommended Models: openai/gpt-5.5, anthropic/claude-opus-4.7
Model Guidance: Choose a strong reasoning model for structured analysis, research synthesis, and documentation generation. Business Analyst delegates research to @explorer, @librarian, and @oracle, then synthesises findings into actionable plans and requirements documents.

Auto-save behavior: The Business Analyst always saves its full analysis output as a .md file under .business-analyst/ and returns only a concise confirmation in chat. This differs from the Planner, which only saves to file when explicitly requested.

---

Subagents

The following agents are delegated to by the primary agents based on task type.

Oracle

Role: Strategic advisor and escalation point for high-stakes decisions, unresolved bugs, and code review
Prompt: oracle.ts
Default Model: openai/gpt-5.5-fast (high)
Recommended Models: openai/gpt-5.5 (high), google/gemini-3.1-pro-preview (high)
Model Guidance: Choose your strongest high-reasoning model for architecture review, escalation debugging, trade-offs, and code review. First-pass bug investigation should go to @debugger — route to Oracle only when the bug persists after initial investigation or has architectural implications.

Debugger

Role: Bug investigation specialist — finds root causes without implementing fixes
Prompt: debugger.ts
Default Model: openai/gpt-5.3-codex (high)
Recommended Models: openai/gpt-5.4-mini, minimax-coding-plan/MiniMax-M2.7
Model Guidance: Choose a capable coding model for systematic debugging. Debugger is read-only and focused on investigation — it traces error paths, analyzes root causes, and reports findings for implementation agents to act on. It does NOT implement fixes.

Explorer

Role: Codebase reconnaissance
Prompt: explorer.ts
Default Model: minimax-coding-plan/MiniMax-M2.7
Recommended Models: fireworks-ai/accounts/fireworks/routers/kimi-k2p5-turbo, openai/gpt-5.4-mini
Model Guidance: Choose a fast, low-cost model. Explorer handles broad scouting work, so speed and efficiency usually matter more than using your strongest reasoning model.

Librarian

Role: External knowledge retrieval (docs, web, Confluence, Jira)
Prompt: librarian.ts
Default Model: minimax-coding-plan/MiniMax-M2.7
Recommended Models: fireworks-ai/accounts/fireworks/routers/kimi-k2p5-turbo, openai/gpt-5.4-mini
Model Guidance: Choose a fast, low-cost model. Librarian handles research and documentation lookups, so speed and efficiency usually matter more than using your strongest reasoning model.

Designer

Role: UI/UX direction, layout/interaction decisions, visual polish, and accessibility judgment
Prompt: designer.ts
Default Model: opencode-go/kimi-k2.6
Recommended Models: google/gemini-3.1-pro-preview, kimi-for-coding/k2p5
Model Guidance: Choose a model strong at UI/UX direction, layout/interaction judgment, visual polish, and design decision-making. Designer serves as the spec/decision authority; implementation work with clear direction goes to @frontend-developer.

Frontend Developer

Role: Client-side implementation and frontend tests — executes what @designer decides
Prompt: frontend-developer.ts
Default Model: opencode-go/deepseek-v4-flash (high)
Recommended Models: google/gemini-3.1-pro-preview, kimi-for-coding/k2p5
Model Guidance: Choose a model strong at client-side implementation, component architecture, and styling execution. Receives bounded frontend tasks from Orchestrator once design direction is established.

Trigger.dev Developer

Role: Trigger.dev implementation specialist — implements tasks, config, schedules, realtime progress, and integrations
Prompt: trigger-developer.ts
Default Model: opencode-go/deepseek-v4-flash (high)
Recommended Models: openai/gpt-5.4-mini, opencode-go/deepseek-v4-flash
Model Guidance: Choose a fast, reliable coding model. Receives bounded Trigger.dev implementation tasks from Orchestrator such as task definitions, trigger configuration, schedule setup, and API integrations.

Backend Developer

Role: Backend implementation specialist
Prompt: backend-developer.ts
Default Model: opencode-go/deepseek-v4-flash (high)
Recommended Models: cerebras/zai-glm-4.7, fireworks-ai/accounts/fireworks/routers/kimi-k2p5-turbo, openai/gpt-5.4-mini
Model Guidance: Choose a fast, reliable coding model for routine backend tasks. Receives bounded server-side tasks from Orchestrator such as API implementation, database work, and service logic changes.

Council

[!NOTE] > Why doesn't Orchestrator auto-call Council more often? This is intentional. Council runs multiple models at once, so automatic delegation is kept strict because it is usually the highest-cost path in the system. In practice, Council is meant to be used manually when you want it, for example: @council compare these two architectures.

Role: Multi-LLM consensus and synthesis
Prompt: council.ts
Guide: docs/council.md
Default Setup: Config-driven — councillors come from council.presets and the Council agent model comes from your normal council agent config
Recommended Setup: Strong Council model + diverse councillors across providers
Model Guidance: Use a strong synthesis model for the Council agent and diverse models as councillors. The value of Council comes from comparing different model perspectives, not just picking the single strongest model everywhere.

Observer

[!NOTE] > Why a separate agent? If your Orchestrator model is not multimodal, enable Observer to handle images, screenshots, PDFs, and other visual files. Observer is disabled by default and gives the Orchestrator a dedicated multimodal reader without forcing you to change your main reasoning model. Set disabled_agents: [] and an observer model in your configuration.

Role: Read-only visual analysis — interprets images, screenshots, PDFs, and diagrams. Returns structured observations to the orchestrator without loading raw file bytes into the main context window.

• Images, screenshots, diagrams → read tool (native image support)
• PDFs and binary documents → read tool (text + structure extraction)
• Disabled by default — enable with "disabled_agents": [] and configure a vision-capable model

Prompt: observer.ts
Default Model: openai/gpt-5.4-mini — configure a vision-capable model to enable
Model Guidance: Choose a vision-capable model if you want the agent to read screenshots, images, PDFs, and other visual files.

---

📚 Documentation

Use this section as a map: start with installation, then jump to features, configuration, or example presets depending on what you need.

🚀 Start Here

Doc	What it covers
Installation Guide	Install the plugin, use CLI flags, reset config, and troubleshoot setup
Low-tech Setup Guide	Step-by-step setup for non-technical users (commands + success checks)
Quick Reference	Jump table for install, configuration, skills, MCPs, tools, and presets

✨ Features & Workflows

Doc	What it covers
Council	Run multiple models in parallel and synthesize a single answer with @council
Session Management	Reuse recent child-agent sessions with short aliases instead of starting over
Codemap	Generate hierarchical codemaps to understand large codebases faster

⚙️ Config & Reference

Doc	What it covers
Configuration	Config file locations, JSONC support, prompt overrides, and full option reference
Skills	Built-in and recommended skills such as simplify, agent-browser, and codemap
MCPs	websearch, context7, grep_app, figma, serena, atlassian, and how MCP permissions work per agent
Tools	Built-in tool capabilities like webfetch, LSP tools, code search, and formatters

💡 Example Presets

Doc	What it covers
$30 Preset	A budget mixed-provider setup for around $30/month
Minimax Preset	Minimax M2.7 Preset

---

📄 License

MIT

---