DISCOVER THE FUTURE OF AI AGENTSarrow_forward

Oh My OpenAgent

calendar_todayAdded Apr 23, 2026
categoryAgent & Tooling
codeOpen Source
TypeScriptMulti-Agent SystemModel Context ProtocolAI AgentsCLIBunAgent & ToolingModel & Inference FrameworkDeveloper Tools & CodingProtocol, API & Integration

An OpenCode-based multi-model orchestrated AI coding agent plugin featuring 11 specialized parallel agents, Hashline safe editing, LSP/AST-Grep toolchain, and a 3-tier MCP system, positioned as an open alternative to Claude Code.

Positioning#

Oh My OpenAgent is an AI coding agent plugin built on the OpenCode framework (v1.0.150+). It is not a standalone LLM — it relies on external model APIs (Claude, GPT, Gemini, GLM, Kimi, MiniMax, etc.) — and is designed as an open alternative to closed solutions like Claude Code.

Multi-Agent Orchestration#

The project features 11 specialized agents, each optimized for specific model characteristics:

  • Sisyphus (master orchestrator): defaults to claude-opus-4-7 / kimi-k2.5 / glm-5, handles planning and delegation
  • Hephaestus (autonomous deep worker): defaults to gpt-5.4, end-to-end exploration and execution
  • Prometheus (strategic planner): interview mode, generates detailed plans before coding
  • Oracle (architecture/debugging), Librarian (doc/code search), Explore (fast codebase search), Atlas, Metis, Momus, Multimodal-Looker, Sisyphus-Junior

Agent category routing: visual-engineering (frontend/UI/UX), deep (autonomous research + execution), quick (single-file edits), ultrabrain (complex logic/architecture decisions).

Key Innovations#

Hash-Anchored Edit Tool (Hashline)#

Inspired by oh-my-pi, each line carries a content hash identifier (e.g., 11#VK| function hello()). Edits verify the hash to prevent stale-edit file corruption. The README claims this improved Grok Code Fast edit success rate from 6.7% to 68.3% (independently unverified).

ultrawork / Ralph Loop#

  • ultrawork (or ulw): single command to activate all agents in parallel
  • /ulw-loop (Ralph Loop): self-referencing loop that doesn't stop until the task is 100% complete

3-Tier MCP System#

  1. Built-in remote MCP: Exa (web search), Context7 (official docs), Grep.app (GitHub search)
  2. Project-level MCP: configured via .mcp.json
  3. Skill-Embedded MCP: skills ship with their own MCP servers, launched on-demand and released after task completion without consuming context window

IDE-Grade Toolchain Integration#

  • LSP: lsp_rename, lsp_goto_definition, lsp_find_references, lsp_diagnostics
  • AST-Grep: pattern-aware code search and rewrite supporting 25 languages
  • Tmux: full interactive terminal integration (REPL, debuggers, TUI apps)
  • IntentGate: analyzes true user intent before classification/execution

Claude Code Compatibility#

Fully compatible with Claude Code's hooks, commands, skills, MCPs, and plugins for seamless migration.

/init-deep#

Automatically generates hierarchical AGENTS.md files (project → directory → component level), with agents auto-loading relevant context.

Built-in Skills#

  • playwright: browser automation
  • git-master: atomic commits, rebase operations
  • frontend-ui-ux: design-first UI development
  • Custom skills supported via .opencode/skills/*/SKILL.md or ~/.config/opencode/skills/*/SKILL.md

Architecture Overview#

Written in TypeScript (1,766 source files, ~377k LOC) using the Bun runtime and build system. CLI framework: Commander.js. Config validation: Zod v4 (32 schemas). MCP protocol: @modelcontextprotocol/sdk. Ships 11 precompiled platform binary packages (including AVX2 + baseline variants) — no additional runtime needed post-install.

Initialization flow: loadPluginConfig()createManagers() (TmuxSessionManager, BackgroundManager, SkillMcpManager, ConfigHandler) → createTools() (26 tools) → createHooks() (52 hooks: Core 43 + Continuation 7 + Skill 2) → createPluginInterface() (10 OpenCode hook handlers).

Runs as an @opencode-ai/plugin via a 6-stage configuration pipeline (provider → plugin-components → agents → tools → MCPs → commands).

Installation & Usage#

Prerequisite: OpenCode (v1.0.150+) must be installed first.

Recommended install (via LLM Agent):

Install and configure oh-my-opencode by following the instructions here:
https://raw.githubusercontent.com/code-yeongyu/oh-my-openagent/refs/heads/dev/docs/guide/installation.md

Manual install: bunx oh-my-opencode install

Supported platforms: macOS (ARM64, x64), Linux (x64, ARM64, Alpine/musl), Windows (x64)

Verify: bunx oh-my-opencode doctor

After install, confirm "oh-my-openagent" exists in the plugin array of ~/.config/opencode/opencode.json. Config files support user-level (~/.config/opencode/oh-my-openagent.json[c]) and project-level (.opencode/oh-my-openagent.json[c]), with backward compatibility for the old name oh-my-opencode.json[c].

External Integrations#

  • OpenClaw: bidirectional external integration (Discord/Telegram/webhook/command)
  • Jobdori: author's AI assistant for real-time build and maintenance

Scale Metrics#

5,237+ commits, 75 branches, 185 tags, 19 feature modules, 104 barrel index.ts files. Current version: v3.17.4.

License#

SUL-1.0 (Sustainable Use License): permits non-commercial/personal use and internal commercial use; prohibits paid distribution.

Unverified Information#

  • No independent website found; docs are embedded in the GitHub repo's docs/ directory
  • npm package publish status: package name is oh-my-opencode per package.json, npm URL inferred but not directly verified
  • Some default model names (e.g., claude-opus-4-7, gpt-5.4) may not have been publicly released at the time of collection — unclear if placeholder or forward-looking config
  • README claims Anthropic blocked OpenCode because of this project — unverifiable here
  • Discord community link not directly provided

Related Projects

View All arrow_forward

STAY UPDATED

Get the latest AI tools and trends delivered straight to your inbox. No spam, just intelligence.

rocket_launch