Oh My OpenAgent — The Best Agent Harness
63.1k stars. 3.1M+ downloads. Battle-tested and production-proven.
$bunx oh-my-openagent install
The Magic Word
ulw
Type Three Letters. Walk Away.
Ultra Work mode triggers maximum precision: auto-planning, deep research, parallel agents, and self-correction loops. The system doesn't stop until it's done. You don't babysit.
Auto-PlanningDeep ResearchSelf-CorrectionParallel Agents
Zero intervention. Full autonomy. Just results.
$~_
Scanning codebase...
Launching 3 explore agents...
Building execution plan...
Deploying 4 parallel agents...
Running verification suite...
okDone. 847 lines changed, 0 issues.
$~_
Primary OrchestratorClaude Opus 4.7 Max
Sisyphus
The CTO Who Never Sleeps
Named after the mythological figure rolling his boulder uphill forever. Sisyphus parses implicit requirements, adapts to your codebase's maturity, and delegates to specialists. Interrupted? The boulder system resumes exactly where you left off.
Parses what you meant, not just what you typed
Maps your architecture before touching a line
Routes to the right specialist agent
Trusts nothing. Verifies everything.
Session Continuity
Active work tracked in boulder.json. Power outage? System crash? Doesn't matter. Pick up exactly where you stopped.
Orchestration Workflow
Think, Then Act
Planning and Execution Are Different Jobs
Prometheus
Strategic Planner
The architect. Interviews you, explores your codebase, and creates detailed battle plans. Never writes code. Uses Metis and Momus as quality gates.
- Intelligent interview mode
- Multi-agent codebase exploration
- Gap analysis with Metis
- Ruthless review with Momus
The builder. Reads verified plans, delegates to specialized agents via category+skills system. Tracks learnings across tasks. Verifies independently — never trusts subagent claims.
- Intent-based task routing
- Wisdom accumulation across tasks
- Boulder session continuity
- Independent result verification
1
Describe your work to Prometheus
3
Momus validates ruthlessly
4
Atlas delegates to specialists
5
Independent verification on everything
The planner doesn't code. The executor follows verified plans. Intelligence lives in the system, not individual agents.
Deep WorkerGPT 5.5 Medium
Hephaestus
Give Him a Goal, Not a Recipe
Named after the Greek god of the forge. Methodical, thorough, obsessive. Fires 2-5 parallel explore agents before writing any code. Built for deep architectural reasoning, complex debugging, and cross-domain synthesis.
EXPLORE — Map the terrain
DECIDE — Commit to the path
EXECUTE — Build with precision
For when good enough isn't.
NEW · v4.0opt-in
Team Mode
Multiple Agents. Side-by-Side. In Parallel.
A lead agent orchestrates a team of category-specialized members, all running in parallel and communicating through dedicated tools. Watch every member work simultaneously in tmux. Opt-in — flip team_mode.enabled and restart.
One coordinator. Multiple specialists. Real-time messaging via team_send_message and team_task_create.
Up to 8 members work simultaneously. No waiting, no queues — every angle covered at once.
Focus and grid windows let you zoom into one teammate without losing the bigger picture.
12 dedicated tools (team_create, team_status, team_send_message, ...) gated behind team_mode.enabled.
5 hostile agents tear apart your plan from orthogonal angles before a single line of code is written.
3 vulnerability hunters + 2 PoC engineers audit your codebase. Severity calibrated by actual exploitability.
team_mode.enabled = true → restart opencode
Off by default. Devastating when on.
Specialized Agents
Five built-in agents for common workflows. Plus a dynamic system that assembles custom agents on demand.
Oracle
Architecture Consultant
Complex debugging and architectural decisions. When the path forward isn't obvious.
Librarian
Docs & Code Search
Finds real GitHub examples and official docs. Evidence-based, with permalinks.
Blazing fast codebase search. Cheap, parallel, always background.
Catches ambiguities in plans before they become production bugs.
Ruthless validation. Only approves when the plan is bulletproof.
Assembled for the task at hand
Dynamic Agent
Matches your request to the right model, loads the necessary skills, and builds an agent in real-time.
Category Routing
visual-engineering
Gemini 3.1 Pro
Skill Injection
"The right model + right expertise, every time."
Built Different
Design decisions that actually matter.
Each agent does one thing exceptionally well. No jacks of all trades.
Orchestrator runs independent verification on everything. Subagents don't get a free pass.
Learnings from each task passed to all subsequent tasks. The system gets smarter as it works.
Expensive models for planning and complex decisions. Cheap models for routine work. Maximum output per dollar.
Intent-based routing. Say what you need — ultrabrain, visual-engineering, quick — and the right model handles it.
Boulder system: interrupted work resumes exactly where it stopped. Zero context loss.
What Devs Actually Say
“It made me cancel my Cursor subscription. Unbelievable things are happening in the open source community.”
— Arthur Guiot
“If Claude Code does in 7 days what a human does in 3 months, Sisyphus does it in 1 hour.”
— B, Quant Researcher
“Knocked out 8000 eslint warnings with Oh My Opencode, just in a day.”
— Jacob Ferrari
“use oh-my-openagent, you will never go back”
— d0t3ch
“You guys should pull this into core and recruit him. Seriously. It's really, really, really good.”
— Henning Kilset
“Oh My OpenCode Is Actually Insane”
— Darren Builds AI (YouTube)
Stop Writing Boilerplate
Install Oh My OpenAgent and let the agents handle it. Your first ulw command is 30 seconds away.
$bunx oh-my-openagent install