25 AI agents. 42 skills. One plugin.
omg turns GitHub Copilot CLI into a multi-agent development platform. Instead of one AI assistant doing everything, omg gives you a team of specialists — an architect who analyzes code, a debugger who traces root causes, a critic who reviews plans, an executor who implements changes — all coordinating autonomously through structured workflows.
| Without omg | With omg |
|---|---|
| One generalist AI | 19 specialists, each best at their job |
| Manual step-by-step prompting | Autonomous multi-phase workflows |
| Hope it works, check manually | Built-in verification with evidence |
| Forget context between sessions | Persistent memory across sessions |
| Single model for everything | Right model per task (haiku for speed, opus for depth) |
One command:
copilot plugin install TheTrustedAdvisor/omgThat's it. 25 agents + 42 skills, ready to use.
Alternative install methods
# SSH (if HTTPS doesn't work)
copilot plugin install git@github.com:TheTrustedAdvisor/omg.git
# From local clone (development)
git clone https://github.com/TheTrustedAdvisor/omg-dev.git
cd omg-dev
./install.sh
# Temporary (no install, session only)
copilot --plugin-dir ./pluginNo special syntax needed. Describe what you want:
# Plan a feature with structured requirements
copilot -i "Plan how to add authentication to this project"
# Fix a bug — root cause analysis, not guesswork
copilot -i "Debug why the tests are failing in src/pipeline/"
# Security audit with severity ratings and fix suggestions
copilot -i "Review the security of src/config.ts"
# Build something from scratch — fully autonomous
copilot -i "autopilot: build a REST API for user management"
# Keep working until it's done — no "it should work"
copilot -i "ralph: add input validation to all CLI commands"
# 5 agents working in parallel on independent tasks
copilot -i "team 5: fix all TypeScript errors across src/"
# Investigate → implement → create PR (Copilot-exclusive)
copilot -i "research-to-pr: fix the auth token expiry bug"Copilot picks the right agents and skills automatically.
Works in any language — speak German, French, Japanese, or any language. omg translates your intent and responds in your language:
copilot -i "omg prüfe die Sicherheit dieses Projekts" # German → security review
copilot -i "omg vérifie la sécurité de ce projet" # French → security audit
copilot -i "omg このプロジェクトのセキュリティを確認して" # Japanese → security check| Say this | What happens |
|---|---|
plan... |
Structured interview → plan with testable acceptance criteria |
autopilot... |
Full lifecycle: plan → implement → test → validate |
ralph... |
Persistent loop until ALL acceptance criteria pass with evidence |
team N... |
N parallel agents, each on independent files, verified together |
trace... |
Competing hypotheses ranked by evidence strength |
deep interview... |
Socratic Q&A with ambiguity scoring — proceed when clarity ≥ 80% |
tdd... |
Red-green-refactor: failing test first, then minimal code |
research-to-pr... |
Investigate → implement → cloud agent creates PR automatically |
deepsearch... |
Multi-angle codebase exploration, not just grep |
ultrathink... |
Deep reasoning on complex architectural decisions |
Not just prompts — agents delegate to each other, persist handoffs to files, and verify each other's work. Here's what happens when you say autopilot: add authentication:
sequenceDiagram
participant U as You
participant A as Autopilot
participant AN as Analyst (opus)
participant AR as Architect (opus)
participant CR as Critic (opus)
participant EX as Executor (sonnet)
participant VR as Verifier (sonnet)
participant SR as Security (opus)
U->>A: autopilot: add authentication
rect rgba(100, 130, 240, 0.15)
Note over A,AN: Phase 1 — Expand
A->>AN: Requirements analysis
AN-->>A: Gaps, edge cases, guardrails
end
rect rgba(100, 200, 100, 0.15)
Note over A,CR: Phase 2 — Plan
A->>AR: Create implementation plan
AR-->>A: Plan with acceptance criteria
A->>CR: Evaluate plan quality
CR-->>A: ACCEPT (or REVISE → loop)
end
Note over A: Plan saved to .omg/plans/
rect rgba(240, 150, 80, 0.15)
Note over A,EX: Phase 3 — Execute
A->>EX: Implement (parallel tasks)
EX-->>A: Code changes + test results
end
rect rgba(180, 100, 220, 0.15)
Note over A,SR: Phase 4 — QA (parallel)
A->>AR: Verify completeness
A->>SR: Security audit
A->>VR: Evidence-based verification
AR-->>A: Architecture OK
SR-->>A: No vulnerabilities
VR-->>A: PASS (all criteria met)
end
A->>U: Done. Evidence in .omg/reviews/
| Traditional AI | omg Orchestration |
|---|---|
| One model does everything | 6 specialists, each best at their job |
| "It should work" | Verifier ran npm test and saw green |
| Forgets context between steps | Plans, research, reviews persist in .omg/ |
| Single pass, hope for the best | Critic rejects → Planner revises → loop until approved |
| Same model for search and architecture | Haiku for speed, Opus for depth |
omg provides several orchestration patterns, each for different needs:
graph TD
U[User Request] --> D{What kind of task?}
D -->|Vague idea| DI[deep-interview]
D -->|Clear task| P[plan]
D -->|Code it now| R[ralph]
D -->|Full lifecycle| AP[autopilot]
D -->|Multiple files| T[team N]
D -->|Bug| DB[debug]
DI -->|Spec ready| P
P -->|Plan approved| R
P -->|Complex| AP
AP --> R
R --> UW[ultrawork]
T --> UW
style AP fill:#4a9eff,color:#fff
style R fill:#ff6b6b,color:#fff
style T fill:#51cf66,color:#fff
style DI fill:#ffd43b,color:#000
style P fill:#cc5de8,color:#fff
style DB fill:#ff922b,color:#fff
style UW fill:#868e96,color:#fff
autopilot wraps ralph (persistence + verification), which wraps ultrawork (parallel execution). Each layer adds capability:
| Layer | What it adds | When to use |
|---|---|---|
| ultrawork | Parallel agent dispatch | Independent tasks, speed matters |
| ralph | Persistence loop + verification | Must complete with evidence |
| autopilot | Full lifecycle (plan→execute→QA) | From idea to working code |
| team N | N parallel workers + coordination | Large tasks, multiple files |
Every claim is backed by evidence. "Tests pass" means the verifier ran npm test and saw green. No "it should work" — only "here's the proof."
Decisions, plans, and review results survive across sessions:
| What | Where | Survives restart? |
|---|---|---|
| Plans | .omg/plans/ |
Yes (file) |
| Research | .omg/research/ |
Yes (file) |
| Reviews | .omg/reviews/ |
Yes (file) |
| Context index | store_memory |
Yes (Copilot-native) |
Features that only work on Copilot, not on other AI coding tools:
| Capability | What it enables |
|---|---|
| Multi-model routing | Haiku for speed, Sonnet for coding, Opus for architecture |
| Parallel agents | 5 workers on 5 files simultaneously (proven: 3s not 9s) |
| Cloud handoff | /delegate sends work to cloud agent → automatic PR |
| GitHub integration | Native MCP access to issues, PRs, code search |
| Reasoning control | --effort low/xhigh per task for cost optimization |
| Agent | Specialty | When it's called |
|---|---|---|
| autopilot | Full lifecycle: idea → working code | "autopilot", "build me", "handle it all" |
| ralph | Persistence loop until criteria pass | "ralph", "don't stop", "must complete" |
| team | N parallel workers on independent tasks | "team 3", "assemble a team" |
| ralplan | Consensus: planner → architect → critic | "ralplan", "consensus" |
| ultrawork | Fire parallel tasks simultaneously | "ultrawork", "ulw" |
| research-to-pr | Investigate → cloud agent → auto PR | "research and fix", "investigate and PR" |
| sciomc | Staged parallel research with scientists | "sciomc", "research" |
| self-improve | Benchmark, tournament-select best approach | "self-improve", "optimize" |
| deep-dive | Trace root cause → crystallize into spec | "deep dive", "investigate deeply" |
| Agent | Specialty | When it's called |
|---|---|---|
| architect | Architecture analysis, trade-offs | Plan review, design decisions |
| critic | Multi-perspective quality gate | Final approval before execution |
| analyst | Requirements gaps, edge cases | Before planning starts |
| planner | Structured plans with acceptance criteria | "plan", "how should we" |
| executor | Code implementation, smallest diff | Actual coding work |
| debugger | Root cause analysis, minimal fix | Bugs, build failures |
| verifier | Evidence-based PASS/FAIL | After every implementation |
| code-reviewer | Severity-rated review, SOLID | Code quality checks |
| security-reviewer | OWASP Top 10, secrets, deps | Security audits |
| test-engineer | TDD, test strategy, flaky tests | Test writing |
| explore | Fast codebase search | Finding files and patterns |
| designer | Production-grade UI | Frontend work |
| git-master | Atomic commits, style-matching | Git operations |
| tracer | Competing hypotheses with evidence | "Why did this happen?" |
| scientist | Statistical analysis with rigor | Data investigations |
| writer | Verified documentation | README, API docs |
| document-specialist | External API references | Library/framework questions |
| code-simplifier | Remove complexity, keep behavior | Refactoring |
| qa-tester | Interactive CLI testing | End-to-end validation |
Execution: autopilot, ralph, team, ultrawork Planning: plan, ralplan, deep-interview, deep-dive Quality: ultraqa, verify, tdd, ai-slop-cleaner Analysis: trace, deep-analyze, deepsearch, ultrathink, sciomc Flagship: research-to-pr (investigate → cloud agent → automatic PR) Utilities: debug, release, cancel, session-memory, ask, ccg, deepinit, external-context, learner, skillify, self-improve, writer-memory, project-session-manager, configure-notifications, visual-verdict Reference: agent-catalog, handoff-protocol
- Architecture — How it works under the hood + empirical findings
- Copilot Capabilities — Every Copilot feature omg uses
- Examples — 10 copy-paste demos with expected output
- Best Practices — Dos, don'ts, documentation standards
- Limitations — Known gaps and platform limitations
- All docs — Full documentation index
omg's agent architecture is inspired by oh-my-claudecode (OMC), adapted and extended for GitHub Copilot's native capabilities. omg adds multi-model routing, parallel execution, cloud delegation, and GitHub-native integration that OMC's Claude Code platform cannot provide.
MIT