Universal AI coding workflow. One install. Every platform.
Works with Claude Code · Codex · Cursor · OpenCode · Gemini CLI · Antigravity
oh-my-colab is a developer workflow framework that adds structured coding methodology, multi-agent execution, persistent memory, and reliable skill triggering to every major AI coding platform — simultaneously.
Your AI agent stops guessing. It starts following a proven workflow: design → plan → implement → verify → ship.
install once → omcl setup
work anywhere → claude, codex, cursor, gemini, antigravity, opencode
keep context → MEMORY.md survives compaction and session switches
never lose flow → 20 skills auto-trigger from natural language
| Without oh-my-colab | With oh-my-colab |
|---|---|
| Agent jumps to code immediately | Design always happens before implementation |
| Platform-locked workflows (OMC is Claude-only, OMX is Codex-only) | Same skills, agents, hooks on all 6 platforms |
| Agent loses context after compact or session switch | MEMORY.md + snapshot hooks survive any interruption |
| Skills fail to trigger 30–50% of the time | 4-layer trigger system: 100% activation in testing |
| 20 terminal windows, no idea what each agent is doing | Unified HUD shows mode, agent, queue, cost |
| Switch platforms, lose your whole setup | One omcl update refreshes everything everywhere |
Step 1: Install
pnpm add -g oh-my-colab
# or: npm install -g oh-my-colabStep 2: Setup (run in your project)
omcl setupDetects your platforms, configures each one, generates context files, installs MCP servers, LSP plugins, and stack-matched skills from autoskills. Takes ~90 seconds.
Step 3: Start coding
design this feature first → feature-design skill triggers
build me a REST API for tasks → autopilot mode activates
why is this test failing → root-cause-debug skill triggers
write tests for the auth module → tdd-execution skill triggers
No commands to memorize. Natural language activates the right workflow automatically.
| Platform | Method |
|---|---|
| Claude Code | /plugin marketplace add iadr-dev/oh-my-colab then /plugin install oh-my-colab |
| Codex CLI | Tell Codex: Fetch and follow https://raw.githubusercontent.com/iadr-dev/oh-my-colab/main/.codex/INSTALL.md |
| Cursor | omcl setup --platform cursor |
| Antigravity | omcl setup --platform antigravity |
| Gemini CLI | gemini extensions install github.com/iadr-dev/oh-my-colab |
| OpenCode | omcl setup --platform opencode |
| All at once | omcl setup (auto-detects everything installed) |
After setup: start a new session and say anything. oh-my-colab activates automatically.
oh-my-colab enforces a structured development process through 7 skill-driven phases. The agent cannot skip phases — every step has a gate.
1. DESIGN feature-design skill
Socratic interview → saves .omcl/specs/feature-YYYYMMDD.md
↓ spec approved by you
2. PLAN implementation-plan skill
Breaks spec into 5-15 min tasks with exact file paths
Saves .omcl/plans/plan-YYYYMMDD.md
↓ plan approved by you
3. BRANCH branch-workflow skill
Creates git worktree + branch, verifies clean test baseline
↓ clean baseline confirmed
4. IMPLEMENT tdd-execution skill (per task)
RED → write failing test first
GREEN → minimal code to pass
REFACTOR → clean up
Runs in parallel via team-mode or sequential per task
↓ all tasks complete
5. VERIFY verify-completion skill
Tests pass, no regressions, plan items satisfied, edge cases handled
Blocks progress until every item checked
↓ verification passes
6. REVIEW pre-review skill
Coverage, naming, docs, no debug code left
↓ review passes
7. SHIP finish-branch skill
Merge / PR / keep / discard — your choice
Cleans up worktree
Skills load automatically when you describe what you want. No /commands required.
| Skill | Triggers on |
|---|---|
feature-design |
"I want to build", "add feature", "implement", "design" |
implementation-plan |
"break this down", "plan the tasks", after spec approval |
tdd-execution |
"write tests", "add test for", "test-driven" |
parallel-development |
"run in parallel", "use multiple agents", "team" |
verify-completion |
"are we done", "check completion", before claiming done |
branch-workflow |
"create branch", "start new feature", "git worktree" |
finish-branch |
"merge this", "create PR", "wrap up branch" |
| Skill | Triggers on |
|---|---|
root-cause-debug |
"error", "failing", "broken", "exception", "why isn't this" |
pre-review |
"before review", "ready for PR", "pre-commit check" |
post-review |
"address review", "reviewer said", "fix feedback" |
security-audit |
"security check", "audit", "vulnerabilities", "OWASP" |
| Skill | Triggers on |
|---|---|
session-pickup |
session start, "I'm back", "where were we", after compact |
context-checkpoint |
"save context", "going to compact", "checkpoint" |
decision-record |
"architecture decision", "we decided to", "record this" |
| Skill | Triggers on |
|---|---|
codebase-map |
first session, "how does this work", "explain this codebase" |
live-docs |
any external library in code (auto-invokes context7) |
| Skill | Triggers on |
|---|---|
team-mode |
"spawn agents", "parallel team", "N agents" |
agent-handoff |
"hand this off", "continue in another session", "pass context" |
| Skill | Triggers on |
|---|---|
write-skill |
"create a skill", "new skill for", "add to oh-my-colab" |
omcl-guide |
"how does oh-my-colab work", "what skills are available" |
Invoke via natural language or slash commands.
| Mode | What It Does | Say |
|---|---|---|
autopilot |
Full autonomous: design → plan → implement → verify | autopilot: build a REST API |
parallel |
N agents working concurrent tasks | run 3 agents on this |
eco |
Token-efficient: routes simple tasks to faster models | eco: refactor this module |
ralph |
Persistence: won't stop until architect signs off | ralph: fix all type errors |
pipeline |
Sequential chain: output of each agent feeds the next | pipeline: analyze → implement → test |
review-loop |
Verify → fix → verify (bounded retries) | review-loop on this PR |
oh-my-colab dispatches the right agent based on the task. Each has a defined role, model tier, and output format.
Build & Analysis: architect · planner · analyst · executor · deep-executor
debugger · verifier · explorer
Review: code-reviewer · security-reviewer · performance-reviewer
quality-reviewer · style-reviewer
Domain: test-engineer · build-fixer · researcher · git-master
dependency-expert · writer
Coordination: critic · orchestrator
Provider routing (configurable in .omcl/config.yaml):
- High tier (architect, critic, planner): GPT-4.5 / Opus 4.7 / Gemini Pro 3.1
- Standard tier (executor, reviewer, debugger): Sonnet 4.7 / GPT-4o / Gemini Flash 3.1
- Fast tier (explorer, writer, style): Haiku / o4-mini / Gemini Flash 2.5
oh-my-colab never loses context, even across compaction or session switches.
Session starts → session-start hook injects SOUL.md + USER.md + MEMORY.md
(agent knows exactly who you are and where work left off)
During session → agent writes to MEMORY.md after significant actions
(current task, branch, last change, known issues)
Before compact → pre-compact hook saves full snapshot to .omcl/memory/
(task state, files changed, test results, next step)
After compact → post-compact hook injects resume context
(agent picks up exactly where it left off)
Session ends → session-end hook updates MEMORY.md + indexes session in SQLite
(searchable: omcl memory search "auth bug we fixed")
oh-my-colab ships 3 MCP servers installed automatically during setup.
| Server | What It Provides |
|---|---|
omcl-memory |
Cross-session memory: set/get/search (FTS5) + snapshot/resume |
omcl-state |
Execution mode tracking, task queue, HUD data feed |
omcl-codebase |
Dependency graph, blast radius, code ownership, health score |
Plus these community MCP servers are installed during omcl setup:
| Server | Why |
|---|---|
| context7 | Up-to-date library docs injected automatically |
| mcp-server-git | git log, diff, blame, branch operations |
| mcp-server-github | PR creation, issues, code search |
| mcp-sequential-thinking | Structured multi-step reasoning |
| mcp-server-filesystem | File operations with access control |
Real-time status line showing what your agents are doing.
omcl hud # interactive Ink TUI
omcl hud --watch # polling mode (refresh every 2s)
omcl hud --json # JSON output for tmux/statusbar integration
omcl hud --preset minimalClaude Code and Codex get a native HUD injected into their status bar via omcl setup.
omcl setup [--yes] [--dry-run] [--platform PLATFORM]
omcl doctor [--platform PLATFORM] [--fix]
omcl skills list
omcl skills validate
omcl sessions list
omcl sessions resume <id>
omcl memory search <query>
omcl memory show
omcl hud [--watch] [--json] [--preset NAME]
omcl analyze [--blast-radius <file>] [--hotspots] [--health]
omcl parallel --n N "task description"
omcl wait [--start] [--stop]
omcl heartbeat install
omcl updateRunning omcl setup generates these files (committed to git):
CLAUDE.md → Claude Code context (identity + memory + skill routing table)
AGENTS.md → Cross-platform brain (Codex, OpenCode, Antigravity)
GEMINI.md → Gemini CLI + Antigravity overrides
And creates .omcl/ (git-ignored) for runtime state:
.omcl/identity/ → SOUL.md, USER.md, PROJECT.md (your profile + project context)
.omcl/memory/ → MEMORY.md + sessions.db (persistent cross-session state)
.omcl/specs/ → feature specs (output of feature-design skill)
.omcl/plans/ → implementation plans (output of implementation-plan skill)
.omcl/tasks/ → cross-platform task queue (for parallel execution)
omcl update # updates npm package, refreshes all platform configsOr via Claude Code plugin:
/plugin marketplace update oh-my-colab
/omcl:setup
- Design before code — The agent clarifies what to build before touching a file
- Memory survives interruption — Compact, switch platforms, reboot: context is never lost
- Workflows, not suggestions — Skills are enforced processes, not optional hints
- One source, all platforms — Skills live once in our repo, symlinked everywhere
- Reliability over magic — 4-layer trigger system, hooks enforce what prompts only suggest
- Test-first always — RED before GREEN, no exceptions
Skills live in skills/. Each is a SKILL.md file you write once and it works on all platforms.
git clone https://github.com/iadr-dev/oh-my-colab
cd oh-my-colab
corepack enable pnpm # once per machine (ships with Node 16.13+)
pnpm install
pnpm run typecheck
pnpm run lint
omcl setup # configure oh-my-colab using itself
omcl skills validate # check your skill descriptionsRead skills/write-skill/SKILL.md for the complete guide on writing skills.
All PRs require:
- Skill description follows the directive format:
"ALWAYS invoke this skill when..." omcl skills validatepasses- Trigger test scenarios documented in PR description
- Tested on at least one platform
MIT — see LICENSE