Oh My Claudecode

OMC's core bet is that the interesting work in a codebase is not one model answering one prompt — it is several models with different strengths working the same repo at once. It executes that literally. Claude decomposes the task; a gemini CLI process spawns in a tmux pane for design review and large-context file analysis; a codex CLI process spawns in another pane for architecture validation and security review. All three are real CLI processes, not API shims, so your existing authenticated CLIs and their quotas carry over. The 55 built-in MCP tools are where the day-to-day leverage sits: LSP integration for hover, go-to-definition, find references, and project-wide type checking; AST grep for structural search and replace by syntax rather than text; and a deliberately sandboxed Python REPL where imports, file I/O, and dynamic code execution are blocked, so it handles calculation and in-memory data analysis but will not run your scripts. Persistent state — notepad, project memory, session state — survives across conversations, which matters more than it sounds once you are running multi-hour autonomous passes. The execution modes are the part worth studying before you commit. Autopilot is full autonomy from idea to tested code. Ralph is a self-referential loop that keeps going until an architect agent verifies completion. Execute is the maximum-parallelism path across worker lanes. Graph adds remote approval gates with workspace checkpoints and rollback — the mode to reach for when a staged change needs sign-off before it lands. Ultragoal keeps a durable ledger under .omc/ultragoal with fail-closed checkpointing, so a long objective survives interruptions rather than silently dying. Deep Interview and Planning invert the usual flow and interrogate a vague request before any code exists. Team coordinates N agents on a shared task list with real-time messaging and replaced the older swarm/ultrapilot models; note that ultrawork is retired and the plan keyword activation was removed — team and the explicit commands are the current interface, and the docs still carry v5.3.0 next to the homepage's v5.7.3. The honest limits: every worker lane needs its underlying CLI installed, authenticated, and inside its context limits, so you are carrying three sets of credentials and three sets of rate limits. Parallel panes need tmux. Native Windows support arrived in the v4.14.x line and is largely hotfix work on hook manifests, cache occupancy, and shell path handling, so Windows users should expect rougher edges. And the version numbering across pages is inconsistent, which is a small signal about the documentation surface you will be living in. Where it fits: engineers doing large refactors, repo-wide type-error sweeps, or multi-step projects that genuinely benefit from delegation. Where it does not: anyone who wants a GUI, a no-code builder, or a single model answering a single prompt.