OmO (omo command): turn tasks into a parallel agent graph with mass ulw

OmO packs every capability into a single omo command. The core usage is to type mass ulw, which turns the whole job into a graph of agents advancing in parallel; the official description of its outputs includes research across ten thousand sources, a deck people actually get, backends, frontends, and code.

Who it's for

• People who need to hand a large task to agents as a whole: official examples include deep research across thousands of sources, a full machine-learning pipeline, a migration nobody wants to do, and backend, frontend, and code development.
• Users subscribed to one of Claude, ChatGPT, Kimi, or GLM: the official text says /login covers these four subscriptions; whether other subscriptions or API-based access are supported is not stated.
• People used to working in a terminal who can accept a task being split into multiple agents running in parallel.
• People who want separate global and per-project configuration: settings live in ~/.omo/omo.jsonc, and a project can add its own .omo/omo.jsonc; the closest file wins.

Cases that do not fit, or that need confirmation first:

• Users who assume the npm package name is omo should beware: the official text states clearly that the omo package on npm is someone else's unrelated package, and OmO's package name is omo-ai.
• Users who must confirm in advance whether omo reads or writes local files, or executes local commands: the official documentation does not state these behaviors, so no conclusion can be given.
• The officially supported range of input and output languages is not stated; do not assume tasks in any language will be supported.

First attempt in 3 minutes

Recommended main path: install the official package omo-ai globally with bun; once installed, the omo command is ready to use.

bun add -g omo-ai

1. Prerequisites: bun is installed on the machine (without bun, npm can be used instead, see the next step), and you have a subscription account with one of Claude, ChatGPT, Kimi, or GLM.
2. Run the command above to install; the equivalent for npm users is npm i -g omo-ai, which is likewise the method given in the official text. One more reminder: the package name is omo-ai, and installing it yields the same omo command either way.
3. Sign in: run /login; the official text says it covers Claude, ChatGPT, Kimi, and GLM subscriptions.
4. Start a low-risk task: type mass ulw followed by a research-style question, for example asking it to gather public material on a topic across sources. Per the official description, this keyword turns the whole job into a graph of agents: each part runs on the model that fits it, executes in parallel, and is checked before anything is called done.
5. Watch for success signals: per the official description, the task is decomposed into multiple parallel agents, each part is checked one by one as it completes, and the results are finally aggregated into output; seeing the parts advance in parallel and receiving an aggregated result that has been checked matches the officially described way of finishing.

Note: the official project also provides an install script that places the native omo binary matching your OS and CPU into ~/.local/bin, verifies it against the release checksums, and falls back to GitHub Releases when the mirror is unreachable; this article does not include that script command, so defer to the repository README. For a first attempt, a research-type analysis task is recommended; the official docs do not say whether omo modifies local files, so confirm engineering tasks in a controlled environment first.

Use cases

Case 1: deep research across sources.

• Scenario: collecting and integrating a large body of material for a technology choice or an industry judgment, for example comparing public material and opinions on several candidate solutions.
• Kick-off: type mass ulw plus a research question in omo.
• Workflow: officially called deep research over thousands of sources — the task is split into a graph of agents that digs across thousands of sources, with the parts running in parallel and being checked.
• Expected result: a research output aggregated after part-by-part checks, usable as a working draft for later decisions; whether the output lands in the terminal or in files is not stated officially, so go by the actual run.

Case 2: turning work into a deck, or taking on engineering tasks.

• Scenario: you need a presentation deck people actually understand, or a full machine-learning pipeline, a migration nobody wants to do, or frontend and backend code.
• Kick-off: started the same way, with mass ulw plus a task description.
• Workflow: the task is split into parallel agents, each part uses the model that fits it, and is checked when done.
• Result and checkpoints: the official text says each part is checked before being marked done; whether it asks for confirmation before generating or modifying code, and where changes are written, is not stated. Validate on a small project or a copy first, then use it for real engineering work.

Usage boundaries

• File and command access: the official documentation does not state whether omo reads or writes local files or executes local commands, nor whether there are confirmation prompts; per the official description, research-type tasks access many external sources.
• Side effects and checkpoints: mass ulw splits a task into multiple parallel agents; the detailed rules of the check-before-done step are not public, and no safety guarantee can be inferred from it.
• Runtime and clients: OmO runs on senpi, a fork of pi; sign-in depends on /login, and the official list covers Claude, ChatGPT, Kimi, and GLM subscriptions, with no statement about other channels. Configuration lives in ~/.omo/omo.jsonc, and a project can add its own .omo/omo.jsonc; the closest file wins.
• Official explicit restriction: the npm package named omo is unrelated to this project; you must use omo-ai.
• Key information not yet confirmed: the list of supported operating systems, the full set of configuration options, where output is saved, the range of language support, and usage and billing rules — all must be verified against the official repository.