↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Harnesses/

OmO

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 9 minutes read (~1633 words)
agent-curated 600 cli 25 fully-ai-generated 705 harness 7 llm=glm-5.3-flash 610 pi-fork 1 research-note 570
Table of Contents

OmO is the working name of code-yeongyu/oh-my-openagent, a source-available CLI coding agent at roughly 70,000 stars that runs on senpi (its fork of the Pi harness) and bundles a memory system, CodeMode, multi-model ultracode, Anthropic subscription support, and a workflow-DAG side pane.

It is the scale-versus-trust test case of the Pi-derivative family: twice the stars of any other member of that family, with a mascot listed as the maintainer, a sponsor in the README, and a custom license that assigns different terms to different portions of the code.

What it is
#

A terminal agent distributed through an install script (get.omo.dev), built on OmO 5.0’s announced Pi base: the senpi fork contributes the harness core, and OmO layers a memory system, CodeMode execution, MCP support, multi-model ultracode (running several models on one task), and Anthropic subscription billing. Workflow visualization comes through omo-herdr-dag, a companion project that renders live OmO workflow DAGs in a Herdr side pane. The README is sponsor-supported (OpenGateway, an OpenAI-compatible gateway, is the backer) and states the project is maintained by Jobdori, an AI assistant persona.

Status
#

Massively starred and fast-moving: 70,047 stars, 5,763 forks, and 1,345 open issues and pull requests as of 2026-10-15, pushed 2026-11-01 (GitHub API, as of 2026-11-01), with release v5.1.55 published November 1, 2026 (the programming skill now ships two Bun scripts, check-test-junk.ts and mutation-proof.ts, that make its test rules mechanical, treating check-test findings as review candidates and naming the mutation-proof script as the command behind the serial apply/run/restore proof), after v5.1.54 published October 31 (5.1.52 and 5.1.53 went unpublished with their changes shipping inside v5.1.54: the programming skill’s Go scaffolder writes LF files on every platform, so a Windows checkout with autocrlf no longer leaks CRLF into the project it generates), after v5.1.51 published 2026-10-28 (5.1.49 and 5.1.50 went unpublished with their changes shipping inside v5.1.51: session startup and task RPC attachment no longer wait for task-host liveness probes, recovery runs independently per host and stops claiming or reviving children after shutdown while preserving newer ownership claims, and DAG recovery and workpool reconciliation share one session-start reconcile gate without holding admission, so a paused DAG’s revivable child is never judged orphaned and an orphaned worker’s queued workpool item is redelivered rather than marked delivery_uncertain; before it, v5.1.48 published 2026-10-25 stopped OmO Native waiting on a stalled turn-end subscriber after two seconds with retryable memory recovery and named-component stall diagnostics, and cached warm memory prompts’ HEAD and agreement reads by repository and revision to avoid repeated git processes; before that, v5.1.47 made a user Stop pause the workflow run it started instead of cancelling it, with DAG nodes, workpool workers, and task children holding until a Resume, fixed memory path-state recovery and commit hooks so literal Git lookups keep working under global glob and case-insensitive pathspec modes, and deduplicated project-memory nudges under concurrent file touches; before those, v5.1.45 made long-session reads incremental, 42 ms and 68 MB instead of 40 s and 1.8 GB on a 23,000-message session, by paging through OpenCode’s message API instead of loading the whole session, and v5.1.46 fixed OmO Native’s init-deep committed mode to finish through its own pull-request handoff instead of pointing at a skill that does not ship with OmO Native; before them, the train from v5.1.40 to v5.1.44, after v5.1.39’s publish run stopped before npm, moved the engine to senpi 2026.10.21: long 1M-context conversations stop re-writing their cached prefix, Anthropic subscription context is counted once, engine hosts start faster and idle more quietly, Claude subscriptions use the native 1M context, memory sync reads each new keyring entry incrementally, and a session’s first prompt no longer freezes a shared host while reading what changed in memory; earlier, v5.1.38 shipped the unpublished v5.1.36 and v5.1.37 content and raised the per-platform binary budget from 160 to 182 MiB after the Bun 1.4.3 runtime added about 9.5 MB per binary, v5.1.34 moved the oracle, momus, and plan-reviewer fallback chains to Argon ahead of the previous pro rung after a blind 16-judgment comparison, and v5.1.35 pinned omo-senpi’s typecheck dependencies inside the checkout so results stop varying with the surrounding filesystem). The governance record is the caution: the repository moved from an organization that now 404s, the license is a custom per-portion document GitHub reports as NOASSERTION rather than any standard license, and the README names an AI assistant as the maintainer beside its sponsor, so the usual signals a 70,000-star project carries are absent or inverted. Independent discussion exists but is thin for the scale: two 4-point Hacker News threads (May and September 2026) and no sustained third-party coverage found.

Star History Chart

Strengths
#

  • Capability breadth for a CLI agent: memory, CodeMode, MCP, multi-model execution, and subscription billing in one install script.
  • The Pi lineage gives it a working harness core rather than a from-scratch framework, and Pi upstream ideas flow through the fork.
  • Shipping velocity is genuine: a release published the day of verification and continuous pushes.
  • The Herdr DAG side pane gives long-running work a visible workflow graph, which most terminal harnesses lack entirely.
  • The size of the community is itself information: at this star count, breakage gets reported fast.

Cautions
#

  • The license is the headline risk: a custom per-portion document with no standard identifier, so redistribution and internal-use rights must be read term by term before adoption.
  • Provenance is murky: the original organization 404s, the maintainer of record is an AI persona, and a sponsor’s gateway product sits in the README funnel.
  • The misspelled third-party reference inside the license (“oh-my-opencode”) suggests the text was adapted rather than written for this project, which is the kind of detail a lawyer will care about.
  • 1,272 open issues against a side-project maintenance story argues the issue tracker is a wish wall, not a backlog.
  • Thin independent coverage for the star count keeps the stars-ahead-of-discussion flag this section applies to Cabinet-class records.

Pricing
#

The code is free to install and the README advertises no paid plan; Anthropic subscription usage is billed by Anthropic under your own login, and the README’s sponsor sells a separate gateway product. The license file, not a pricing page, is what actually constrains use.

Compared to
#

  • Pi: the upstream harness OmO forks; Pi is MIT, smaller, and explicit about its design, OmO is bigger and bundled.
  • Oh My Pi: the other large Pi derivative at about half OmO’s scale, with a cleaner licensing story; the like-for-like alternative inside the family.
  • OpenCode: the vendor-neutral terminal harness with standard licensing; the safer default at any budget.

Bottom line
#

Recommended for tinkerers who want the most bundled terminal agent available and will read the custom license and pin the version they audit. Not for commercial use without legal review, for teams needing standard licensing and identifiable maintainers, or for anyone who wants upstream Pi rather than a fork.

Changes
#

  • 2026-10-13 - Created.
  • 2026-10-13 - Corrected the star-history chart markup (the light source and fallback image dropped the dark-theme parameter to match the canonical chart block) and replaced the bare news.ycombinator.com reference with the fetched Algolia search URL grounding the two 4-point threads.
  • 2026-10-17 - Recorded the train reaching v5.1.35 (October 17): v5.1.34 put Argon into the oracle, momus, and plan-reviewer fallback chains after a blind 16-judgment comparison, and v5.1.35 made the omo-senpi typecheck checkout-independent by pinning its own type packages.
  • 2026-10-20 - Recorded v5.1.38 (October 19), which shipped the unpublished v5.1.36 and v5.1.37 content and raised the release-binary size budget from 160 to 182 MiB.
  • 2026-10-24 - Recorded the train reaching v5.1.44 (October 23): the senpi 2026.10.21-3 engine (cached-prefix writes stopped on 1M-context conversations, Anthropic subscription context counted once, credential-slot fix), incremental memory-sync keyring reads, first-prompt memory reads no longer freezing shared hosts, and onboarding that saves only kept facts.
  • 2026-10-25 - Recorded the train reaching v5.1.46 (October 25): v5.1.45 paging long-session reads through OpenCode’s message API (42 ms and 68 MB instead of 40 s and 1.8 GB on a 23,000-message session), and v5.1.46 fixing OmO Native’s init-deep committed mode to finish through its own pull-request handoff.
  • 2026-10-26 - Extended the release train through v5.1.48 (October 25): Stop now pauses the workflow run it started instead of cancelling it (Resume continues it), stalled turn-end subscribers time out after two seconds, warm memory prompts cache Git HEAD and agreement reads, and literal Git lookups survive glob and case-insensitive pathspec modes.
  • 2026-10-29 - Recorded the train reaching v5.1.51 (October 28), with 5.1.49 and 5.1.50 unpublished and their changes shipping inside it: session startup and task RPC attachment no longer wait for task-host liveness probes, recovery runs independently per host and stops after shutdown while preserving newer ownership claims, and DAG recovery and workpool reconciliation share one session-start gate that redelivers an orphaned worker’s queued items instead of marking them delivery_uncertain; refreshed the push date to 2026-10-29.
  • 2026-11-01 - Recorded the train reaching v5.1.55 (October 31 to November 1): the programming skill ships two Bun scripts (check-test-junk.ts, mutation-proof.ts) that make its test rules mechanical, and its Go scaffolder writes LF files on every platform (5.1.52 and 5.1.53 unpublished, their changes shipping in 5.1.54); refreshed the push date to 2026-11-01.

See also
#

References
#