Oh My Antigravity (OmA)

A multi-agent engineering workflow pack for Antigravity CLI that turns planning, read-only review, sealed scope, real verification evidence, and small reviews into a gated team process.

Stars
★ 221
Last updated
2d ago
License
NOASSERTION
Primary language
JavaScript

At a glance

How it runs
Agent plugin / skill
Works with
Platform-specific
Cost
Free software; you pay for model usage
Setup effort
Low · running in minutes
You'll need
Antigravity CLINode.js 20+Python 3.9+ (only when the sealed scope contains Python files)Shell / CLINetwork accessLocal filesystem
Typical use
Developers doing real engineering changes on Antigravity CLI who want design-first, sealed-scope execution with auditable verification evidence rather than a one-shot auto-patch
Not a fit if
  • Developers not using Antigravity CLI or the Google Gemini ecosystem
  • Users who want agents with free write/shell access — shipped workers exclude write, shell, MCP, and delegation tools
  • Teams wanting a fully autonomous single-turn coder rather than a gated, multi-turn workflow
Maintenance
Commits in the last 30 dayslast commit 2026-11-20
Source review
86/100 · Good

What does this agent do, and when should you use it?

oh-my-antigravity (OmA) is a native Antigravity CLI plugin that organizes engineering work as a context-engineering-powered multi-agent workflow. It separates design from execution: /oma-team-plan records a Scope Brief and seals the file scope and checks, a read-only oma-design-auditor independently audits the design, and /oma-team-exec applies the sealed plan only in the next user turn. The plugin ships 64 short skills (including 53 /oma-<name> command entrypoints) and 16 capability-limited agents: oma-doc-reader reads specifications, oma-planner and oma-architect assist design, oma-executor proposes patches that only the owning main conversation applies, oma-reviewer examines one bounded section, and oma-verifier checks acceptance evidence. Every shipped worker excludes write, shell, MCP, and delegation tools. Authoritative workflow state lives in .omg/state/guard.json, execution logs in .omg/evidence/, and reviewer receipts are captured by a native hook from another conversation using the real sender identity, so the parent cannot author PASS. The model catalog follows the Gemini model guide with Neon, Carbon, and Hydrogen lanes and blocks unapproved model identities before invocation.

Installed via agy plugin install as a native Antigravity CLI plugin, OmA exposes a set of hook-enforced commands: /oma-read pages source snapshots fully and returns a documentation reader's summary; /oma-probe performs actual bounded CLI help introspection; /oma-brief records objective, constraints, non-goals, decision owners, and source conflicts; /oma-design-check audits the exact sealed plan before the design turn can close; /oma-team-exec applies the sealed scope in the next turn; /oma-team-verify collects section reviews and runs real checks; /oma-team-fix runs a Stop-selected batch of up to three repair units. Validator logs are surfaced as integrity-checked windows with a bounded UTF-8 byte cursor; validator scripts and check inputs carry recorded hashes, and dependency changes block start, verification, or acceptance of stale evidence, with drift reported in Session Trail. Missing evidence yields blocked or unknown — success is never a required answer. Session Trail (/oma-checkpoint, /oma-resume) keeps immutable JSON/Markdown checkpoints of intent and next action and detects source/plan drift on resume. Passed files stay locked; reopening a lock or changing sealed scope requires an explicit operator action outside the guarded agent.

  1. Developers doing real engineering changes on Antigravity CLI who want design-first, sealed-scope execution with auditable verification evidence rather than a one-shot auto-patch
  2. Teams worried about fabricated "checks passed": every review receipt is captured by a hook from a child conversation with its real identity, and missing evidence reports unknown
  3. Editors of large files or critical paths: care profiles (routine/elevated/critical) bound sections to 120/90/60 lines and repair batches to 3/2/1 units
  4. Recovery after a long session interruption: /oma-checkpoint and /oma-resume resume from immutable checkpoints while detecting source drift
  5. Existing v1.0 or Gemini CLI extension users migrating: the docs document native commands (/oma-team-exec replacing /oma:team-exec) and the legacy protocol path

How do you install or deploy this agent?

Prerequisites: a current Antigravity CLI and Node.js 20+; Python 3.9+ is additionally needed when the sealed scope contains Python files.

agy plugin install https://github.com/Joonghyun-Lee-Frieren/oh-my-antigravity
agy plugin list

Inside the CLI, inspect /plugin list and /hooks to confirm the plugin and hooks are active.

How do you use this agent?

Start a design turn in the Antigravity CLI describing the change and its acceptance criteria:

/oma-team-plan Describe the change and its acceptance criteria

Use /oma-read for full source coverage and /oma-probe for CLI help, record a Scope Brief, seal the file scope and checks, and obtain an independent Design Check via /oma-design-check. After it passes, end the design turn, then in your next turn:

/oma-team-exec Implement the sealed plan

Run /oma-team-verify to collect remaining section reviews; on failure, Stop schedules a /oma-team-fix repair batch. Use /oma-status and /oma-doctor for phase, evidence gaps, and locks, and /oma-model for runtime model selection. Models can also be chosen with agy --model "Orchestrator (Neon)" or /model.

What are this agent's strengths and limitations?

Pros
  • Independent review receipts are captured by a hook using the real child-conversation identity; the parent cannot author PASS or steer the review
  • Every shipped worker physically excludes write, shell, MCP, and delegation tools; patches are applied only by the owning main conversation
  • Validator scripts and declared check inputs carry recorded hashes; dependency changes block start, verification, or acceptance of stale evidence, and Session Trail reports drift
  • Honest failure semantics: missing evidence or permissions yields blocked/unknown, and success is never a required answer
Limitations
  • Deeply tied to Antigravity CLI and the Gemini model catalog; unapproved models are blocked before invocation and no other provider is supported
  • The harness constrains agent tools but is not an OS sandbox; validators are trusted project code, and semantic fabrication still depends on behavioral checks and reviewers
  • Real learning curve: multi-turn operation (plan turn + execution turn), sealing, review handoffs, and repair batches differ from single-shot autonomous coding
  • Passed files stay locked across subsequent workflows; reopening a lock, changing sealed scope, or recovering blocked state requires an explicit operator action outside the guarded agent
  • Migrating from v1.0 requires resealing plans for the new input protection; the old TOML protocol path is legacy-only

How does this agent compare with similar options?

The project grew from Jeongkyu Shin's observation that the engineering harness around a model matters alongside the model itself. It positions itself as a gated multi-agent workflow pack inside the Gemini/Antigravity CLI ecosystem rather than a portable plugin for other agent hosts; the source provides no comparison data against specific competing products.

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
Oh My Antigravity (OmA) This agent 86 · Good Agent plugin / skillFree + model costs ★ 221 2d ago JavaScript —
Crewplane 76 · Some gaps CLIFree + model costs ★ 56 5d ago Python Codex · Claude Code
multi 56 · Major gaps — ★ 25 4d ago Shell Codex · Claude Code · Claude API
Finding-Unknowns Skills 67 · Some gaps Agent plugin / skillFree ★ 374 1mo ago Python Claude Code · Claude.ai

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Good
86/ 100 5-point scale 4.3 / 5
Trust 27/29
Reliability 11/14
Adaptability 16/18
Convention 16/18
Effectiveness 10/13
Verifiability 6/8
Why each dimension lost points
Trust27 / 29 · 4.7/5

Evidence shows clear privilege layering: all worker agents exclude write, shell, MCP, and delegation tools; only the owning main conversation writes code. Sealed plans, explicit design turns and user-confirmation gates, and lock persistence (reopening requires explicit operator action) are documented; rollback/recovery paths including evidence preservation and resume drift detection are written down. Dependency changes block start, validator hashes are recorded. Deductions: no dedicated sensitive-data handling section, only implicit bounds around log/evidence windows; dependency security relies on declared hashes and manifests with no lockfile or vulnerability-scanning evidence in the packet.

Reliability11 / 14 · 3.9/5

Typed honest failure exits (blocked/unknown, Stop-scheduled repair batches) are well described — failure messaging is a strength. Deductions: README and installation.md contradict each other on model recommendations (README recommends Neon/Carbon/Hydrogen and explicitly blocks Orchestrator (Argon); installation.md recommends Gemini 4 Argon and Gemini 3.8 Flash); behavior when runtime hook support is missing degrades to 'unknown', and dependency availability rests on documentation with no degradation-execution evidence.

Adaptability16 / 18 · 4.4/5

Audience and scenarios are clear (engineering teams, multi-agent workflows); 64 granular skills with explicit instruction to use only relevant procedures; capability boundaries are honestly stated ('the harness is not an OS sandbox', 'does not guarantee zero defects', 'cannot prove the model understood the source'); command trigger semantics are precise. Deductions: deeply bound to Antigravity CLI and its hook runtime — not portable — and parts of CLI model availability are self-declared 'unverified'.

Convention16 / 18 · 4.4/5

Good information architecture: README, multi-language docs, installation guide, guarded-workflow contract, session-continuity guide layered clearly; installation notes are concrete (agy plugin install, hook validation, removal of stale copied hooks); naming migration is explicit (/oma:name → /oma-name, legacy TOML protocols quarantined); known limitations are honest (unverified model availability, static validation ≠ runtime behavior); LICENSE is complete MIT with derived-work attribution to everything-claude-code; version 1.2.1 is consistent across badge, changelog, package.json, with CI checks. Deductions: concrete command outputs/JSON examples live in referenced guides not included in the sample, and no FAQ is visible; maintenance responsibility is a single personal repo with no governance or response commitment.

Effectiveness10 / 13 · 3.8/5

Output usability is well designed: compact status/context packs, bounded log windows, Session Trail checkpoints, review receipts. Deductions: marginal value depends on native Antigravity hook/send_message runtime support — when missing, core safeguards degrade to unknown; the scale (64 skills, 16 agents) plus mandatory two-turn sealed-plan flow is heavyweight for ordinary tasks, and no lightweight path for small work is documented.

Verifiability6 / 8 · 3.8/5

Claim/inference separation is strong: the docs explicitly mark 'unverified' (Flash-Lite CLI availability), 'unknown' (missing hooks), and implemented behavior separately, and state that safeguards do not prove the model understood the source. Deductions: key safety claims (hook capture of real sender identity, validator hashes blocking stale evidence) are not traceable to implementation files within the sample (hooks.json and plugin.json are not included); cross-source corroboration is limited — the CI workflow confirms test scripts exist, but test content is absent, so claims like coverage of actor spoofing and replay cannot be checked.

Risks and how to mitigate them
  • README and installation.md contradict each other on model recommendations (Argon is blocked in one and recommended in the other); verify with agy models before relying on either.
  • The safeguards depend on native Antigravity CLI hook runtime; if hooks are missing or disabled, the controls do not apply, and the project itself warns not to assume safety controls are active.
  • The authors state the harness is not an OS sandbox and validators are trusted project code; behavioral checks and reviewers still need human oversight for semantic fabrication.
  • Verified claims (hook sender-identity capture, validator hashes) could not be traced to implementation files in this static sample; no runtime validation was performed.
  • Publisher identity is unverified by the curated registry; license metadata is NOASSERTION while the in-repo license is MIT (with derived-work attribution) — treat the LICENSE file as authoritative for compliance.
Evidence confidence: Low Reviewed Nov 21, 2026 Reviewed revision db5901120ad5
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/oh-my-antigravity.svg)](https://followagents.com/en/agents/oh-my-antigravity)
<a href="https://followagents.com/en/agents/oh-my-antigravity"><img src="https://followagents.com/badge/oh-my-antigravity.svg" alt="FARS review"></a>

FAQ

Does it only work inside Antigravity CLI?
Yes. OmA is a native Antigravity CLI plugin installed via agy plugin install. The model catalog follows the Gemini model guide, and the plugin blocks unapproved or missing model identities before invocation. This check applies only while the plugin and hook are enabled and does not change the CLI model picker.
Does using it cost money?
The plugin itself is MIT-licensed, but it runs on Gemini-family models (Neon, Carbon, Hydrogen), so usage costs depend on your Antigravity/Gemini account eligibility. The README advises running agy models and checking account eligibility before selecting a model.
Does it guarantee defect-free code?
No. The README states explicitly that these safeguards do not guarantee zero defects or prove the model understood the delivered source; validators are trusted project code, and semantic fabrication must still be caught by behavioral checks and reviewers.
What happens when a check fails or evidence is missing?
The outcome is blocked or unknown, never a forced success. Stop schedules /oma-team-fix batches of up to three repair units, focusing on one group when repair oscillates. Recovering blocked or foreign-owned state requires an explicit operator action outside the guarded agent, preserving prior evidence.
Can I run multiple worktrees in one workflow?
One workspace and one writer are allowed per workflow; use separate workflows for isolated worktrees. Authoritative state is .omg/state/guard.json and logs live in .omg/evidence/.
Is Oh My Antigravity (OmA) open source?
The source code is public on GitHub (Joonghyun-Lee-Frieren/oh-my-antigravity), but the repository does not declare a recognised licence, so the rights to reuse or modify it are unclear. Ask the author or check the README before building on it.
Is Oh My Antigravity (OmA) safe to use?
FollowAgents' FARS review (Nov 21, 2026) scored Oh My Antigravity (OmA) 27/29 on trust and safety. None of the eight safety controls the review checks for was missing outright.
Is Oh My Antigravity (OmA) still maintained?
Yes. The latest commit on GitHub is from 2026-11-20, within the last 30 days.
Who makes Oh My Antigravity (OmA)?
Oh My Antigravity (OmA) is published on GitHub by the account Joonghyun-Lee-Frieren. FollowAgents has not verified which person or company is behind that account.
What is Oh My Antigravity (OmA)'s official website?
The project's GitHub repository lists https://joonghyun-lee-frieren.github.io/oh-my-antigravity/ as its website. That link is set by the repository owner; the source code itself is on GitHub (Joonghyun-Lee-Frieren/oh-my-antigravity).
Was this page helpful?
View on GitHub ↗ Install ↓

Host ecosystems

Related concepts