oh-my-agent
A portable multi-agent harness that keeps project skills, workflows, and engineering roles consistent across AI coding runtimes.
Per-dimension scores and reasoning
Evidence shows: SECURITY.md explicitly warns that oma-config.yaml is code-equivalent and can execute arbitrary binaries, advising review; backup directories (migration 008) and rollback mechanisms exist; but no user confirmation mechanism, data flow transparency, or sensitive data handling details are provided. Deductions: lack of user confirmation, data flow transparency, and sensitive data handling implementation evidence.
Evidence shows: README and package.json are consistent, version numbers match; but dependency availability (e.g., Node 26+, bun, uv, serena) is unverified, and failure messages are not provided. Deductions: dependency availability and failure messages lack evidence.
Evidence shows: README describes multiple scenarios (multi-IDE, presets, languages), capability boundaries are clear (each agent's role), trigger precision is measured (0% missed-fire), and environment fit is broad (multi-IDE, multi-OS). Deductions: no major deductions, but trigger precision data is not independently verified.
Evidence shows: README is well-structured, install instructions are detailed, naming is stable (oma- prefix), examples and FAQ exist (docs links), MIT license is clear; but known limitations are not explicitly listed, version changelog is not provided, and maintenance responsibility is unclear. Deductions: known limitations, version changelog, and maintenance responsibility are insufficient.
Evidence shows: output usability (generated code, docs) is described, marginal value (multi-agent collaboration, token savings) is explained; but cost-benefit (e.g., 75% token savings) lacks specific data. Deductions: cost-benefit lacks data support.
Evidence shows: README cites arXiv papers, but no specific verification data is provided; cross-source corroboration is insufficient; facts and inferences are not clearly separated. Deductions: claim traceability, cross-source corroboration, and fact-inference separation are insufficient.
- oma-config.yaml can execute arbitrary binaries; review before running.
- Install scripts auto-install bun, uv, and serena, potentially introducing unverified dependencies.
- No user confirmation mechanism is provided; agents may execute actions automatically.
What does this agent do, and when should you use it?
oh-my-agent is a project-scoped multi-agent harness that treats `.agents/` as the single source of truth for skills, workflows, and rules. It ships role-oriented agents for architecture, backend, frontend, databases, QA, security, infrastructure, project management, and other specialist work. Its slash-command workflows include `/plan`, `/work`, `/orchestrate`, `/review`, and `/debug`, while keyword detection can activate workflows from ordinary chat. `oma emit` projects the shared source into Agent Skills-conformant folders, `.claude-plugin/marketplace.json`, and `AGENTS.md`, alongside runtime-native or adapter layouts. The documented quality model includes `oma verify <agent>`, session quota caps, and the iterative `ralph` workflow, making it a fit for teams that want reusable engineering guidance to travel with a repository across supported coding tools.
oh-my-agent keeps skills, workflows, and rules in .agents/, then projects that source into detected runtime layouts. Users can invoke /deepinit, /brainstorm, /architecture, /plan, /work, /orchestrate, /ultrawork, /review, /debug, /docs, and /scm; the workflows coordinate roles such as oma-pm, oma-orchestrator, oma-backend, oma-frontend, oma-db, and oma-qa. oma emit produces Agent Skills-conformant skill folders, .claude-plugin/marketplace.json, and AGENTS.md; oma doctor --profile prints the resolved per-role model matrix; and oma verify <agent> runs shared and agent-type-specific checks. In .agents/oma-config.yaml, model_preset selects a model configuration and agents can override an individual role, for example backend: { model: openai/gpt-5.5, effort: high }.
- A full-stack team using Claude Code, Codex CLI, or Cursor that wants frontend, backend, database, and QA guidance maintained once in
.agents/. - An engineering lead planning a feature who needs
/architectureto assess module boundaries before using/planto create prioritized tasks. - A team implementing independent workstreams that wants
/orchestrateto run multiple agents while enforcing documented session quota limits. - Maintainers of an existing repository who want
/deepinitto map the codebase, then/review,/debug, and/docsfor quality, regressions, and documentation work. - A monorepo team using pnpm, Nx, Turbo, or Lerna that needs agents routed to the appropriate workspace.
What are this agent's strengths and limitations?
- Uses
.agents/as a single source of truth andoma emitto produce open-standard skill folders, a Claude plugin marketplace manifest, andAGENTS.md. - Provides named specialist roles across architecture, implementation, databases, mobile, QA, security, observability, infrastructure, and project management, with matching workflows.
- Documents concrete control mechanisms:
oma verify <agent>, shared and type-specific checks,session.quota_cap, independent JUDGE verification inralph, and hypothesis exploration after repeated retries. - Explicitly supports several coding runtimes, including Codex CLI and Claude Code, and allows role-level model selection in configuration.
- It requires Node.js 26+; manual setup also requires bun, uv, and serena, which can add environment setup work.
- The multi-runtime approach uses generated native or adapter layouts, and the repository advises choosing either APM or CLI distribution per project to avoid drift.
- The number of agents, workflows, presets, and quality controls creates upfront adoption choices around scope, configuration, and budget caps.
- The supplied material does not document direct integration paths for ChatGPT, the OpenAI API, or the Claude API.
How do you install or deploy this agent?
Node.js 26+ is required. On macOS or Linux, run:
curl -fsSL https://raw.githubusercontent.com/first-fluke/oh-my-agent/main/cli/install.sh | bashThe script auto-installs bun, uv, and serena when missing. On Windows PowerShell, run:
irm https://raw.githubusercontent.com/first-fluke/oh-my-agent/main/cli/install.ps1 | iexWith bun, uv, and serena already available, run:
bunx oh-my-agent@latestMicrosoft APM can install all skills with apm install first-fluke/oh-my-agent, but APM distributes skills only; it does not include workflows, rules, oma-config.yaml, keyword-detection hooks, or the oma agent:spawn CLI.
How do you use this agent?
After installation, maintain shared project content in .agents/. Start with /deepinit to map an existing codebase, then use /plan followed by /work or /orchestrate for implementation. Run oma verify <agent> for that agent type's deterministic checks, and run oma doctor --profile to inspect the resolved role-to-model matrix. Configure .agents/oma-config.yaml with a preset such as model_preset: mixed, then add per-role model and effort overrides under agents when needed.
How does this agent compare with similar options?
Compared with installing through Microsoft's APM, bunx oh-my-agent@latest includes workflows, rules, oma-config.yaml, keyword-detection hooks, and the oma agent:spawn CLI; APM distributes skills only. The project names Claude Code, Codex CLI, Cursor, OpenCode, and Qwen Code among runtimes that can share the .agents/ source rather than requiring one IDE.
FAQ
Am I locked into one model provider or IDE?
.agents/oma-config.yaml.Can I install it through both APM and the CLI?
oma agent:spawn.How are parallel-agent costs or usage constrained?
session.quota_cap in oma-config.yaml can cap per-session tokens, spawns, and per-vendor budgets; orchestrate blocks the next spawn when a cap is exceeded.Does it guarantee that review findings are fixed?
/review, /deepsec, oma-qa, and oma verify <agent> workflows and checks. The supplied material does not guarantee that every issue in a repository will be found or corrected automatically.