AI DevOps Framework (aidevops)
An OpenCode plugin and AI DevOps framework that puts AI to work across code, infrastructure, business, and content projects — with security, token efficiency, and 24/7 autonomous delivery built in.
Strong evidenced security design: sudo-signed, session/content-digest-bound source-read approvals; an admin-only ai-approved label gate with automated reversal; secretlint and unpinned-dependency scanning; gopass-encrypted secrets. But the capability surface is enormous (credential access, infrastructure modification, pulse auto-merge, auto-update) and the README itself shifts responsibility to users ("Your responsibility"); external_effects relies on documented dry-run/rollback contracts that cannot be statically verified, hence the deduction.
Tests exist with deterministic isolation (tempdirs, network-free branches, characterisation pinning before refactor); CLI output follows structured conventions (OK:/ERROR:/SKIP:); CI ratchets AGENTS.md size and validates referenced headings. However dspy==3.1.0b1 is a beta pin and a known unpatched diskcache transitive dependency is documented as unresolved, so dependency_availability is not fully earned.
Audience is clear (developers/business users on OpenCode), platform constraints explicit (darwin/linux, node>=20), triggers are precise slash commands and agent pickers. But claims of coverage across "every business domain" outpace the evidenced paths, and the 100x aim is self-declared unsubstantiated, so capability_boundaries is thin.
Excellent information architecture: AI-CONTEXT block, purpose.md, DESIGN.md, exact auditable inventory counts, SPDX headers, LICENSE, SECURITY.md, semantic version 3.32.335 with release badges. Install notes give npm/Homebrew paths but under-document prerequisites for such a complex stack (OpenCode plugin + CLI + model routing); maintenance responsibility rests on a single author plus an email disclosure channel.
Output usability evidenced by the report rendering pipeline with verified/partial/inferred/missing evidence badges and structured CLI output; marginal value claims (token efficiency, tiered model routing, compaction control) are concrete and operable; cost/benefit is honestly qualified ("savings are not guaranteed"). Yet nothing is execution-verified and the learning surface (2,000+ scripts, 100+ commands) is heavy, warranting conservative deductions.
Claims bind to concrete file paths and line numbers; inventory counts are auditable via a shipped script; the README explicitly separates measured results from ambition ("not a guaranteed or measured result") and self-limits the harness evaluation ("not a leaderboard-superiority claim") — good fact/inference separation. External badge data and sealed-replay results cannot be corroborated from the repository alone, and referenced document contents cannot be fully confirmed statically.
- The framework grants AI broad access to credentials, infrastructure, and external APIs; pilot in isolation and review pulse/auto-merge and auto-update defaults before enabling.
- The dspy dependency is a beta pin with a known unpatched transitive dependency (diskcache); assess supply-chain risk before production use.
- This is a static review with no execution; README performance claims (100x, model-routing savings) are unverified by this assessment.
- Source-access approvals rely on sudo and 12-hour expiry; verify the helper's logic matches your host hardening policy before deployment.
- macOS/Linux only; maintenance depends heavily on a single author — enterprises should plan a fork contingency.
What does this agent do, and when should you use it?
aidevops (marcusquinn/aidevops, MIT licensed, npm package 'aidevops') is an OpenCode plugin and AI DevOps framework founded by Marcus Quinn on 9 November 2025. It ships 14 main agents (Build+, Automate, Product, SEO, Marketing-Sales, etc.), 2,200+ individually addressable Markdown sub-agent modules, 2,000+ helper scripts, and 100+ slash commands. The stack is Bash scripts and TypeScript (Bun), with 20 MCP servers and 30+ API service integrations. It emphasizes deliberate token spend via tiered model routing, secrets kept out of chat via gopass encryption, parallel work in isolated git worktrees, and a Pulse supervisor loop (running every 2 minutes via launchd) that dispatches workers, merges ready PRs, and recovers stuck work. Deployment is local-first: agents install to ~/.aidevops/agents/ and the aidevops CLI manages project registration, updates, and security scanning. The README explicitly frames its "100x more capable" ambition as unverified, not guaranteed.
Concretely, aidevops: (1) initializes features in any git repo with aidevops init (planning, git-workflow, code-quality, security, beads, etc.), scaffolding .aidevops., .agents/AGENTS.md, TODO.md, PLANS.md, and optional .beads/ task graph; (2) manages credentials via aidevops secret (gopass encrypted) and runs aidevops security for posture checks, plaintext secret detection, and supply chain scans including a .pth auditor; (3) manages OAuth account pools for OpenAI, Anthropic, Cursor, and Google via aidevops model-accounts-pool with rotation, cooldowns, and live validity checks; (4) runs /pulse for autonomous supervision and /mission to split goals into milestones with budget tracking and automatic advancement; (5) verifies risky operations with a second model from a different provider, and offers /cross-review to diff multiple models against the same prompt; (6) isolates agent work in git worktrees with SQLite mailbox messaging and Beads task dependency graphs; (7) renders evidence bundles into HTML/PDF reports with verified/partial/inferred/missing evidence badges for SEO scorecards, delivery reviews, and client handoffs.
- A solo developer who wants AI to build, review, and release features across multiple repos in parallel while protecting main branches and automating PR merges
- A one-person operator who needs AI to handle hosting, DNS, monitoring, deployment, SEO, and email deliverability checks for their sites
- A team that wants budget tracking and cross-provider verification to control token costs and de-risk destructive AI operations like production deploys
- A content/marketing professional who wants recurring AI-generated SEO scorecards and campaign reports exported as branded PDFs for clients
- Existing ChatGPT or Claude Pro/Max subscribers who want to reuse subscription quotas through OAuth account pools instead of buying API keys
- A developer who wants to codify repetitive browser workflows, report rendering, or mobile simulator testing into reusable slash commands
What are this agent's strengths and limitations?
- Scaled by auditable counts: an exact inventory of 14 main agents, 2,236 sub agents, 2,017 helper scripts, and 106 slash commands, with a public audit script (readme-helper.sh counts --inventory)
- Deep security engineering: gopass-encrypted secrets, 12-hour exact-path session-bound source-read approvals, supply chain dependency pinning (== versions), a .pth attack auditor, and cross-provider multi-model review of destructive operations
- Cost and context engineering: tiered model routing (Luna for bounded work, Terra for implementation, Astra for consequential reasoning), a 300K context-window guard for GPT-5.6, and per-session token-use reports with API-equivalent cost comparison
- Complete execution loop: worktree isolation, 2-minute Pulse supervision, mission budget tracking, SQLite observability with causal worker lineage
- Strong OpenCode dependency: all features are designed and tested for OpenCode first; other editors only get pointer files (.cursorrules etc.) and are not installed or configured
- The recommended model stack is OpenAI GPT-5.6 Luna/Terra and GPT-6 Astra — the authenticity of these model names should be independently verified; Claude models are only the supported fallback
- Wide local footprint: requires shell, gopass, gh CLI, git worktrees, optional launchd timers; setup proactively installs recommended tools (Tabby, Zed, Oh My Zsh, RTK), which is invasive for existing environments
- The README itself states the "100x more capable" ambition is not a measured result, and its local harness evaluation is a pilot rather than leaderboard-grade evidence
- The Claude Code integration adds an alias that runs claude --dangerously-skip-permissions, a risk users must consciously accept or disable
How do you install or deploy this agent?
Recommended: npm install -g aidevops && aidevops update (npm suppresses postinstall output; update deploys agents to ~/.aidevops/agents/). Alternatives: bun install -g aidevops && aidevops update; brew install marcusquinn/tap/aidevops && aidevops update; bash <(curl -fsSL https://aidevops.sh/install); or git clone https://github.com/marcusquinn/aidevops.git ~/Git/aidevops && ~/Git/aidevops/setup.sh. The setup script installs the CLI, configures AI assistants, and installs recommended token-efficiency tooling (RTT/RTK) by default. OpenCode is the required runtime; authenticate with aidevops model-accounts-pool add openai or add anthropic (browser OAuth, no API key needed), then restart OpenCode.
How do you use this agent?
- After install, run aidevops status and aidevops doctor to verify health. 2. In a project, run aidevops init (optionally with feature flags like planning, git-workflow, or beads). 3. Launch OpenCode and run /onboarding to configure services interactively; in each repo run /setup-git afterward for per-repo platform secrets like SYNC_PAT. 4. Work via slash commands: /pulse for the supervisor loop, /mission for multi-day projects, /review and /cross-review for quality, /skills recommend "TASK" for capability discovery, /report-render for reports. 5. Maintain with aidevops update (also checks registered projects), aidevops auto-update, and periodic aidevops security. For auth errors like "Anthropic Key Missing", use aidevops model-accounts-pool status/check/rotate.
How does this agent compare with similar options?
The README positions aidevops as going beyond single-task AI harnesses: normal tools excel at one job, while aidevops covers the surrounding discipline — context, routing, safety, git hygiene, collaboration, verification, and memory. Its relationship to Claude Code is complementary rather than competitive: Claude models remain a fully supported fallback provider, and the framework even installs a permission-skipping alias for Claude Code when present. No other specific competitors are named in the source.