The ProCoders AI-Native Stack: Tools by Product Stage
The ProCoders AI-native stack is the practical, opinionated toolset we actually install and run on projects, organized by where each tool fits in the product cycle: coding, browser automation, product, design, SEO/GEO, anti-AI-writing, and voice-first input. Each tool has a type (Skill, MCP, Plugin, or App), a clear job, and a priority.
What this stack is (and isn't)
This is the practical stack ProCoders actually recommends — the tools we install and run on real projects — laid out by where they fit in the product cycle. It's the concrete implementation of our 17 AI-native skills and the engine behind the L1 to L7 levels: the skills define what an AI-native developer does; this stack is how we do it.
It is deliberately different from a "best tools" listicle. If you want the broad, ranked, by-level survey of coding agents, see our coding tools by level deep dive. What follows is narrower and more honest: an internal recommendation with types, jobs, and priorities, not a market roundup. Every tool here is something we'd put on a project tomorrow.
The priority labels mean what they say. Must-have is non-negotiable for the level it applies to; recommended is what we'd add next; optional is situational. Types are Skill / MCP / Plugin / App (a finished product). Star counts, where given, are approximate indicators of maturity and adoption, not precise figures.
Coding
The core loop: agentic framework, fresh docs, structural search.
| Tool | Type | What it gives | Priority |
|---|---|---|---|
| Superpowers | Plugin / Skills | Base agentic framework: TDD, plans, debugging, review, verification. | must-have |
| Superpowers-V (by ProCoders) | Plugin | Our fork/overlay: phased workflow, code archaeology, parallel dispatch, Context7 validation. | must-have |
| ECC | Plugin / Orchestrator | Harness-native operator system across Claude Code, Codex and Cursor: the orch-* orchestrator family, subagents, skills, hooks and session memory. Install v2 and wire it to the Hermes operator layer. | recommended |
| Context7 | MCP | Up-to-date library docs straight into context — kills stale, deprecated solutions. ~54k★ (approx). | must-have |
| ast-grep + skill | CLI + Skill | Structural code search over the AST, not text: "find every async with no error handling." | recommended |
| skill-creator | Skill | Build your own skills — turn repeated work into a reusable harness (skill #7). | recommended |
Browser automation
Let the agent drive a real browser for navigation, checks, and E2E.
| Tool | Type | What it gives | Priority |
|---|---|---|---|
| axi | App / Browser | Fast browser built for agents, saves tokens. | recommended |
| dev-browser | Tool | Browser environment for agent-driven development. | optional |
| Playwright MCP | MCP | Agent-driven browser control: navigation, clicks, UI checks, E2E. | recommended |
Product
The product/marketing layer that powers cross-functional L7 work.
| Tool | Type | What it gives | Priority |
|---|---|---|---|
| marketingskills | Skills | CRO, copywriting, SEO, analytics, growth — the product-marketing layer for L7. | recommended |
| premortem-skill (by ProCoders) | Skill | Stress-test a plan or decision: "assume we failed" → surface failure causes up front (skill #5). | recommended |
Design
From PRD to layout, plus agent access to design files.
| Tool | Type | What it gives | Priority |
|---|---|---|---|
| Claude.ai/design | App | Highly effective design prep: PRD → full layout; restyle and improve existing apps. | recommended (strongly) |
| Figma MCP | MCP | Agent access to Figma: read/assemble mockups, design tokens. | recommended |
| OpenDesign | Tool | Open design toolkit for agentic UI/UX. | optional |
SEO / GEO
Make the agent write for both classic search and AI answer engines.
| Tool | Type | What it gives | Priority |
|---|---|---|---|
| claude-seo | Skill | SEO optimization tuned for Claude — sharper and more complete. | must-have |
| geo-seo-claude | Skill | GEO/SEO as a softer alternative to claude-seo. | optional |
| DataForSEO MCP | MCP | Real SEO data (rankings, keywords, SERPs) into the agent's context. | recommended |
Avoid AI Writing
Strip the AI tells out of generated text before it ships.
| Tool | Type | What it gives | Priority |
|---|---|---|---|
| humanizer | Skill | Removes AI-writing tells (33 patterns). One of the most popular — ~24k★ (approx). | recommended |
| avoid-ai-writing | Skill | A solid alternative — genuinely works, if less hyped. | optional |
Input / voice-first (dictation)
Voice instead of keyboard. Typing is a hidden time sink most people just ignore. The ideal pattern: the developer dictates the ticket, request, or stream of thought, and the AI orders and structures it. It's faster, and it removes the "too lazy to type it out long-form" barrier. It directly backs the L7 idea of director, not coder: voice is the natural interface for orchestration.
| Tool | Type | What it gives | Priority |
|---|---|---|---|
| superwhisper | App | Dictation with offline models, free for offline use. Personal pick. | recommended |
| handy | App | Free dictation. | recommended |
| Apple Dictation | App (built in) | macOS dictation out of the box, but basic in quality. | optional |
Voice-first is mandatory from L3
This isn't a nice-to-have. The KPI: ≥80% of long requests and tickets are dictated, not typed. And it becomes mandatory at Middle (L3) — by that stage dictation saves too much time to stay optional. Dictating a messy stream of thought and letting the agent structure it is faster than typing a clean one, and it's the most natural way to drive an agent fleet.
Gaps we're closing (candidates for the stack)
After scanning the leaders, these fill a real skill gap rather than padding the stack:
| Tool | Type | Status | Why |
|---|---|---|---|
| Supabase / Postgres MCP | MCP | taking it | Agent writes and runs SQL, reads the result — for DB-backed projects (skill #2). Works well in most cases. |
| Sentry / logs MCP | MCP | desirable | Access to prod/staging errors and logs — "from the tray to the fix" for systematic debugging and verification (skills #9, #15). |
| tdd-guard | Skill / Hook | recommended | Hard TDD enforcement: blocks code without a failing test (skill #9). Complements Superpowers. |
| — | not needed | PRs/issues/repo work is covered by the GitHub CLI (gh) — a separate MCP isn't required. |
We keep the candidate list short on purpose: we add an MCP or skill only for a real gap in a skill, not "just in case."
The wider landscape (leaders by adoption)
Separate from our opinionated stack above, here's the broader OSS landscape we draw from — the most starred, most mature tools that map onto our skill system. Stars are approximate and drift over time; treat them as a maturity signal, not a scoreboard. The selection principle: official first (Anthropic + our own procoders/*), then leaders by adoption, and always check compatibility with our stack before rolling something in.
Base frameworks (install first).
- obra/superpowers (~225k★ approx) — the main agentic-skills framework: TDD, planning, brainstorming, debugging, verification, code review, subagents, worktrees. In Anthropic's official marketplace. github
- anthropics/claude-code-skills (official) — Anthropic's official, vetted skills index (brainstorming, test-driven-development, systematic-debugging). The quality benchmark. github
Orchestration / subagents / worktrees.
- ruvnet/ruflo (ex claude-flow, ~59k★ approx) — meta-harness for agent swarms: swarm coordination, memory, autonomous workflows on top of Claude Code. The most adopted multi-agent OSS project of 2026. github
- wshobson/agents (~37k★ approx) — 192 specialized subagents across 84 plugins (researcher, reviewer, tester, security, and more). A multi-harness marketplace. github
- affaan-m/ecc (v2, MIT) — harness-native operator system: an orch-* orchestrator family, dozens of subagents, reusable skills, hooks and persistent memory across Claude Code, Codex, Opencode and Cursor. We recommend the v2 line, wired to its Hermes operator layer. github
Verification / review.
- nizos/tdd-guard (popular) — automatic TDD enforcement: blocks writing code without a failing test. Works with Claude Code, Codex, Copilot, any language/runner. github
- Anthropic Code Review (official plugin) —
/code-reviewon a PR branch: analyzes changes, rates issues, auto-comments on GitHub. claude.com/plugins/code-review
Planning / research.
- superpowers
brainstorming/writing-plans(~225k★ approx) — the structured pre-implementation cycle: research context, plan, risks. github - procoders/premortem-skill (ours) — plan stress-test: "assume we failed" → failure causes surfaced in advance. Claude/Codex/Cursor. github
- Deep-Research-skills / DAG deep research (OSS) — breaks a question into a dependency graph, runs parallel subagents, writes a sourced report. github
Artifacts / demo.
- Anthropic built-in
pptx/docx/xlsx/pdf(official) — generate finished artifacts (decks, docs, sheets, PDFs) straight from the agent. - procoders/claude-ui-recorder (ours) — branded UI video demo: record → narrate → render →
demo.mp4. github - Remotion skills (OSS) — programmatic video generation from code.
npx skills add remotion-dev/skills. remotion.dev - NotebookLM (Google) — narrated audio overview / interactive report from your materials.
Curated lists (for finding your own). hesreallyhim/awesome-claude-code (~37k★ approx), rohitg00/awesome-claude-code-toolkit (135 agents, 35 skills, 176+ plugins), travisvn/awesome-claude-skills, ComposioHQ/awesome-claude-skills.
Where to start
You don't install all of this at once — you add tools as you climb. The minimum to be agentic is the Coding row plus voice-first; everything else slots in as your level demands it. Find your level in 3 minutes and you'll know exactly which rows of this stack you should be running next.
Related: the anatomy of a compound system shows how ProCoders' Compound V assembles these ideas into a real, reviewed multi-agent workflow.
FAQ
- What's the minimum stack to start?
- The Coding row plus voice-first input. Concretely: Superpowers (the agentic framework) and Context7 (fresh library docs into context), driven by dictation (superwhisper or handy) instead of typing. That gets you a real agentic loop — plan, build, verify — with current docs and a fast input method. Everything else (browser automation, design, SEO, the wider landscape) layers on as your level and project demand it. Don't install all of it on day one; add a tool only when a skill needs it.
- Skill vs MCP vs plugin vs app — what's the difference?
- A Skill is a packaged capability that teaches the agent a workflow (e.g. claude-seo, humanizer, skill-creator). An MCP is a Model Context Protocol server that connects the agent to a live data source or tool (e.g. Context7 for docs, Figma MCP, DataForSEO MCP, Supabase/Postgres MCP). A Plugin bundles skills and configuration into an installable unit (e.g. Superpowers, Superpowers-V). An App is a finished standalone product you run alongside the agent (e.g. superwhisper, axi, Claude.ai/design). In this stack, each tool's Type column tells you which it is.
- Do I need all of these tools?
- No. This is an opinionated menu organized by product stage and priority, not a checklist. Must-have items (Superpowers, Context7, claude-seo) are the core; recommended and optional items slot in by need. You add tools as you climb the levels — a Level 3 developer needs the coding core and voice-first; browser automation, design, and multi-agent orchestration become relevant at Level 4 and up. Add an MCP or skill for a real gap in a skill, never 'just in case.'
- How is this different from your 'best AI coding tools' page?
- This page is the practical stack ProCoders actually installs and runs, organized by where each tool fits in the product cycle, with types and priorities. The coding-tools-by-level deep dive at /agentic-coding-tools is a broader, ranked survey of coding agents matched to maturity levels. Think of this as our internal recommendation and that as the market roundup. They complement each other.
- Why is dictation mandatory from Level 3?
- Typing is a hidden time sink, and by Middle (L3) the savings are too large to leave optional. The pattern: dictate the ticket, request, or stream of thought and let the agent structure it — faster than typing it clean, and it removes the 'too lazy to write it long-form' barrier. The KPI is ≥80% of long requests and tickets dictated rather than typed. It's also the natural interface for the L7 'director, not coder' way of working: voice is how you orchestrate an agent fleet. Tools: superwhisper or handy.