Agentic Stack

Six battle-tested plugins that turn Claude Code into a research & decision machine. One-command install · works in Claude Code, Cursor & Codex · free & MIT.

[](https://github.com/justsima/agentic-stack/stargazers) [](https://github.com/justsima/agentic-stack/network/members) [](https://hits.sh/github.com/justsima/agentic-stack/) [](traffic/REPORT.md)

[](https://github.com/justsima/agentic-stack/commits/main) [](https://github.com/justsima/agentic-stack/issues) [](https://github.com/justsima/agentic-stack) [](LICENSE) [](https://claude.com/claude-code) [](#-contributing) [](https://github.com/justsima/agentic-stack/releases)

[🚀 Quick start](#-quick-start) · [🧩 Plugins](#-the-plugins) · [🧠 How it works](#-how-it-works) · [📊 Analytics](#-analytics--traffic) · [🧰 Any tool](#-install-in-your-tool)

---

✨ Why agentic-stack?

Most "AI skills" are a single clever prompt. These are full multi-agent pipelines with the engineering that separates a demo from a tool:

• 🧪 Evidence in, judgment out. Research runs split evidence-gathering (parallel sub-agents) from judgment (a deterministic Python scoring engine for product decisions, an adversarial red-team for research). The conclusion is grounded — not vibes, not a hallucinated ranking.
• 🔁 They get smarter. ultradeep writes a transferable lesson back into its own program.md after every run. Your tooling compounds.
• 🗂️ They compound into a knowledge base. Reports file themselves into an optional Obsidian wiki, cross-linked and searchable next time.
• 🛡️ Privacy-first by design. Ships the engine, never your data. No secrets, no personal content — your wiki/memory start as empty scaffolds.
• ⚡ Zero-to-running in one command. An interactive installer wires the free MCP servers and asks what you want. Required pieces are free and key-less.

> Built by @justsima — the agentic research workflow behind real deep-research, buying decisions, and high-stakes calls, packaged so you can run it too.

How it compares

| | Single-prompt "skill" | Vanilla Claude Code | agentic-stack | | --- | :---: | :---: | :---: | | Multi-agent fan-out | ❌ | ⚠️ manual | ✅ | | Adversarial red-team verification | ❌ | ❌ | ✅ | | Deterministic scoring — no hallucinated rankings | ❌ | ❌ | ✅ | | Self-learning across runs | ❌ | ❌ | ✅ | | Files into a knowledge base | ❌ | ❌ | ✅ | | One-command install + MCP wiring | — | — | ✅ | | Privacy-first — ships the engine, not your data | — | — | ✅ |

---

🧩 The plugins

| | Plugin | What it gives you | Try it | |---|---|---|---| | 🔬 | ultradeep | Multi-agent deep research: 8–10 parallel explorers → STORM-style questioning → adversarial red-team → tiered search (Exa / SearXNG / WebSearch) → cited, wiki-filed report. Self-tunes via program.md. | /ultradeep | | 🛒 | market-scout | "Best X on the market" across Amazon/Best Buy/Walmart/Newegg. Parallel explorers + a deterministic decision-matrix engine + red-team → a ranked, segment-aware buying report. | /market-scout best | | 🏛️ | llm-council | A decision through 5 advisors, each a distinct reasoning method (inversion, decomposition, analogy, naive questioning, dependency graphing) → anonymized peer review → chairman synthesis. | /llm-council | | 🧠 | adhd | Divergent ideation — N isolated idea branches under different cognitive frames, scored, pruned, deepened. Beats one-shot brainstorming on open-ended problems. | /adhd | | 📝 | job-application-helper | Natural, tailored application answers from your own profile (a private, git-ignored profile.md you create from the template). | shares a JD → drafts answers | | 🛡️ | agentic-config | Safety hooks (block dangerous bash, block secret leaks), a session-context hook, a verify-before-stop hook + a recommended-settings fragment. No secrets. | auto-loads |

---

🎬 In action

You ▸ /market-scout best 5G router

agentic-stack ▸ dispatching explorers… (expert reviews · contrarian/reliability · live prices)
              ▸ scoring 7 candidates across 4 weight profiles (deterministic MCDA)
              ▸ red-team: PASS-with-revisions (caught 2 inflated ratings, 1 missing candidate)

🏆 No single winner — 4 picks by use case:
   • Whole-home   → Ubiquiti UDR-5G-Max ($499)
   • Mobile (AT&T)→ NETGEAR Nighthawk M7 Pro ($449, locked)
   • Travel/eSIM  → NETGEAR Nighthawk 5G M7 ($499)   ← carrier-free pick
   • Value/control→ GL.iNet Puli AX ($379, OpenWrt+VPN)
   ▸ filed → wiki/research-reports/best-5g-router-2026.md

💡 Want a real demo GIF here? Record with asciinema or vhs and drop it in assets/.

---

🧠 How it works

flowchart LR
    Q([Your question]) --> P[Intake & scope]
    P --> W[Wiki + memory pre-search]
    W --> F{Fan out}
    F --> E1[Explorer: experts]
    F --> E2[Explorer: contrarian]
    F --> E3[Explorer: live data]
    F --> E4[Explorer: verify specs]
    E1 & E2 & E3 & E4 --> S[Synthesize / score]
    S --> R[["⚔️ Adversarial red-team"]]
    R -- revise --> S
    R -- pass --> O[Cited report + decision]
    O --> WIKI[(Wiki + memory)]
    WIKI -. compounds .-> W
    O --> L["Self-learning:lesson to program.md"]
    L -. tunes next run .-> P

The loop is the point: pre-search → fan-out → verify → file → learn, and every run leaves the system a little sharper.

---

🚀 Quick start

git clone https://github.com/justsima/agentic-stack.git
cd agentic-stack
./install.sh

The installer is interactive, idempotent, and sudo-free. It auto-wires the required (free, key-less) MCP servers — Exa · agentmemory · Context7 — asks about optional ones (SearXNG, Jina, the claude-obsidian wiki engine, graphify), and creates an empty wiki scaffold.

Prefer the in-app plugin marketplace?

/plugin marketplace add justsima/agentic-stack
/plugin install ultradeep@agentic-stack
/plugin install market-scout@agentic-stack
/plugin install llm-council@agentic-stack
/plugin install adhd@agentic-stack
/plugin install agentic-config@agentic-stack

Then add the required MCP servers:

claude mcp add --scope user --transport http exa https://mcp.exa.ai/mcp
claude mcp add --scope user agentmemory -- npx -y @agentmemory/agentmemory
claude mcp add --scope user --transport http context7 https://mcp.context7.com/mcp

Other flags

./install.sh --dry-run   # show everything it would do, change nothing
./install.sh -y          # non-interactive: required = yes, optional = no
./install.sh --help

Requirements: Claude Code · node/npx (for agentmemory) · Docker (optional, for SearXNG). Using Cursor, Codex or Antigravity instead? → [Install in your tool](#-install-in-your-tool).

---

🧰 Install in your tool

Works in any agent that supports the open Agent Skills (SKILL.md) + MCP standards — two pieces: (A) copy the skills, (B) add the MCP servers.

| Your tool | Install the skills (one command) | Skills land in | |---|---|---| | Claude Code | ./install.sh (or /plugin marketplace add justsima/agentic-stack) | ~/.claude/skills/ | | Cursor | ./skills-portable/sync-skills.sh cursor | .cursor/skills/ | | Codex | ./skills-portable/sync-skills.sh codex | ~/.agents/skills/ | | Antigravity | ./skills-portable/sync-skills.sh antigravity | ~/.gemini/antigravity/skills/ | | Gemini CLI | ./skills-portable/sync-skills.sh gemini | ~/.gemini/skills/ |

Then add the 3 MCP servers (any tool — all free, no keys for 2 of 3)

| Server | Transport | Endpoint / command | |---|---|---| | exa | HTTP | https://mcp.exa.ai/mcp | | context7 | HTTP | https://mcp.context7.com/mcp | | agentmemory | stdio | npx -y @agentmemory/agentmemory |

Add them in your tool's MCP settings (Claude Code / Codex: mcp add; Cursor / Antigravity / Gemini: MCP config UI). Full per-tool steps in [skills-portable/](skills-portable/README.md).

> ultradeep is a Claude Code slash-command (+ sub-agents), so it's Claude Code-only. market-scout · llm-council · adhd · job-application-helper run everywhere.

---

📊 Analytics & traffic

Everything you need to measure reach — and the honest truth about each metric:

| Signal | How it's tracked | Where | |---|---|---| | ⭐ Stars / forks | Live shields.io badges | top of this README | | 👁️ README views | Live hits.sh counter badge | top of this README | | ⬇️ "Downloads" (git clones) | Self-archived daily by a GitHub Action (GitHub deletes traffic after 14 days) | [traffic/REPORT.md](traffic/REPORT.md) | | 📈 Clones + views history | Permanent CSVs, upserted daily | [traffic/views.csv](traffic/views.csv) · [traffic/clones.csv](traffic/clones.csv) | | 🌟 Star growth over time | star-history (live, below) | [chart](#-star-history) | | 🔥 Contributor/issue/PR activity | Repobeats embed (activate once — see below) | this section |

> Why a self-hosted archiver? GitHub has no plugin-install counter, and native traffic (views + clones) vanishes after 14 days. The bundled [.github/workflows/analytics.yml](.github/workflows/analytics.yml) snapshots views + clones every day into traffic/, turning that 14-day window into a permanent record. Runs automatically; works with the built-in token (add a GH_TRAFFIC_TOKEN PAT secret if your account 403s the traffic API).

Activate the Repobeats activity graph (30 seconds)

1. Go to repobeats.axiom.co → add justsima/agentic-stack.
2. Paste the snippet it gives you here:

```html

```

---

🧪 The science behind it

These aren't arbitrary — each pattern is grounded in published work:

• ultradeep — Anthropic's multi-agent Research, Stanford STORM (perspective questioning), GPT-Researcher (plan→fan-out→synthesize), adversarial verification.
• market-scout — classic MCDA (multi-criteria decision analysis): min-max normalization + weighted scoring + value-per-dollar, so the ranking is reproducible.
• llm-council — Karpathy's LLM Council + DMAD (ICLR 2025): distinct reasoning methods beat distinct personas.
• adhd — isolated parallel branches + separated generate/critique phases to fight premature convergence.

---

🔧 Tuning

• ultradeep: edit ~/.claude/deep-research/program.md (seeded by the installer) — priorities, depth, source policy, search backends. Its Domain Notes section grows itself.
• market-scout: edit plugins/market-scout/skills/market-scout/references/criteria.json — add product categories + weight profiles (default / value / performance / travel).

---

🔒 Privacy & security

• No secrets, no personal data ship here. Free MCP endpoints need no keys; agentmemory runs locally.
• Your wiki & memory are empty scaffolds — the engine, not anyone's data.
• job-application-helper reads a git-ignored profile.md you create locally.
• Safety hooks block destructive bash + secret leaks. Full notes: [docs/SECURITY.md](docs/SECURITY.md).

---

❓ FAQ

Do I need any paid API keys? No. The required MCP servers — Exa (public), agentmemory, Context7 — are free and key-less. The optional ones (SearXNG, Jina) are free too.

Does it really work in Cursor & Codex? The skills use the open Agent Skills SKILL.md standard. Run ./skills-portable/sync-skills.sh <dir> to copy them into another tool's skills folder. (ultradeep is a slash-command + sub-agents, so it's Claude Code-only for now.)

Is my data private? Yes — nothing personal ships. Your wiki & memory are empty scaffolds, and job-application-helper reads a git-ignored profile.md you create locally. See SECURITY.md.

How are "downloads" counted? GitHub has no plugin-install counter, so downloads ≈ git clones. GitHub deletes traffic after 14 days, so the bundled Action archives views + clones daily into traffic/.

Will it use a lot of tokens? The deep pipelines (ultradeep, llm-council) spawn many sub-agents by design — powerful but token-heavy. market-scout and adhd are lighter. Each skill documents its own cost.

🗺️ Roadmap

• [ ] Real demo GIFs per plugin
• [ ] More market-scout categories (TVs, headphones, monitors, GPUs)
• [ ] ultradeep knowledge-graph view
• [ ] Submit to public Claude Code plugin directories
• [ ] One-click cross-tool installer for Cursor & Codex

Ideas? Open an issue 🙌

---

🤝 Contributing

PRs welcome! Add a plugin under plugins// with a .claude-plugin/plugin.json, list it in .claude-plugin/marketplace.json, and run claude plugin validate .. Keep the privacy rule sacred: ship the engine, never the data.

---

👥 Contributors & activity

Want your face here? PRs welcome.

⭐ Star history

[](https://star-history.com/#justsima/agentic-stack&Date)

If this saves you time, a ⭐ helps others find it.

---

MIT © justsima · Companion tools (claude-obsidian, graphify, agentmemory) are separate open-source projects.

Built with Claude Code 🤖

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: justsima
• Source: justsima/agentic-stack
• License: MIT
• Homepage: https://github.com/justsima/agentic-stack

Install and usage instructions live in the source repository linked above.