# Agentic Stack > Portable agentic research stack for Claude Code β€” multi-agent deep research (ultradeep), best-on-market product research (market-scout), 5-advisor council, divergent ideation, + a one-command installer. - **Type:** MCP server - **Install:** `agentstack add mcp-justsima-agentic-stack` - **Verified:** Yes β€” security-reviewed for prompt injection and unsafe behavior - **Seller:** [justsima](https://agentstack.voostack.com/s/justsima) - **Installs:** 0 - **Category:** [AI & ML](https://agentstack.voostack.com/c/ai-and-ml) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [justsima](https://github.com/justsima) - **Source:** https://github.com/justsima/agentic-stack - **Website:** https://github.com/justsima/agentic-stack ## Install ```sh agentstack add mcp-justsima-agentic-stack ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About **Six battle-tested plugins that turn Claude Code into a research & decision machine.** One-command install Β· works in Claude Code, Cursor & Codex Β· 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) [**πŸš€ Quick start**](#-quick-start) Β· [**🧩 Plugins**](#-the-plugins) Β· [**🧠 How it works**](#-how-it-works) Β· [**πŸ“Š Analytics**](#-analytics--traffic) Β· [**🧰 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: - πŸ§ͺ **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 β€” not vibes, not a hallucinated ranking. - πŸ” **They get smarter.** `ultradeep` writes a transferable lesson back into its own `program.md` after every run. Your tooling compounds. - πŸ—‚οΈ **They compound into a knowledge base.** Reports file themselves into an optional Obsidian wiki, cross-linked and searchable next time. - πŸ›‘οΈ **Privacy-first by design.** Ships the *engine*, never your data. No secrets, no personal content β€” 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**](https://github.com/justsima) β€” 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 | βœ… | | Adversarial red-team verification | ❌ | ❌ | βœ… | | Deterministic scoring β€” no hallucinated rankings | ❌ | ❌ | βœ… | | Self-learning across runs | ❌ | ❌ | βœ… | | Files into a knowledge base | ❌ | ❌ | βœ… | | One-command install + MCP wiring | β€” | β€” | βœ… | | Privacy-first β€” ships the engine, not your data | β€” | β€” | βœ… | --- ## 🧩 The plugins | | Plugin | What it gives you | Try it | |---|---|---|---| | πŸ”¬ | **ultradeep** | Multi-agent **deep research**: 8–10 parallel explorers β†’ STORM-style questioning β†’ **adversarial red-team** β†’ tiered search (Exa / SearXNG / WebSearch) β†’ cited, wiki-filed report. Self-tunes via `program.md`. | `/ultradeep ` | | πŸ›’ | **market-scout** | **"Best X on the market"** across Amazon/Best Buy/Walmart/Newegg. Parallel explorers + a **deterministic decision-matrix engine** + red-team β†’ a ranked, segment-aware buying report. | `/market-scout best ` | | πŸ›οΈ | **llm-council** | A decision through **5 advisors, each a distinct reasoning method** (inversion, decomposition, analogy, naive questioning, dependency graphing) β†’ anonymized peer review β†’ chairman synthesis. | `/llm-council ` | | 🧠 | **adhd** | **Divergent ideation** β€” N isolated idea branches under different cognitive frames, scored, pruned, deepened. Beats one-shot brainstorming on open-ended problems. | `/adhd ` | | πŸ“ | **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 β†’ drafts answers | | πŸ›‘οΈ | **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 ```text You β–Έ /market-scout best 5G router agentic-stack β–Έ dispatching explorers… (expert reviews Β· contrarian/reliability Β· live prices) β–Έ scoring 7 candidates across 4 weight profiles (deterministic MCDA) β–Έ red-team: PASS-with-revisions (caught 2 inflated ratings, 1 missing candidate) πŸ† No single winner β€” 4 picks by use case: β€’ Whole-home β†’ Ubiquiti UDR-5G-Max ($499) β€’ Mobile (AT&T)β†’ NETGEAR Nighthawk M7 Pro ($449, locked) β€’ Travel/eSIM β†’ NETGEAR Nighthawk 5G M7 ($499) ← carrier-free pick β€’ Value/controlβ†’ GL.iNet Puli AX ($379, OpenWrt+VPN) β–Έ filed β†’ wiki/research-reports/best-5g-router-2026.md ``` πŸ’‘ Want a real demo GIF here? Record with [asciinema](https://asciinema.org) or [vhs](https://github.com/charmbracelet/vhs) and drop it in `assets/`. --- ## 🧠 How it works ```mermaid 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[["βš”οΈ 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 β†’ fan-out β†’ verify β†’ file β†’ learn**, and every run leaves the system a little sharper. --- ## πŸš€ Quick start ```bash 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 β€” **Exa Β· agentmemory Β· Context7** β€” asks about optional ones (SearXNG, Jina, the claude-obsidian wiki engine, graphify), and creates an empty wiki scaffold. Prefer the in-app plugin marketplace? ```text /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: ```bash 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 ```bash ./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 Β· `node`/`npx` (for agentmemory) Β· Docker *(optional, for SearXNG)*. **Using Cursor, Codex or Antigravity instead?** β†’ [**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 β€” 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 β€” 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 Β· llm-council Β· adhd Β· job-application-helper** run everywhere. --- ## πŸ“Š Analytics & traffic Everything you need to measure reach β€” and the honest truth about each metric: | Signal | How it's tracked | Where | |---|---|---| | ⭐ **Stars / forks** | Live shields.io badges | top of this README | | πŸ‘οΈ **README views** | Live [hits.sh](https://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) | | πŸ“ˆ **Clones + views history** | Permanent CSVs, upserted daily | [`traffic/views.csv`](traffic/views.csv) Β· [`traffic/clones.csv`](traffic/clones.csv) | | 🌟 **Star growth over time** | star-history (live, below) | [chart](#-star-history) | | πŸ”₯ **Contributor/issue/PR activity** | Repobeats embed *(activate once β€” 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](https://repobeats.axiom.co)** β†’ add `justsima/agentic-stack`. 2. Paste the snippet it gives you here: ```html ``` --- ## πŸ§ͺ The science behind it These aren't arbitrary β€” each pattern is grounded in published work: - **ultradeep** β€” Anthropic's multi-agent Research, Stanford **STORM** (perspective questioning), GPT-Researcher (planβ†’fan-outβ†’synthesize), adversarial verification. - **market-scout** β€” classic **MCDA** (multi-criteria decision analysis): min-max normalization + weighted scoring + value-per-dollar, so the ranking is reproducible. - **llm-council** β€” Karpathy's *LLM Council* + **DMAD** (ICLR 2025): distinct reasoning methods beat distinct personas. - **adhd** β€” isolated parallel branches + separated generate/critique phases to fight premature convergence. --- ## πŸ”§ Tuning - **ultradeep:** edit `~/.claude/deep-research/program.md` (seeded by the installer) β€” priorities, depth, source policy, search backends. Its **Domain Notes** section grows itself. - **market-scout:** edit `plugins/market-scout/skills/market-scout/references/criteria.json` β€” add product categories + weight profiles (`default` / `value` / `performance` / `travel`). --- ## πŸ”’ Privacy & security - **No secrets, no personal data** ship here. Free MCP endpoints need no keys; agentmemory runs locally. - Your **wiki & memory are empty scaffolds** β€” 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 β€” Exa (public), agentmemory, Context7 β€” 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 β€” 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 β‰ˆ 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 β€” powerful but token-heavy. market-scout and adhd are lighter. Each skill documents its own cost. ## πŸ—ΊοΈ 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](https://github.com/justsima/agentic-stack/issues) πŸ™Œ --- ## 🀝 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.** --- ## πŸ‘₯ 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** Β© [justsima](https://github.com/justsima) Β· Companion tools (claude-obsidian, graphify, agentmemory) are separate open-source projects. Built with Claude Code πŸ€– ## Source & license This open-source MCP server is cataloged on AgentStack and links to its original source β€” we do not rehost the code. - **Author:** [justsima](https://github.com/justsima) - **Source:** [justsima/agentic-stack](https://github.com/justsima/agentic-stack) - **License:** MIT - **Homepage:** https://github.com/justsima/agentic-stack Install and usage instructions live in the source repository linked above. ## Pricing - **Free** β€” Free ## Security capabilities Automated source analysis of v0.1.0 β€” what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source β€” more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** β€” security scan: passed β€” Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/mcp-justsima-agentic-stack - Seller: https://agentstack.voostack.com/s/justsima - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack β€” the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.