Quick start · How it works · AGENTS.md · llms.txt · Skills · Discussions · All projects
Where agents gear up: 60,000+ open-source agent components, ranked on public evidence.
Armory is an agent employee, a packaged specialist agent with its own skills, memory and tests. It owns one job: MCPs and the component registry. Its persona is Armiger. Coding agents use it through its skills, its MCP server and its command line. A nightly job grows its catalogue, reruns every gate, and commits the result to main only when all of them pass.
Armory is part of an agent-native play for agent-native systems that can eventually be recursively self-improving and evolving, operated by agent employees.
🔗 Live registry → armory-murex.vercel.app · Browse 60,000+ components · Stack · Components
Last updated: 2026-11-14T14:01:50.765Z (UTC), when catalog.json was last generated.
Armory is the open-source, ever-evolving registry of every agent-harness component: MCP servers, skills, hooks, sub-agents, identity files, memory patterns, rules, CLIs, eval frameworks, observability, infrastructure, and the workflows that compose them. It is for coding agents, and the people who run them, who would rather reach for proven gear than hand-roll it: one armory your agent reaches into to grab exactly what it needs. For agents, by agents, of agents.
Claude Code: the plugin brings the skills, sub-agents, hooks and MCP server.
/plugin marketplace add naman10parikh/armory
/plugin install armory@armory
Its MCP server, Codex, Cursor, Gemini CLI, one skill and the command line
Its MCP server, in any MCP client:
claude mcp add --transport stdio armory -- npx -y -p github:naman10parikh/armory armory-mcp
codex mcp add armory -- npx -y -p github:naman10parikh/armory armory-mcpCursor (.cursor/mcp.json), Gemini CLI (.gemini/settings.json) and most other clients take the same command as JSON:
{
"mcpServers": {
"armory": {
"command": "npx",
"args": [
"-y",
"-p",
"github:naman10parikh/armory",
"armory-mcp"
]
}
}
}Gemini CLI: the extension brings the context and the MCP server.
gemini extensions install https://github.com/naman10parikh/armoryOne skill (4 in all, in skills/), in any agent that reads Agent Skills:
gemini skills install https://github.com/naman10parikh/armory --path armory-skill
git clone --depth 1 https://github.com/naman10parikh/armory /tmp/armory && cp -r /tmp/armory/armory-skill ~/.claude/skills/The command line:
npx -y -p github:naman10parikh/armory armory --helpCodex, Cursor, OpenCode, GitHub Copilot and other agents read AGENTS.md; Claude Code reads CLAUDE.md and Gemini CLI reads GEMINI.md, which import it. llms.txt maps every surface for agents that read first.
These commands were run for this README. The npm packages start fastest; the GitHub form above builds this
repository's latest code on first use, which took 90 seconds when this was written. npm 11.6.1 and 11.6.2 can stop
that first build with "Lock compromised", an npm bug fixed in 11.6.3: run npm i -g npm@latest first.
npx -y @namanparikh/armory search "browser automation" # rank across 60,000+ components
npx -y @namanparikh/armory install github-mcp --dry-run # see what it would write
npx -y @namanparikh/armory install github-mcp # wire GitHub's own MCP server into this project
npx -y @namanparikh/armory init --claude # add the Armory MCP server (--cursor|--codex|--opencode|--gemini)Install it once with npm i -g @namanparikh/armory; the command is armory. The MCP server alone is
npx -y armory-mcp. The newest commands (sandbox-run, memory-search, runs) are on GitHub ahead of the next npm
release: npx -y -p github:naman10parikh/armory armory --help.
🧰 COMPONENT TYPES: every harness component is its own root folder
mcps/ (~21,850) skills/ (~900) subagents/ (~991) workflows/ (~716)
claudemd-rules/ hooks/ clis-tools/ evals/ observability/
infrastructure/ memory/ identity/
└─ each <type>/ holds installable files; full metadata → brain/components/<type>/
📚 brain/ The catalog's source of truth (Obsidian vault): a detail page per component
catalog.json └─ generated index (never hand-edited; counts grow nightly)
🔧 ARMORY'S OWN HARNESS: install Armory itself
armory-mcp/ armory-skill/ cli/ skills/armory-* .claude/agents/armory-*
.claude-plugin/ .codex-plugin/ .hermes-plugin/ .cursor/ .codex/
.opencode/ .gemini/ opencode.json .mcp.json install.sh install.ps1
⚙️ THE MACHINERY: grow and self-improve
ingest/ (crawlers, promote, catalog, validate, test-gate, surface)
incoming/ (ephemeral staging) repos/ (source-repo watchlist) autolab.yml dsip.json
🌐 SITE + DOCS
web/ (the registry website → armory-murex.vercel.app) docs/ scripts/ (dev helpers)
Where are the MCPs?
mcps/<slug>.jsonholds install configs; details are inbrain/components/mcps/<slug>.md. The full map of every folder is STRUCTURE.md.
Armory mirrors and de-duplicates the biggest open-source catalogs into one connected brain, so it holds more harness components than any single registry:
| Category | What it holds |
|---|---|
| MCPs | PulseMCP, Smithery, Glama, mcp.so, awesome-mcp-servers and the official modelcontextprotocol/servers, deduped into ~61,700 (26 September 2026) |
| Skills | anthropics/skills, obra/superpowers, community skill packs, awesome-claude-code |
| Sub-Agents | wshobson, VoltAgent, davila7 and community collections |
| Rules | awesome-cursorrules, multi-language rule packs, CLAUDE.md collections |
| Hooks | disler hooks-mastery, decider, awesome-claude-code and community hook sets |
| CLIs & Tools · Evals · Observability · Infrastructure | the four categories every other list under-covers, first-class here |
| Identity · Memory · Workflows | SOUL and persona specs, memory layers, command and composition recipes |
Every entry is de-duplicated by name, carries provenance (source_repo, license, stars, verified_at), and is
wired into a knowledge graph through related: links, so following one component leads you to the next.
Curated lists get starred and forgotten in a browser tab. Armory is built to be used by the agents themselves:
- 🤖 Agents read it. Your coding agent pulls the right component straight into its harness.
- 🔎 Agents query it. The Armory MCP server, the
armoryCLI and the search site all read one generatedcatalog.json. - 🛠️ Agents install from it.
armory install <name>fetches the real thing (throughghor git) and drops it into whatever harness you are in. - ♻️ Agents improve it. It restocks itself nightly (
autolab.yml, a remote cron, no laptop needed).claude -pdiscovers new sources (Darwin DSIP Performance-Novelty selection); every candidate passes the Hamel test-gate, functional and behavioral, before it enters, so skills cannot drift or go stale. Then dedup, validate, commit.
armory install detects your harness (.claude/, .cursor/, .codex/, .opencode/, .gemini/), fetches the
component from its source repository, and installs it the right way for that CLI: MCP servers into the MCP config,
skills into the skills folder, sub-agents, hooks, rules and commands into their slots. One command, any harness, any
component type. Target another harness with --cli cursor (or codex, opencode, gemini).
Before you install a stranger's code, armory sandbox-run <name> clones it inside an isolated E2B sandbox and reports
its manifests, install-time scripts and tests (it needs E2B_API_KEY). armory runs shows every install, submit and
sandbox run, and armory memory-search searches Armory's own notes.
Install it once and your agent can search and pull from all 60,000+ components (61,739 on 28 September 2026), with the ~3,170 vendored skills, sub-agents, workflows, hooks and rules right there.
| Harness | Install |
|---|---|
| Codex | codex plugin marketplace add naman10parikh/armory |
| OpenCode | add armory-mcp to opencode.json, see PLUGIN.md |
| Hermes | hermes plugin add naman10parikh/armory |
| Anything | npx @namanparikh/armory init (detects the harness; --claude --cursor --codex --opencode --gemini to pick) or bash install.sh |
The plugin wires in the Armory MCP server (armory-mcp: live search at runtime; installing is CLI-only) and the
armory skill. Full per-harness commands are in PLUGIN.md; Armory's own harness against the catalog
is explained in HARNESS.md.
The catalog is not only metadata. The real component files are vendored in:
skills/ subagents/ workflows/ hooks/ claudemd-rules/: ~3,170 actual files at the repository root (content verbatim, with a provenance header).- Harness-native drops:
.claude/{skills,agents,commands,hooks}and.cursor/rules/, ready to use as they are. brain/stays the metadata knowledge graph (the source of truth). MCP details live inbrain/components/mcps/<slug>.md. MCP servers are cataloged as install specs incatalog.json; the servers themselves run elsewhere.
brain/ is an Obsidian vault: Markdown, frontmatter and [[wikilinks]]. Open the folder in Obsidian and explore the
whole graph; the related: edges are the connections between gear. One source of truth generates the site, the MCP
server and the CLI (counts are computed, never hand-typed).
Armory is a registry: add a component through a pull request or armory submit. CONTRIBUTING.md
has the one-file format, and CODE_OF_CONDUCT.md the ground rules. Every contribution restocks
the armory for every agent.
Armory is for agents, by agents, of agents: agent-native and recursively self-improving, built almost entirely by
agents. None of it would exist without the open-source catalogs, collections and reference repositories it mirrors,
de-duplicates and learns from. The self-improving loop keeps tabs on them so their newest work
keeps flowing in; the live list with freshness is repos/watchlist.json.
- Navigation and awesome-lists: hesreallyhim/awesome-claude-code, VoltAgent/awesome-claude-code-subagents, PatrickJS/awesome-cursorrules, vijaythecoder/awesome-claude-agents
- Mega-collections: davila7/claude-code-templates, affaan-m/ecc (ECC), wshobson/agents and wshobson/commands, dl-ezo/claude-code-sub-agents
- Anthropic official: anthropics/skills, anthropics/anthropic-cookbook, anthropics/anthropic-quickstarts, modelcontextprotocol/servers
- Hooks and skills: disler/claude-code-hooks-mastery, decider/claude-hooks, obra/superpowers
- MCP registries: PulseMCP, Smithery (smithery-ai/registry), Glama, mcp.so
- Protocol: google/A2A
Every cataloged entry carries its source_repo and license provenance back to one of these. Thank you. 🙏
Every project in this fleet is graded on the same 11 harness components and on self-improvement. Each mark is backed by a file you can open.
| Component | Grade | Where |
|---|---|---|
| Identity | ✅ | SOUL.md:8 |
| Memory | ✅ | cli/src/memory.ts:1 |
| Skills | ✅ | skills/armory-vet-component/SKILL.md:1 |
| Tools/CLIs | ✅ | cli/package.json:7 |
| Hooks | ✅ | .claude/settings.json:4 |
| Sub-agents | ✅ | .claude/agents/armory-curator.md:1 |
| MCPs | ✅ | armory-mcp/src/index.ts:45 |
| Dispatch | ✅ | .github/workflows/autolab.yml:17 |
| Evals | ✅ | .github/workflows/ci.yml:29 |
| Observability | ✅ | cli/src/runlog.ts:1 |
| Sandbox/Infra | 🟡 | cli/src/sandbox.ts:63 (live boot needs the founder's E2B key) |
| Self-improvement | ✅ | dsip.json:1 |
✅ real and tested · 🟡 partial, with the reason in the row · ❌ absent
It carries its own evals and a self-improvement contract, dsip.json, which says what it may change and how a change is judged. A nightly job grows its catalogue, reruns every gate, and commits the result to main only when all of them pass. Lessons it learns are kept as rules in its memory files.
One of the projects by Naman Parikh. Each owns one part of an agent's harness, or of the product around an agent, and all of them follow one standard.
| Related project | What it owns |
|---|---|
| AgentDial | Identity |
| WikiMem | Memory |
| Skillsmith | Skills |
| Agent Booster | Tools / CLIs |
| Hooksmith | Hooks |
| AgentCast | Sub-agents |
| Plugsmith | Plugins |
| AgentGrid | Dispatch |
| AgentSwarm | Dispatch |
| AgentBench | Evals |
| AgentMoney | Observability |
| Sandforge | Sandbox / Infra |
| Claude Harness | The whole harness, as a template |
Questions and ideas go to Discussions; bugs go to issues. Follow the work on X and LinkedIn.
Read CONTRIBUTING.md. AI-assisted pull requests are welcome. Report a vulnerability privately, as SECURITY.md explains.
MIT, © 2026 Naman Parikh. See LICENSE and NOTICE.
If it is useful to you, a star helps other people find it.