agntcy/dir/skill
Claude Code and agent skill by agntcy/dir
- community
Checking that the SKILL.md file is still in the repository.
Use when the user asks to discover, browse, search, suggest, recommend, or install agents and agentic resources. Author, validate, or import OASF records; push, sign, and publish records; discover, search, browse, suggest, or recommend agents like MCP servers, A2A agents, or agent skills; claim and prove record ownership or identity (DNS, HTTPS, DID, SPIFFE, ANS); verify signatures, ownership claims, or security scans; list running agent workloads; configure node content policies and prune jobs…
Install
npx vskill@latest install agntcy/dir/skillSKILL.md
AGNTCY Directory (dirctl)
The AGNTCY Directory (DIR) is a distributed, peer-to-peer registry of AI agent records described in OASF. Records are content-addressed by CID and carry skills, domains, modules (MCP, A2A, Agent Skills), locators, signatures, and security scan results.
This skill turns natural-language requests into dirctl workflows. It is a router: read the matching reference file below _before_ acting on a workflow. Read more than one when the request spans workflows (e.g. "import and publish" → authoring + publishing).
Dispatch
| User intent (examples) | Read first | |
|---|---|---|
| "install dirctl", "set up a local directory", "start the daemon", "configure contexts", "connection issues" | references/setup.md | |
| "log in", "am I authenticated?", "switch to another directory node", "permission denied / forbidden" | references/setup.md | |
| "create a record", "validate my record", "import MCP servers / A2A cards / skills into DIR" | references/authoring.md | |
| "push", "publish to the network", "sign my record" | references/publishing.md | |
| "prove I own this", "claim ownership", "publish my key", "why is my claim failed/not verified" | references/identity.md | |
| "find/search/suggest agents, MCP servers, skills", "browse the directory", "pull a record" | references/discovery.md | |
| "what agents are running", "find a running workload", "agents reachable over SLIM" | references/discovery.md | |
| "is this record signed/safe/verified?", "check scan reports", "verify signature" | references/verification.md | |
| "sync with another directory", "mirror records", "sync from a public registry/GHCR" | references/sync.md | |
| "install <record> into my editor/agent", "uninstall", "export as mcp.json / SKILL.md / A2A card" | references/install.md | |
| "content policy", "hide/block records on my node", "shadow/enforce", "prune untrusted records", "upgrade my node to v1.8" | references/policy.md |
Vocabulary
- Record — one OASF JSON document describing an agent (name, version, skills, domains, locators, modules, annotations).
- CID — content identifier; hash of the record bytes. Same bytes → same CID. Any change → new CID.
- Reference — how commands address records:
<cid>,<name>,<name>:<version>,<name>@<cid>,<name>:<version>@<cid>(the@<cid>forms are hash-verified and fail on mismatch). - Skill / Domain — OASF taxonomy nodes describing what an agent does and where.
- Module — typed payload on a record:
integration/mcp(MCP server),integration/a2a(A2A card),core/language_model/agentskills(Agent Skill). - Locator — pointer to a runnable artifact (
container_image,source_code,helm_chart,binary, …; each with aurlslist). - Referrer — Artifact attached to a record: signature, public key, security scan report, identity claim.
- Subject / Claim — a subject is who a record says it belongs to (
dns:example.com,did:web:…,spiffe://…,ans://…), declared in theagntcy.dir/owner/agntcy.dir/identityannotations. A claim is a signature by that subject's key; the reconciler re-verifies it against the subject's published keys on every run. Record names are not verified — claims are. - Content policy — a node operator's CEL/OPA rule; records failing an enforced policy are withheld by that node only.
Prerequisites
dirctl must be on PATH — probe with dirctl version (works in every shell; POSIX: command -v dirctl, PowerShell: Get-Command dirctl). If missing, follow the install section of references/setup.md — never suggest building from source.
A reachable Directory server is required for everything except context, validate, version, and the local-state install commands (install agents, install list, install pin, install unpin, install prune, uninstall). Default is localhost:8888 (local daemon). Server selection order: --context <name> flag → DIRECTORY_CLIENT_CONTEXT → current_context in config → --server-addr / DIRECTORY_CLIENT_SERVER_ADDRESS overrides.
CLI conventions (apply everywhere)
- Parse, don't scrape: request
-o json(or-o jsonlfor streams) when
you need to read results; -o raw for CIDs in shell pipelines. Structured formats write data to stdout and messages to stderr, so piping to jq is safe. validate, context, auth, daemon, and version do not take -o.
- Quote every argument — record names contain
/and:. - Capability probing: command surface varies by version. Before relying on
newer commands (install, init, NL search), check dirctl <cmd> --help; if absent, use the documented fallback in the relevant reference file.
- Auth: never invent auth flags. On an auth error, surface the raw error
and ask the user how to authenticate (see setup reference: auth login, --auth-mode, --auth-token).
- Authentication is not authorization: a valid token from
auth status
says nothing about what the user may do. Gateways enforce per-method access control, so PermissionDenied / Forbidden on push, sign, routing publish, delete, or sync is an entitlement problem — logging in again cannot fix it. Name the denied method, say it needs elevated rights on that deployment, and stop; only Unauthenticated and expired-token errors are login problems.
- Destructive ops (
delete,uninstall,sync delete,--force, a
reconciler prune job with dry_run: false, switching a content policy to enforce): always confirm with the user first.
- Never fabricate CIDs, names, versions, scores, or scan results — every
value shown must come from real command output.
Platforms & shells
dirctl ships for Linux and macOS (amd64 + arm64) and Windows (amd64 only — no windows-arm64 build; the binary is dirctl.exe). Homebrew covers macOS/Linux only; on Windows install from GitHub Releases (see setup reference). dirctl behaves identically everywhere — only the wrapping shell syntax and paths differ.
Emit commands for the shell the host actually runs — never assume bash. Code blocks in this skill are written for POSIX shells (bash/zsh) and run as-is on Linux/macOS. On Windows (PowerShell), translate before running:
| bash idiom | PowerShell equivalent | ||
|---|---|---|---|
CID=$(dirctl … -o raw) … "$CID" | $CID = dirctl … -o raw … $CID | ||
for f in ./out/*.json; do … done | `Get-ChildItem ./out/*.json \ | ForEach-Object { … $_.FullName … }` | |
command -v dirctl | Get-Command dirctl | ||
mktemp | (New-TemporaryFile).FullName | ||
chmod +x dirctl | not needed (.exe) | ||
cmd1 && cmd2 | run sequentially as separate commands |
Path rules: ~ in this skill means the user's home directory — %USERPROFILE% on Windows (e.g. ~/.agntcy/dir/ → %USERPROFILE%\.agntcy\dir\). When unsure where the effective client config lives, run dirctl context show instead of guessing paths.
Environment adaptivity
Maximize the host's native capabilities; degrade gracefully:
| Capability | VS Code Copilot | Claude Code | Copilot CLI / other CLIs | |
|---|---|---|---|---|
| Confirm selections & choices | ask-questions tool (question card with options) | AskUserQuestion tool | Plain-text numbered list; wait for reply | |
| Display result sets | Markdown table | Markdown table | Markdown table (narrow columns) | |
| File links in summaries | Clickable relative Markdown links | Plain paths | Plain paths | |
| Post-install activation | Suggest "MCP: List Servers" → restart | Suggest restarting the session | Suggest restarting the tool |
Other agentic hosts (Cursor, Windsurf, Zed, Cline, Roo, Gemini CLI, opencode, Continue, Codex, …) are not listed per-column — detect capabilities instead of matching product names: if an interactive question tool exists, use it; otherwise behave like the CLI column. You generally know which host you are running in; when you don't, infer it from your available tools, not from guesses.
Rules:
- Use an interactive question tool whenever one exists for: picking records to
install, choosing install scope/agents, and confirming destructive actions. Only fall back to free-text prompts when no such tool is available.
- Never promise webviews, side panels, or native dialogs — chat surfaces
render Markdown only.
- Long operations (
initdownloads ~89 MB; imports and syncs can run
minutes): warn the user before starting and report progress from command output.
- When the user's host is itself an install target (e.g. VS Code →
vscode,
Claude Code → claude-code, Cursor → cursor), default installs to that host's agent ID (see install reference).
- Headless / CI runs (no human available to answer): never block on
interactive prompts. Use --yes for init and install, authenticate via DIRECTORY_CLIENT_AUTH_TOKEN or auth login --device / --no-browser, and when a decision genuinely requires a human, fail fast with a clear message instead of waiting.
Pointers
- CLI reference: <https://dir.agntcy.org/latest/dir/dir-cli-reference/>
- Quickstart: <https://dir.agntcy.org/latest/dir/dir-quickstart/>
- OASF schema: <https://schema.oasf.outshift.com>
- Repo: <https://github.com/agntcy/dir>
Security
- Security scanPassed, 100 of 100Pattern checks on the skill's files.
- Model reviewNot runA model review is scheduled for the next rescan.
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryChecking.skill/SKILL.md
Popularity
- Repository stars
- 199
- Repository forks
- 61
- 7-day trend score
- 26
Stars and forks are for the whole GitHub repository, not this skill alone. Refreshed 17 Sep 2026, a month ago.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/agntcy/dir/skill)