| name | agntcy-dir |
| description | 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; verify signatures, name ownership, or security scans; synchronize data between servers; or install/uninstall agents and agentic resources into coding agents like VS Code Copilot, Claude Code, or Cursor. |
| metadata | {"author":"AGNTCY Contributors","version":"1.0.0"} |
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 |
| "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", "prove name ownership" | references/publishing.md |
| "find/search/suggest agents, MCP servers, skills", "browse the directory", "pull a record" | 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 into my editor/agent", "uninstall", "export as mcp.json / SKILL.md / A2A card" | references/install.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 (
docker-image, source_code, helm_chart, …).
- Referrer — Artifact attached to a record: signature, public key, security scan report.
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, install list, and version. 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 jsonl for 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, mcp serve, 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).
- Destructive ops (
delete, uninstall, sync delete, --force):
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 (
init downloads ~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.
MCP-native alternative
dirctl mcp serve starts a built-in MCP server exposing DIR and OASF
operations as tools (push, pull, search, validate, taxonomy browsing). When a
host is configured with this server, prefer its tools over shelling out to
equivalent CLI commands. To wire it up, add a stdio server entry that runs
dirctl mcp serve to the host's MCP config, following the merge-never-
overwrite rules in references/install.md.
Pointers