Skip to main content
guidesSource-backed

Agent Skills in Claude Agent SDK Applications

A practical walkthrough of using Agent Skills in the Claude Agent SDK: how skills are discovered from the filesystem via settingSources, the skills option to enable or filter them, tool access, and troubleshooting discovery.

by JPette1783·added 2026-06-05·
Review first review before installing

Open the source and read safety notes before installing.

Citation facts

Source-backed facts for citing this resource, derived directly from the registry — also available as plain text for AI assistants.

Source URLs
https://code.claude.com/docs/en/agent-sdk/skills, https://github.com/JSONbored/awesome-claude/blob/main/content/guides/agent-skills-in-claude-agent-sdk-applications.mdx
Safety notes
The skills option is a context filter, not a sandbox: unlisted skills are hidden from the model but their files remain on disk and are reachable via Read and Bash., Skills are model-invoked; pair them with a tight allowedTools list (and dontAsk where appropriate) so an invoked skill cannot use more tools than intended., The allowed-tools frontmatter in SKILL.md does not apply through the SDK; control tool access with the main allowedTools option.
Privacy notes
Skill descriptions are loaded so the model can decide when to use them; keep sensitive workflow detail and secrets out of descriptions., Skills sourced from outside your project run their instructions in your sessions; review them before enabling., Skill content is sent to the model provider when a skill is invoked; treat it like any other prompt content.
Author
JPette1783
Submitted by
JPette1783
Claim status
unclaimed
Last verified
2026-06-05

Decision playbook

Review trust signals before you adopt

Signals are present but mixed. Use the checklist below to confirm the source and operational safety for your environment.

Compare context
Selected

0

Current score

63

Baseline

Delta

No baseline selected

No major trust-signal divergence detected in the current selection.

Source and provenance checks

Needs review

Confirm ownership and provenance before trusting install instructions.

  • Source link availableRequired

    Open the canonical repository and verify ownership.

    Done
  • Source provenance statusRequired

    Marked as source-backed.

    Done
  • Metadata reviewed

    No reviewed flag detected in metadata.

    Pending

Safety and privacy checks

Complete

Validate risk disclosures before installation or API wiring.

  • Safety notes presentRequired

    Review the listed safety guidance before running commands.

    Done
  • Privacy notes presentRequired

    Review data handling notes before connecting accounts or secrets.

    Done
  • Trust level risk gateRequired

    Trust level does not block evaluation.

    Done

Package and install checks

Needs review

Check package metadata and artifact integrity signals.

  • Install payload available

    Install or copy payload is available for review.

    Done
  • Package verification flag

    No package verification flag provided.

    Pending
  • Checksum metadata

    No checksum provided for downloaded artifact.

    Pending

Compare-driven decision checks

Needs review

Use compare context to validate trade-offs before adoption.

  • Compare tray has multiple entries

    Add at least one more entry to compare trust differences.

    Pending
  • Baseline comparison available

    No baseline peer selected yet.

    Pending
  • Diverging trust signals identified

    No major trust-signal divergence found.

    Pending

Setup at a glance

Copy & paste

Copy-ready — paste the snippet to get started.

Adoption plan

Balanced adoption plan

Current risk score 24/100. Use staged verification before broader rollout.

Risk 24

Pre-adoption checks

Validate source and review signals before any execution.

  • Confirm source provenanceRequired

    Source URL/provenance metadata is present.

    Done
  • Confirm metadata review state

    No review metadata found; increase manual validation.

    Pending
  • Verify install payload

    Install/config payload exists and can be inspected.

    Done

Security checks

Confirm safety, privacy, and package integrity signals.

  • Review safety notesRequired

    Safety notes are present.

    Done
  • Review privacy notesRequired

    Privacy notes are present.

    Done
  • Verify package integrity metadata

    No package verification/checksum metadata.

    Pending

Rollout

Adopt in controlled steps based on the selected plan.

  • Run in isolated sandbox firstRequired

    Use a constrained sandbox and observe behavior across multiple tasks.

    Pending
  • Roll out graduallyRequired

    Roll out to a small cohort before wider usage.

    Pending
  • Set monitoring and fallback

    Define rollback path and monitor errors after adoption.

    Pending

Evidence readiness

Evidence readiness matrix · balanced

Missing required evidence: Metadata review. Risk score 31.

Risk 31

Source provenance

Present

Source repository/provenance is listed.

Required in this preset

Metadata review

Missing

Review metadata is missing.

Required in this preset

Safety notes

Present

Safety notes are present.

Required in this preset

Privacy notes

Present

Privacy notes are present.

Optional in this preset

Package integrity

Missing

Package integrity metadata is missing.

Optional in this preset

Install payload

Present

Install payload is available.

Required in this preset

Required gaps: Metadata review

Decision timeline

Decision timeline · balanced

Blocking gaps: Check metadata review status. Risk 28.

Risk 28

triage

Confirm source provenanceRequired

Source/provenance metadata is available.

Done

triage

Check metadata review statusRequired

Review metadata is missing.

Pending

verify

Review safety notesRequired

Safety notes are available.

Done

verify

Review privacy notes

Privacy notes are available.

Done

verify

Validate package integrity metadata

Package integrity metadata is missing.

Pending

rollout

Verify install payload and commandsRequired

Install payload is available.

Done

Blockers: Check metadata review status

Prerequisite readiness

Prerequisite readiness

3 prerequisites to line up before setup.

0/3 ready
Install & runtime1General2

Safety & privacy surface

Safety & privacy surface

3 safety and 3 privacy notes across 4 risk areas. Review closely: credentials & tokens, third-party handling.

4 areas
  • SafetyLocal filesThe skills option is a context filter, not a sandbox: unlisted skills are hidden from the model but their files remain on disk and are reachable via Read and Bash.
  • SafetyGeneralSkills are model-invoked; pair them with a tight allowedTools list (and dontAsk where appropriate) so an invoked skill cannot use more tools than intended.
  • SafetyGeneralThe allowed-tools frontmatter in SKILL.md does not apply through the SDK; control tool access with the main allowedTools option.
  • PrivacyCredentials & tokensSkill descriptions are loaded so the model can decide when to use them; keep sensitive workflow detail and secrets out of descriptions.
  • PrivacyCredentials & tokensSkills sourced from outside your project run their instructions in your sessions; review them before enabling.
  • PrivacyThird-party handlingSkill content is sent to the model provider when a skill is invoked; treat it like any other prompt content.

Safety notes

  • The skills option is a context filter, not a sandbox: unlisted skills are hidden from the model but their files remain on disk and are reachable via Read and Bash.
  • Skills are model-invoked; pair them with a tight allowedTools list (and dontAsk where appropriate) so an invoked skill cannot use more tools than intended.
  • The allowed-tools frontmatter in SKILL.md does not apply through the SDK; control tool access with the main allowedTools option.

Privacy notes

  • Skill descriptions are loaded so the model can decide when to use them; keep sensitive workflow detail and secrets out of descriptions.
  • Skills sourced from outside your project run their instructions in your sessions; review them before enabling.
  • Skill content is sent to the model provider when a skill is invoked; treat it like any other prompt content.

Prerequisites

  • The Claude Agent SDK installed for Python or TypeScript.
  • SKILL.md files in .claude/skills/ (project) or ~/.claude/skills/ (user).
  • A cwd that points at or below the directory containing .claude/skills/, within the same repository.

Schema details

Install type
copy
Troubleshooting
Yes
Full copyable content
## Overview

Agent Skills extend Claude with specialized capabilities packaged as `SKILL.md`
files that Claude invokes autonomously when relevant. In the Agent SDK, skills are
filesystem artifacts discovered through setting sources; there is no programmatic
API for registering them.

## How discovery works

Skills load from filesystem locations governed by `settingSources` (TypeScript) /
`setting_sources` (Python). With default `query()` options, the SDK loads `user`
and `project` sources, so skills in `~/.claude/skills/`, `<cwd>/.claude/skills/`,
and `.claude/skills/` in parent directories up to the repo root are available. If
you set `settingSources` explicitly, include `'user'` or `'project'` to keep
discovery.

## Enable and filter skills

Use the `skills` option to control which skills are available. Omit it to enable
discovered skills (matching CLI behavior), pass `"all"` for every discovered
skill, a list of names to enable only those, or `[]` to disable all.

```typescript
for await (const message of query({
  prompt: "Help me process this PDF document",
  options: {
    cwd: "/path/to/project",
    settingSources: ["user", "project"],
    skills: "all",
    allowedTools: ["Read", "Write", "Bash"],
  },
})) { /* ... */ }
```

Enable specific skills by name (matching the `name` in SKILL.md or the directory
name; use `plugin:skill` for plugin skills):

```typescript
const options = { skills: ["pdf", "docx"] };
```

When you set `skills`, the SDK adds the Skill tool to `allowedTools` automatically.
If you also pass an explicit `tools` list, include `"Skill"`.

## Tool access

The `allowed-tools` frontmatter in SKILL.md is honored only by the CLI, not the
SDK. Control tool access through the main `allowedTools` option; without a
`canUseTool` callback, anything not listed is denied. Pair with
`permissionMode: "dontAsk"` to deny anything outside the allow list.

## Troubleshooting

- **Skills not found**: confirm `settingSources` includes `user`/`project` (an
  empty array disables discovery), and that `cwd` points at or below the directory
  containing `.claude/skills/`.
- **Skill not used**: confirm its name is in the `skills` list (an empty list
  disables all) and that its description is specific.
- Verify on disk with `ls .claude/skills/*/SKILL.md`.

## Source

- Agent Skills in the SDK: https://code.claude.com/docs/en/agent-sdk/skills

About this resource

Overview

Agent Skills extend Claude with specialized capabilities packaged as SKILL.md files that Claude invokes autonomously when relevant. In the Agent SDK, skills are filesystem artifacts discovered through setting sources; there is no programmatic API for registering them.

How discovery works

Skills load from filesystem locations governed by settingSources (TypeScript) / setting_sources (Python). With default query() options, the SDK loads user and project sources, so skills in ~/.claude/skills/, <cwd>/.claude/skills/, and .claude/skills/ in parent directories up to the repo root are available. If you set settingSources explicitly, include 'user' or 'project' to keep discovery.

Enable and filter skills

Use the skills option to control which skills are available. Omit it to enable discovered skills (matching CLI behavior), pass "all" for every discovered skill, a list of names to enable only those, or [] to disable all.

for await (const message of query({
  prompt: "Help me process this PDF document",
  options: {
    cwd: "/path/to/project",
    settingSources: ["user", "project"],
    skills: "all",
    allowedTools: ["Read", "Write", "Bash"],
  },
})) { /* ... */ }

Enable specific skills by name (matching the name in SKILL.md or the directory name; use plugin:skill for plugin skills):

const options = { skills: ["pdf", "docx"] };

When you set skills, the SDK adds the Skill tool to allowedTools automatically. If you also pass an explicit tools list, include "Skill".

Tool access

The allowed-tools frontmatter in SKILL.md is honored only by the CLI, not the SDK. Control tool access through the main allowedTools option; without a canUseTool callback, anything not listed is denied. Pair with permissionMode: "dontAsk" to deny anything outside the allow list.

Troubleshooting

  • Skills not found: confirm settingSources includes user/project (an empty array disables discovery), and that cwd points at or below the directory containing .claude/skills/.
  • Skill not used: confirm its name is in the skills list (an empty list disables all) and that its description is specific.
  • Verify on disk with ls .claude/skills/*/SKILL.md.

Source

Source citations

Add this badge to your README

Show that Agent Skills in Claude Agent SDK Applications is listed on HeyClaude. Paste this Markdown into your README — it renders the badge and links back to this page.

Listed on HeyClaude
[![Listed on HeyClaude](https://heyclau.de/badge/guides/agent-skills-in-claude-agent-sdk-applications.svg)](https://heyclau.de/entry/guides/agent-skills-in-claude-agent-sdk-applications)

How it compares

Agent Skills in Claude Agent SDK Applications side by side with 3 alternatives on trust, install, platform support, and disclosed safety notes — all from reviewed registry metadata.

Field

A practical walkthrough of using Agent Skills in the Claude Agent SDK: how skills are discovered from the filesystem via settingSources, the skills option to enable or filter them, tool access, and troubleshooting discovery.

Open dossier

A practical walkthrough of loading local plugins in the Claude Agent SDK: the plugins option with type local and a path, what plugins bundle (skills, agents, hooks, MCP servers), verifying loads via the init message, and invoking namespaced plugin skills.

Open dossier

A practical walkthrough of subagents in the Claude Agent SDK: defining them programmatically with the agents option and AgentDefinition, context isolation, parallelization, tool restrictions, invocation, and resuming.

Open dossier

A practical walkthrough of token and spend accounting in the Claude Agent SDK. Read total_cost_usd from the result message, deduplicate parallel tool calls that share an assistant id, break spend down per model with modelUsage, sum cost across query() calls yourself, and read cache_creation/cache_read tokens.

Open dossier
Next steps
Trust
Review statusNot reviewedNot reviewedNot reviewedNot reviewed
Package trustPackage not verifiedPackage not verifiedPackage not verifiedPackage not verified
Source provenanceSource-backedSource-backedSource-backedSource-backed
SubmitterJPette1783JPette1783JPette1783JPette1783
Install riskReview firstReview firstReview firstReview first
Notes Safety ✓ Privacy ✓ Safety ✓ Privacy ✓ Safety ✓ Privacy ✓ Safety ✓ Privacy ✓
Brand
Categoryguidesguidesguidesguides
SourceSource-backedSource-backedSource-backedSource-backed
AuthorJPette1783JPette1783JPette1783JPette1783
Added2026-06-052026-06-052026-06-052026-06-05
Platforms
Harness
Source repo
Safety notesThe skills option is a context filter, not a sandbox: unlisted skills are hidden from the model but their files remain on disk and are reachable via Read and Bash. Skills are model-invoked; pair them with a tight allowedTools list (and dontAsk where appropriate) so an invoked skill cannot use more tools than intended. The allowed-tools frontmatter in SKILL.md does not apply through the SDK; control tool access with the main allowedTools option.The SDK only accepts type: local plugins; to use a marketplace plugin, download it first and point at the local directory. Plugins can bundle hooks, MCP servers, and agents that run with your application's privileges; review a plugin's contents before loading it. Check the system init message's plugin_errors to fail fast in CI when a plugin or a dependency version did not load.Restrict each subagent's tools via the tools field (for example read-only Read, Grep, Glob) so it cannot take unintended actions. When the parent uses bypassPermissions, acceptEdits, or auto, subagents inherit that mode and it cannot be overridden per subagent; subagents may be less constrained, so inheriting bypassPermissions grants full autonomous access. Subagents cannot spawn their own subagents; do not include Agent in a subagent's tools array.total_cost_usd / costUSD are client-side estimates from a bundled price table, not authoritative billing; do not bill end users or trigger financial decisions from them. Estimates can drift when pricing changes or the SDK version does not recognize a model; use the Usage and Cost API or Console for real billing. Both success and error result messages include usage and cost; read cost regardless of subtype so failed runs are still accounted for.
Privacy notesSkill descriptions are loaded so the model can decide when to use them; keep sensitive workflow detail and secrets out of descriptions. Skills sourced from outside your project run their instructions in your sessions; review them before enabling. Skill content is sent to the model provider when a skill is invoked; treat it like any other prompt content.Bundled MCP servers and hooks can access local files and external services; confirm what a plugin's components reach before deploying it. Plugin skills and commands are sent to the model as available capabilities; keep secrets out of plugin definitions. Loading a plugin from a path you do not control is a supply-chain decision; prefer plugins you have reviewed and pinned.A subagent's context starts fresh; the only channel from parent is the Agent tool prompt, so pass only the context it needs and avoid leaking unrelated data. The parent receives the subagent's final message; intermediate tool results stay in the subagent, which helps keep sensitive detail out of the main conversation. Subagent transcripts persist within the session and are cleaned up per cleanupPeriodDays.Usage data is token counts and cost, not content; it is safe to log, though it can reveal activity volume. Per-model and end-user attribution may be sent to an observability backend if you also enable telemetry; govern that data accordingly. The SDK uses prompt caching automatically; cache token fields reveal reuse patterns but not content.
Prerequisites
  • The Claude Agent SDK installed for Python or TypeScript.
  • SKILL.md files in .claude/skills/ (project) or ~/.claude/skills/ (user).
  • A cwd that points at or below the directory containing .claude/skills/, within the same repository.
  • The Claude Agent SDK installed for Python or TypeScript.
  • One or more plugin directories on the local filesystem (the parent of skills/, agents/, hooks/, or .claude-plugin/).
  • For marketplace plugins, the plugin downloaded locally so you can provide a path.
  • The Claude Agent SDK installed for Python or TypeScript.
  • Include Agent in allowedTools so subagent invocations are auto-approved.
  • Clear descriptions per subagent so Claude can match tasks to the right one.
  • The Claude Agent SDK installed for Python or TypeScript.
  • An async loop over query() results so you can read assistant and result messages.
  • For authoritative billing, access to the Usage and Cost API or the Console.
Install
Config
Citations
ClaimUnclaimedUnclaimedUnclaimedUnclaimed
Open 4 picks in the interactive comparison tool

Related guides

Signals

Loading live community signals…

More like this, weekly

A short, calm digest of reviewed Claude resources. Unsubscribe any time.