A2X

Agent (Server)

Agent Skills (Claude open standard)

A2X supports the open Claude Agent Skills format. A "skill" is a directory with a SKILL.md describing when to use it, plus optional reference files and scripts. When you point the SDK at a skill root (or hand it inline skills), the agent automatically:

  • Injects each skill's name + description into the system prompt.
  • Registers three builtin tools — load_skill, read_skill_file, run_skill_script — that the LLM uses to pull the body, read bundled files, and run scripts.
  • Applies variable substitution (${A2X_SKILL_DIR}, ${A2X_SESSION_ID}, $ARGUMENTS, $0, $1, …, plus the ${CLAUDE_*} aliases for Anthropic skill compatibility).

The same skill directory behaves identically under the Anthropic, OpenAI, and Google providers.

One-line activation

import { LlmAgent } from '@a2x/sdk';
import { AnthropicProvider } from '@a2x/sdk/anthropic';
 
const agent = new LlmAgent({
  name: 'my-agent',
  provider: new AnthropicProvider({ model: 'claude-sonnet-4', apiKey: process.env.ANTHROPIC_API_KEY! }),
  instruction: 'You are an assistant with access to the skills below.',
  skills: { root: '/absolute/path/to/skills' },
});

skills.root must be an absolute path. Use __dirname / import.meta.url to derive it so the code stays bundler-neutral:

import * as path from 'node:path';
import { fileURLToPath } from 'node:url';
 
const here = path.dirname(fileURLToPath(import.meta.url));
const skillsRoot = path.resolve(here, '../skills');

Inline skills

Use defineSkill when you cannot ship a directory of .md files at runtime (serverless, Next.js bundling, edge runtimes):

import { defineSkill, LlmAgent } from '@a2x/sdk';
 
const commitHelper = defineSkill({
  name: 'commit-helper',
  description: 'Generate concise commit messages from staged changes.',
  body: [
    '# commit-helper',
    '',
    '1. Read the staged diff.',
    '2. Produce a conventional commit message.',
  ].join('\n'),
  resources: {
    'FORMAT.md': 'Use type(scope): subject.',
  },
  scripts: {
    // Function-backed scripts run in-process — no child process is spawned.
    'scripts/preview.js': async ({ arguments: argv }) => ({
      stdout: `preview for ${argv.join(' ')}`,
      exitCode: 0,
    }),
  },
});
 
const agent = new LlmAgent({
  name: 'my-agent',
  provider: /* any provider */,
  instruction: 'x',
  skills: { inline: [commitHelper] },
});

Inline skills can coexist with a filesystem root.

Script execution policy

Scripts run from the skill's scripts/ directory. Three modes are supported:

Mode Behaviour
allow Default. Scripts execute immediately; the audit hook runs for logging.
confirm Scripts only execute if onScriptExecute returns a truthy value.
deny Script execution is blocked; the audit hook still fires for auditing. 
skills: {
  root: skillsRoot,
  scriptMode: 'confirm',
  onScriptExecute: async (meta) => {
    console.warn('[audit]', meta.skillName, meta.scriptRelativePath, meta.arguments);
    return await approveWithOperator(meta);
  },
},

The hook can also throw — doing so aborts the execution regardless of the mode, so you can use it as a uniform kill-switch.

Frontmatter reference

The SDK accepts every frontmatter field used by the open Claude Agent Skills format. Required fields are name and description; everything else is optional. Fields the runtime does not act on (e.g. model, effort, paths, context, agent) are preserved on AgentSkillMetadata.unknownFields for audit purposes.

  • See the Claude Agent Skills open standard for the file format.
  • load_skill, read_skill_file, and run_skill_script are ordinary tools; they flow through the same provider converters as any FunctionTool you register yourself.