SkillManifest (draft)

Schema for agent skill packages at packages/skills/<id>/. Defines runtime compat, required scopes, policy template, consent summary, trust tier.

The OSS agent-skill manifest. Every package at packages/skills/<id>/ ships one of these as src/manifest.ts.

Canonical URL

https://glide.co/schemas/agent-banking/v1/skill-manifest.json

Required fields

Field	Type	Notes
schemaVersion	'v1'	Locked.
slug	string (lowercase kebab-case)	Package directory name.
displayName	string	Catalog display name.
tagline	string	One-sentence pitch.
longDescription	string	Multi-paragraph description rendered in the install consent flow.
category	'ap' | 'treasury' | 'consumer' | 'payroll' | 'x402'	Catalog filter.
runtimeCompat	Runtime[]	claude-desktop / chatgpt-apps / vertex / openclaw / hermes.
requiredScopes	Scope[]	Closed-vocab MCP scopes the skill calls.
policyTemplate	SkillPolicyTemplate	Per-skill envelope defaults. Strict subset of AgentPolicyEnvelope.
consentSummary	string	Plain-English copy surfaced at install. Must mention payments / money / cards / $ / usd if requiredScopes includes a money-touching scope.
trust	'community' | 'verified' | 'core'	Trust tier.
publisher	string	Maintainer org or individual.
iconPath	string (defaults ./icon.svg)	Relative path.
packageVersion	string	Semver.
blockingDependency	string | null	Free-text label. null = unblocked. Non-null = catalog renders “wait for V3” instead of “install now”.
homepageUrl	string (URL, optional)
changelogUrl	string (URL, optional)

Closed vocabularies

Adding to any of these requires @glideco/policy-engine to understand the same vocabulary — CODEOWNERS-protected change.

SkillScope

accounts:read
agents:read
audit:stream
beneficiary:write
cards:manage
payments:initiate
payments:simulate
treasury:yield-allocate
x402:pay
x402:receive

SkillRuntime

claude-desktop
chatgpt-apps
vertex
openclaw
hermes

SkillCategory

ap
treasury
consumer
payroll
x402

Consent under-disclosure guard

The SkillContractTestSuite enforces a soft guard: if requiredScopes includes any of payments:initiate, cards:manage, or x402:pay, the consentSummary MUST mention payments / money / cards / $ / usd. Soft guard against community-tier skills that under-disclose at the install consent step.

The contract test catches the obvious cases. Reviewers catch subtler ones during the verified-tier promotion review (e.g. consent copy that mentions money but lies about caps).

Example

import { SkillManifest } from '@glideco/skills-base';

export const manifest = SkillManifest.parse({
  schemaVersion: 'v1',
  slug: 'ap-agent-claude-quickbooks',
  displayName: 'AP Agent (Claude + QuickBooks)',
  tagline: 'Let Claude draft vendor payments from QBO invoices.',
  longDescription: '...',
  category: 'ap',
  runtimeCompat: ['claude-desktop'],
  requiredScopes: [
    'accounts:read',
    'payments:initiate',
    'payments:simulate',
    'beneficiary:write',
    'audit:stream',
  ],
  policyTemplate: {
    perTxMaxUsdCents: 250_000,
    dailyCapUsdCents: 1_000_000,
    counterpartyAllowlist: [],
    velocityCap: { windowMinutes: 60, maxCount: 20 },
  },
  consentSummary: 'Claude reads your QuickBooks Online invoices and drafts USD payments to your existing vendors, capped at $2,500 per transaction and $10,000 per day. Each payment requires your explicit approval before broadcast.',
  trust: 'core',
  publisher: 'Glide',
  packageVersion: '0.1.0',
  blockingDependency: 'V3 Bucket 3.3 (QBO integration) + 3.4 (bill pay)',
});

Reading list

• Authoring a skill — partner-PR flow.
• AgentPolicyEnvelope — the full envelope a skill template constrains.
• Skill catalog — every skill currently in the repo.