capability-intake

name	capability-intake
description	Security-gated intake workflow for external plugins, skills, MCP servers, and agent tools.

Capability Intake

Description

Run the deny-by-default review workflow for external capabilities before agents enable plugins, skills, MCP servers, connectors, or CLIs.

Tags

plugins, skills, tools, mcp, security, admission, supply-chain, allowlist, import

When to Use

Use this skill before adding, enabling, updating, or delegating review of any external plugin, skill, MCP server, agent connector, or CLI capability.

Use it for:

• Comparing candidate tools before installation.
• Auditing MCP/tool metadata and declared permissions.
• Producing reviewer-ready admission reports.
• Fan-out delegation for deeper source, security, and integration review.

Workflow

1. List candidates:

   scripts/ai/aq-capability-intake list
   

2. Audit all candidates:

   scripts/ai/aq-capability-intake audit --all --json
   

3. For a specific candidate:

   scripts/ai/aq-capability-intake audit <candidate-id> --json
   

4. Do not enable a candidate unless its report is low-risk or a reviewer explicitly accepts needs-review with mitigation.
5. If a candidate has network, secret, write, shell, package-manager, or browser permissions, require a follow-up slice with:
   • pinned version or digest
   • tool allowlist
   • SBOM/dependency scan
   • sandbox/permission boundary
   • dashboard or aq-report visibility
   • rollback path

Usage

Start with scripts/ai/aq-capability-intake audit <candidate-id> --json, review admission status and mitigations, and only promote candidates that are low-risk or accepted-with-mitigations.

Files

• Candidate registry: config/agent-capability-intake-candidates.json
• CLI: scripts/ai/aq-capability-intake
• PRD: .agent/PROJECT-CAPABILITY-INTAKE-PRD.md
• Test: scripts/testing/test-capability-intake.py

Delegation Rule

When delegating review, pass only candidate id and this skill name. The receiving agent should load this skill, run the audit CLI, inspect upstream source, then report PASS/FAIL/REQUEST_REVISION with evidence.