Skip to content

TESTING_AND_CAPABILITIES.md

Latest commit

 

History

History
127 lines (93 loc) · 6.05 KB

TESTING_AND_CAPABILITIES.md

File metadata and controls

127 lines (93 loc) · 6.05 KB
Raw

AgentStack plugin (Claude Code) — testing and capabilities

How to verify the plugin works

1. Install the plugin

Option A: from marketplace (after publish)

  • Claude Code → Plugin manager / Discover → search "AgentStack" → Install.

Option B: locally (before publish)

  • Run with plugin directory: 
    claude --plugin-dir ./provided_plugins/claude-plugin
    Or copy the plugin to the directory Claude Code loads plugins from (see Claude Code — Plugins).

2. MCP connection (required for calling tools)

The plugin provides Skills, but calls to projects, logic, buffs, etc. go through MCP. Without MCP configured, tools will not be called.

  1. Preferred: OAuth Device Code
    Run /agentstack:login, approve in the browser, then add MCP with the returned Bearer token:

    claude mcp add agentstack --transport http https://agentstack.tech/mcp \
  --header "Authorization: Bearer <access_token>" \
  --header "Content-Type: application/json"

  2. Fallback: API key
    Create an anonymous project/API key and configure:

    claude mcp add agentstack --transport http https://agentstack.tech/mcp \
  --header "X-API-Key: YOUR_KEY" \
  --header "Content-Type: application/json"

  3. Verify:

    claude mcp list
/mcp

  4. Reload plugins after local edits:

    /reload-plugins

Details: MCP_QUICKSTART.md.

3. Testing in chat

In Claude Code chat ask the agent:

  • "Create a project in AgentStack named Test Project"
    → Expected: call to projects.create_project_anonymous (or projects.create_project when authenticated).
  • "Show my AgentStack projects"
    → Expected: projects.get_projects.
  • "Get stats for project <project_id>"
    → Expected: projects.get_stats.

If the agent calls MCP tools and returns a sensible answer — the plugin and MCP are working.

4. Checking Skills

  • Skills are picked up by Claude Code and available under the plugin namespace (e.g. /agentstack:agentstack-8dna, /agentstack:agentstack-projects, /agentstack:agentstack-rules-engine, /agentstack:agentstack-payments, /agentstack:agentstack-auth). Run /help and /reload-plugins after local edits. The plugin ships 8 skills.

5. Common issues

Symptom What to check
Agent does not call MCP MCP added via claude mcp add, correct URL and Authorization or X-API-Key header, claude mcp list shows agentstack.
401 / 403 on call Key is valid, not expired; some operations require a subscription (e.g. Professional for add_user).
"Tool not found" Tool name matches docs/MCP_CAPABILITY_MATRIX.md or GET https://agentstack.tech/mcp/actions.
Skills not visible Plugin is loaded (--plugin-dir or installed from marketplace). Run /reload-plugins; check /help and namespace (e.g. /agentstack:agentstack-8dna).

Plugin capabilities

What the plugin includes

Component Purpose
Manifest (.claude-plugin/plugin.json) Name, description, keywords for Claude Code and marketplace.
Skills (8) Teach the agent when and how to use AgentStack: 8DNA, projects, Rules Engine, auth, RBAC, payments, buffs, assets.
Commands (commands/agentstack-login.md) Device Code login and Claude MCP setup instructions.
Documentation README, MCP_QUICKSTART, this file.

Capabilities via MCP (after MCP setup)

The plugin does not call the backend itself — the AgentStack MCP server does. After adding MCP in Claude Code the agent gets access to tools such as:

  • Projects: create (including anonymous), list, details, update, delete, stats, users, settings, activity, API keys, attach anonymous project to user.
  • Logic and rules: create/update/delete rules, list, execute, processors, commands.
  • Buffs: create, apply, extend, rollback, cancel, list active, effective limits, temporary and persistent effects.
  • Payments: create, get, refund, list transactions, balance.
  • Auth: login, register, assign role, profile.
  • Scheduler: create/cancel/get/list tasks, etc.
  • Analytics: usage, metrics.
  • API keys: create, list, delete, etc.
  • Webhooks, notifications, wallets — as implemented on backend and in MCP.

Full tool list and parameters: generated MCP_CAPABILITY_MATRIX in the AgentStack repo or GET https://agentstack.tech/mcp/actions.

Skills capabilities

  • agentstack-8dna: design and query data with hierarchy (parent_uuid) and evolution (generation), work with data/config/protected structure and genetic coding.
  • agentstack-projects: create and manage projects and API keys via MCP, anonymous projects, attach to user.
  • agentstack-rules-engine: configure server logic without code (when/do), use Logic Engine and rules via MCP, link with buffs and commands.
  • agentstack-auth: login, register, profile, session.
  • agentstack-rbac: roles, permissions, membership.
  • agentstack-payments: payments, refunds, balance, transactions.
  • agentstack-buffs: trials, subscriptions, temporary/persistent effects.
  • agentstack-assets: assets, inventory, digital goods.

Summary

  • Testing: install plugin → run /reload-plugins → configure MCP via Device Code or API key → verify claude mcp list and /mcp → ask to create/list projects and verify MCP calls.
  • Capabilities: access to the live AgentStack MCP action catalog (projects, logic, buffs, payments, auth, scheduler, analytics, agents, storage, support, etc.), plus eight Skills for consistent use of 8DNA, projects, Rules Engine, auth, RBAC, payments, buffs, and assets.

Latest Local Smoke Snapshot

2026-05-11:

  • node provided_plugins/scripts/validate-all-plugins.mjs — passed with 3 warnings for Cursor placeholder screenshots.
  • Claude manifest and eight-skill count were checked by the shared validator.
  • Manual runtime checks still require Claude Code: /reload-plugins, /agentstack:login, claude mcp list, and /mcp.