Skip to content

Latest commit

 

History

History
105 lines (75 loc) · 6.56 KB

File metadata and controls

105 lines (75 loc) · 6.56 KB

Plugins: Claude, Cursor, GPT, VS Code — Differences and choices

Version: 0.4
Date: 2026-06-28
Platform: 0.4.14 · MCP catalog via GET /mcp/actions (live discovery; see capability matrix in repo docs) Context: AgentStack plugins for Claude Code, Cursor, GPT (OpenAI), and VS Code; alignment and philosophy. Plugin index: docs/plugins/README.md.


Where they live

Plugin source on GitHub: cursor-plugin, claude-plugin, gpt-plugin, vscode-plugin.

  • Cursor: cursor-plugin
  • Claude: claude-plugin
  • GPT (OpenAI): gpt-plugin
  • VS Code: vscode-plugin

One plugin — one artifact (Decomposition). Shared MCP endpoint and ecosystem; manifests and configs differ by platform.


GPT (OpenAI) — GPT Actions

  • Integration model: not a "plugin package" but artifacts for GPT Actions. The user creates a Custom GPT in ChatGPT and attaches the OpenAPI 3.1 schema + instructions.
  • Manifest: OpenAPI 3.1 schema (provided_plugins/gpt-plugin/openapi/agentstack-mcp.yaml) + reference instructions (GPT_INSTRUCTIONS.md). Install = create Custom GPT per GPT_QUICKSTART.md.
  • MCP: same endpoint https://agentstack.tech/mcp; auth — API Key in header X-API-Key (set in Custom GPT → Action → Authentication).
  • OAuth (optional): Custom GPT can use OAuth2 (Authorization Code) with AgentStack as IdP:
    • Authorization URL: https://agentstack.tech/api/oauth2/authorize
    • Token URL: https://agentstack.tech/api/oauth2/token
  • More: artifact layout in provided_plugins/gpt-plugin/ARTIFACTS.md; quick start in provided_plugins/gpt-plugin/GPT_QUICKSTART.md.

Component comparison

Component Cursor Claude Code GPT (OpenAI) VS Code
Manifest .cursor-plugin/plugin.json .claude-plugin/plugin.json OpenAPI 3.1 schema + GPT_INSTRUCTIONS.md package.json + contributes.mcpServerDefinitionProviders
Install Copy plugin + MCP config Install plugin + claude mcp add Create Custom GPT, paste schema and instructions Marketplace/VSIX + one-time API key entry
MCP config mcp.json (HTTP: type, baseUrl, headers) HTTP via user setup (see below) API Key or OAuth in Action settings HTTP via extension (auto-registration)
Skills skills/*/SKILL.md24 domain routers (incl. CRM, AgentNet, storefront, wallet, guidance) 11 synced skills + uplift path for messenger/rag/sdk gaps No equivalent; context in Custom GPT instructions No equivalent; context in MCP and README
Rules rules/*.mdc (Cursor-specific) No equivalent; knowledge in Skills + doc links No equivalent No equivalent

MCP: HTTP in Claude Code

  • In the Claude Code plugin (.mcp.json) official docs only describe running MCP via command (stdio). For remote HTTP servers the user configures MCP manually.

  • Recommended flow for AgentStack: after installing the plugin the user runs once:

    claude mcp add --transport http agentstack https://agentstack.tech/mcp --header "X-API-Key: <YOUR_API_KEY>"

    Or configures MCP in Claude Code UI (if available) per MCP_QUICKSTART.md.

  • Summary: we do not add .mcp.json with HTTP in provided_plugins/claude-plugin/ (HTTP format in plugin bundle is not fixed); all MCP connection steps are described in MCP_QUICKSTART.md and README (Elegant Minimalism: minimal steps, one API key).

Claude and OAuth

  • Current recommended path for Claude Code is API key (X-API-Key) for MCP HTTP.
  • If the Claude platform/MCP catalog requires OAuth, AgentStack already provides standard OAuth2 endpoints (/api/oauth2/authorize, /api/oauth2/token) without a separate "for Claude" implementation.

Cursor gen3 domain parity (2026-06)

Domain Cursor skill Claude (sync target)
Meta-router agentstack-backend Same pattern
Data / 8DNA agentstack-data agentstack-data
CRM agentstack-crm agentstack-crm (stub)
AgentNet economy agentstack-agentnet agentstack-agentnet (stub)
Storefront studio agentstack-storefront-studio agentstack-storefront-studio (stub)
Project wallet agentstack-project-wallet agentstack-project-wallet (stub)
Platform guidance agentstack-guidance agentstack-guidance (stub)
Hosting agentstack-hosting Add or merge in claude-plugin
Support agentstack-support Add or merge
Storage agentstack-storage Add or merge
Auth/RBAC agentstack-auth-rbac Existing claude skills
Logic / Commerce / RAG / Signals / Projects matching agentstack-* Partial parity today
Messenger / Integrations / Discovery gen3 only on Cursor Optional claude sync (11 WARN gaps without --strict)

Hooks and Device Code install are Cursor-only. Claude uses claude mcp add per MCP_QUICKSTART.


What is reused

  • Skills text and structure — copy from provided_plugins/cursor-plugin/skills/ (gen3 decision-first names: agentstack-data, agentstack-logic, …); replace "Cursor" with "Claude Code" in instructions.
  • GPT: same MCP endpoint and API key; key acquisition text reused from MCP_QUICKSTART; Custom GPT instructions reference MCP_CAPABILITY_MATRIX.
  • Production MCP URL: https://agentstack.tech/mcp.
  • Documentation: links to MCP_CAPABILITY_MATRIX, 8DNA, ecosystem without duplication.

Skills: syncing Cursor and Claude

  • Source of truth: provided_plugins/cursor-plugin/skills/ (gen3). Copy to provided_plugins/claude-plugin/skills/ on release; retire gen1 folder names (agentstack-8dna, etc.).
  • Claude adaptation: in copied SKILL.md replace "Cursor" with "Claude Code" in the body (e.g. "add MCP in Cursor" → "add MCP in Claude Code"). Frontmatter (name, description) unchanged.
  • Links in Claude version: References to MCP_QUICKSTART and README point to artifacts in claude-plugin root (MCP_QUICKSTART.md, README.md — same plugin). Repo links (MCP_CAPABILITY_MATRIX, philosophy) stay shared.
  • Versioning: when changing skills, update CHANGELOG in both plugins (Time Processes Philosophy). See also SKILLS_AUTHORING_GUIDE.md.

Genetic code (reference): docs.plugins.claude_vs_cursor_plugin.gen2