MCP, ChatGPT Actions & IDE agents
Toolcore publishes stable HTTPS JSON and plain-text prompts so different client types can integrate without scraping HTML. This page compares MCP, Custom GPT / Actions-style HTTP, and browser URL handoff.
Client type × typical wiring
| Client / pattern | How Toolcore fits | First URLs to fetch |
|---|---|---|
| Cursor / Claude Code (stdio MCP) | Run the repo's MCP server; it exposes manifest + URL builder tools. | bun run mcp — see /ai-agents |
| HTTP-capable agent (no MCP) | Poll JSON manifests; build tool links or call /api/agent/v1 when needed. | /agent-tools.json, /llm-prompt.txt |
| ChatGPT Custom GPT–style Actions | Point Actions at public GET endpoints; return links or fetch capabilities JSON for your action schema. | /api/agent/v1/capabilities, /agent-tools.json |
| Human + assistant in browser | Assistant returns prefilled tool URLs; user clicks and runs locally. | /ai-agents, /tools/cli |
Cross-reference: Integration matrix, Offload compute, Server API matrix.
Common use cases
- Document internally which base URL (production vs preview) your GPT Action or script should call.
- Train support staff to distinguish MCP (stdio) from plain HTTPS manifest integration.
Common mistakes to avoid
Hard-coding only ChatGPT or only MCP
The manifests are client-agnostic; the same agent-tools.json supports multiple front-ends once your wrapper fetches it.
FAQ
Does Toolcore host a public Streamable HTTP MCP endpoint?
Integration is manifest + HTTPS (and optional local stdio MCP in the repo). There is no requirement for a hosted MCP wire server on toolcore.dev for core flows.