Prodhub Agent Kit MCP Server

ProdHub Agent Kit

Give any MCP-compatible coding harness a safe, repeatable path from source code to a ProdHub Preview Space.

This kit does not contain credentials. ProdHub uses browser-based OAuth, scopes each session to the approved organization, and applies the same role controls as the product UI.

Install

The canonical remote MCP endpoint is:

https://app.prodhub.in/mcp

ProdHub is described for machine discovery in server.json and published under io.github.SwaDeshiTech/prodhub in the official MCP Registry. MCP clients and registry-backed catalogs can use that entry to discover the hosted Streamable HTTP endpoint without relying on a general web search.

Codex

codex mcp add prodhub --url https://app.prodhub.in/mcp

Merge the delivery instructions in AGENTS.md into the repository where Codex should use ProdHub; preserve existing project instructions. Copy the supporting docs too, or retain links to this kit.

Gemini CLI

gemini mcp add prodhub https://app.prodhub.in/mcp --transport http --scope user

Copy GEMINI.md into that repository or merge its instructions with an existing project file.

Claude Code and other MCP clients

Use the remote-server entry in mcp.json, then copy or merge CLAUDE.md into the project guidance file used by your client.

ChatGPT and Codex plugin

This repository is also a portable Agent Plugins package:

• plugin.json contains the public listing, onboarding, review, and publication metadata.
• mcp.json declares the remote ProdHub MCP server and OAuth with dynamic client registration.
• skills/get-started/SKILL.md supplies the post-install setup and safe-delivery workflow.
• assets/prodhub-icon.svg is the square directory and composer icon.

Generate the ZIP to upload in the OpenAI Plugins submission portal:

./scripts/package-chatgpt-plugin.sh

Before final submission, record a reviewer-accessible walkthrough and add its URL to extensions.com.openai.review.demo_recording_url in plugin.json. Enter the dedicated test-account credentials only in the secure review form; never commit them to this repository.

Follow the complete ChatGPT submission checklist for the review fixture, domain verification, upload, and publication steps.

First connection

1. Add the MCP endpoint in the client.
2. Start the OAuth connection. Sign in to ProdHub and choose an organization.
3. Approve the consent screen.
4. Ask the coding harness to deploy a branch or create a Preview Space.

The agent should begin with get_prodhub_delivery_capabilities, call list_available_product_areas, inspect the repository and the prodhub_build contract, show a plan, obtain approval, and then create resources. The product-area response is filtered by the same role-based tab configuration used by the ProdHub UI and does not require organization selection.

Read the full workflow in docs/preview-delivery.md.

Understand the product before deploying

Start with product concepts to distinguish applications, profiles, provider connections, Preview Spaces, and pipeline executions. Then follow the end-to-end Preview Space runbook. It covers new applications, existing profiles, credential sync, branch/revision verification, planning, explicit profile selection, provisioning, monitoring, retrying in the same environment, and lifecycle handoff.

Use profile field definitions while inspecting source, and failure recovery when delivery fails. These guides describe the current implementation; the connected server's tools/list is authoritative for tool availability.

Example requests

• “Deploy this repository's feature/login branch to a new Preview Space for one day.”
• “Reuse the existing catalog application's profiles and give me a verified preview URL.”
• “Check why this Preview Space failed, fix the configuration, and retry in the same environment.”

The agent should retain the application, profile, environment, and execution IDs returned by tools. Users should not have to discover these IDs manually. Local uncommitted changes are not included in a remote build.

Important boundaries

plan_application_preview plans delivery for an existing application with saved profiles. New application onboarding needs its own proposal before those records are created. The planner does not validate the source code or reserve resources.

A url_available response means a URL exists, not necessarily that the application has finished deploying. Inspect the current execution and verify the application response before reporting success. Early environment deletion and generic profile updates currently need a supported product flow; do not invent MCP tools for them.

mcp.json is a configuration example. Clients have different file names and schemas; cloning this repository alone does not register the MCP server.

Files

• server.json: official MCP Registry metadata for the hosted server.
• mcp.json: portable remote-MCP configuration.
• plugin.json: ChatGPT and Codex public plugin metadata.
• skills/get-started/SKILL.md: plugin onboarding workflow.
• docs/chatgpt-submission.md: public directory submission checklist.
• AGENTS.md: Codex-oriented project instructions.
• CLAUDE.md: Claude Code project instructions.
• GEMINI.md: Gemini CLI project instructions.
• docs/preview-delivery.md: safe delivery workflow.

For the browser setup page, see prodhub.in/setup.

Publishing discovery metadata

The Publish MCP Registry metadata GitHub workflow validates and publishes server.json using GitHub OIDC. No registry token or long-lived secret is required. Increment the manifest version whenever the public MCP identity or tool contract changes; merging that update to main republishes the entry.