Custom ChatGPT Plugins (MCP Integration)
Your systems, inside ChatGPT.
A custom ChatGPT plugin is an MCP server. We build yours over the API you already run, over a legacy system that has no API at all, or as a new backend that exists to do one job — and the same server works in Claude and Cursor without a second build.
What you get
- A ChatGPT plugin your team installs and uses
- One MCP server, also usable in Claude and Cursor
- Over your existing API, a legacy system or a new backend
- Interactive tables and forms in the chat, not text
- You own the server, the widgets and the specs
Start a conversation
Request a callback
Leave your details and a named senior engineer gets back to you within one business day. No sales script, no obligation.
Thanks — we have your details
A named senior engineer gets back to you within one business day.
That did not go through
Please email us directly at info@techthos.net and we will pick it up.
What this is
Your people already work in ChatGPT. What they cannot do there is touch your business: ChatGPT can discuss your orders, invoices, or stock levels in the abstract, but it cannot look one up, and it must not change one blindly. A custom plugin closes that gap — without touching what runs in production today.
Underneath, that plugin is a Model Context Protocol server. We put one in front of your existing REST API, GraphQL endpoint, database, or legacy application, expose a deliberate set of tools, and — where it earns its place — give those tools a real interface. Because MCP is the same standard Claude and Cursor speak, the plugin you commission for ChatGPT is not a ChatGPT-only asset.
What a ChatGPT plugin actually is in 2026
It is an MCP server with a listing. That is the whole answer, and it is why we sell plugin work and MCP integration as one service rather than two.
The 2023 plugin platform — the ai-plugin.json manifest and the OpenAPI spec behind it — is gone; OpenAI retired it in 2024. What replaced it arrived in stages. In March 2025 OpenAI added MCP support across the Agents SDK, the Responses API, and the ChatGPT desktop app. In October 2025 the Apps SDK shipped in preview, built on MCP, together with a developer mode that gives ChatGPT full MCP client support for read and write tools. In December 2025 the App Directory opened. On 9 July 2026 OpenAI migrated that directory to the Plugin Directory, where a single plugin listing can bundle apps, skills, and app templates, and appears across ChatGPT web, ChatGPT desktop, ChatGPT Work, and Codex.
So there is no longer a fork in the road. Commissioning “a ChatGPT plugin” and commissioning “an MCP integration” produce the same artifact: a server that speaks MCP, deployed remotely, packaged for the directory if you want a public listing and kept private if you do not.
Four ways we build it
Which shape fits depends on what already exists on your side. Discovery ends with that decision written down.
Wrapper server
A server in front of what you already expose.
- Speaks to your existing REST, GraphQL or SQL surface
- Zero changes to the running application
- Fastest path to a working plugin
- Tool surface designed, not auto-generated from endpoints
Legacy bridge
For systems that never had a usable API.
- SOAP endpoints, on-premise databases, file drops
- Desktop and terminal applications driven directly
- We build the missing interface, then the tools on top
- The legacy system stays untouched and unrewritten
Embedded mode
MCP as a second mode of your own binary.
- The application itself answers as an MCP server
- One codebase, one deployment, no translation layer
- The binzaar pattern: one Go binary, three modes
- Best when we also own or maintain the application
Purpose-built backend
A new backend that exists to do one job.
- No existing system to wrap — we build the whole thing
- Scoped to a single use case, not a general platform
- Its own storage, its own auth, its own deployment
- The smallest surface that answers the question well
Interfaces in the chat, not walls of text
A tool that returns JSON forces the assistant to describe your data in prose. A tool that returns an interface lets your people act on it. When the assistant lists open invoices, it can hand back a sortable table with per-row actions; when it needs structured input, a validated form the user submits straight back as a tool call.
This is a standard, not a trick that breaks on the next host update. On 26 January 2026, Anthropic and OpenAI standardised interactive UI as MCP Apps (SEP-1865) — the first official extension to the Model Context Protocol, carried by the io.modelcontextprotocol/ui namespace. A widget built once renders in any compliant host: ChatGPT, Claude, VS Code, Cursor, Postman, or the private workspace we build for you.
Our open-source library gomukit is how we ship those interfaces. It generates table, form, card, and card-list widgets as self-contained HTML resources embedded in the server binary — themed to the host, formatted to the host’s locale and timezone, locked down by a strict Content Security Policy, with no external CDN dependency at all.
Proof: the MCP servers we ship in the open
We point at working software. Every pattern on this page is one we run in our own repositories, under the MIT licence, in public.
binzaar
Open sourceA single-binary local app store: browse a GitHub-hosted catalog, install release binaries with SHA-256 verification, update, and scaffold new microapps. One binary runs as a terminal UI, an MCP stdio server, and an HTTP catalog server — the embedded-mode pattern described above, including in-chat widgets.
gomukit
Open source · pre-releaseThe Go library behind our in-chat interfaces: table, form, card and card-list widgets emitted as self-contained HTML, SDK-agnostic at the core with an adapter for the official MCP go-sdk. Compliant with the io.modelcontextprotocol/ui extension, revision 2026-01-26.
clockwork
Open sourceOur flagship dual-mode time tracker: an MCP server that lets an assistant log work in natural language and turn git commits into worklogs, plus a keyboard-driven terminal UI over the same data.
The standards we build to
Building on the published specification is what keeps your plugin portable. A bespoke assistant integration is tied to one vendor’s host; an MCP server is not.
Security and what the assistant is actually allowed to do
The reason most internal systems are still not reachable from ChatGPT is not capability, it is permissions. We treat the tool surface as an authorization boundary, not a prompt.
How we engage & pricing
Per-milestone and stop-anywhere, billed by effort at €85/hour. What a plugin costs depends on its complexity — how many tools, how clean the existing interface is, whether write operations and interactive widgets are in scope — so we scope it after Discovery rather than guessing at a fixed price.
- DiscoveryFree
- Plugin specBy effort · €85/h
- Pilot pluginBy effort · €85/h
- Full buildBy effort · €85/h
- Maintenancefrom €70 / mo
Billed by effort at €85/hour; the total follows complexity and is fixed per SOW after the free Discovery. You own the server source, the widget definitions, the specifications and the deployment. Stop after any milestone — the tool map alone is a usable artifact.
