Build in-house, hire a studio or use an MCP platform: who should build your connector
Should you build your MCP connector in-house, use an MCP platform or hire a studio?
Build in-house if your team can own tool design, OAuth, evals and each directory's review over time. Use an MCP platform if a server generated from your API is enough and directory listings matter less. Hire a studio such as Invokary for a connector built to each directory's published rules, on a fixed price and timeline.
- Published
Which to choose
Build in-house
Choose it when:
- Your engineers can own the server after launch: spec revisions, changes to directory rules and incidents.
- You already run OAuth for third-party apps.
- The connector is core to your product, and you want the expertise inside the company.
Use an MCP platform
Choose it when:
- You need a working server soon, and your API maps cleanly to a few actions.
- Access from clients your users configure themselves matters more than directory listings.
- Tools that mirror your endpoints, rather than your users' jobs, are acceptable for now.
Hire a studio
Choose it when:
- You want a connector built to each directory's published review rules, at a fixed price.
- Your team's time is committed to the core product.
- You want the server in your repository, with a runbook, so your team can own it later.
The three options differ less in what gets built than in what your team owns afterwards. An MCP server is not a one-time project: the protocol has revisions, each directory changes its rules, and a published connector needs someone to watch its errors and its listing. Choose by who should own that work.
What every option has to get right
Whoever builds it, the connector has to pass the same review. The directories publish their rules: tool titles and hints, short names, descriptions that don't instruct the model, separate read and write tools, working reviewer credentials, and on ChatGPT, exactly five positive and three negative test cases. Our guide to why submissions get rejected lists them with their sources, and the Claude Connectors Directory page has every requirement with the date we checked it.
Build in-house
You keep the knowledge and the control. The cost is attention: someone on your team learns the MCP spec, OAuth with Client ID Metadata Documents, tool design and each directory's submission process, then keeps up as they change. It fits companies where the connector is part of the product, not a channel next to it.
Use an MCP platform
Hosted MCP platforms and OpenAPI-to-MCP generators produce a server from your API quickly. What they produce mirrors your endpoints. Anthropic's guidance on writing tools for agents recommends the opposite: a few tools shaped around workflows, not one tool per endpoint. A generated server can be a good start where the API is small and the goal is access from clients your users set up themselves. Check whether the server, the domain and the OAuth setup are yours, and what the platform does when a directory's rules change.
Hire a studio
A studio builds the connector to the published rules on a fixed price and timeline, then hands it over. In a Connector Launch Sprint that means a remote server in your repository and on your domain, OAuth with a DCR fallback, 8 to 20 annotated tools, evals in CI, the submissions for the platforms you choose, and a runbook with a handover call. You pay for focus and for not learning each directory's process from scratch. You still own the result, and someone on your side should be ready to take it over.
Not sure which fits? The Agent Distribution Audit maps your API to tools and gives a fixed quote, which also tells you what building it in-house would involve.
Questions
How long does building an MCP connector take?
A Connector Launch Sprint is 3 to 4 weeks of build, plus each platform's review, which the platform sets. In-house timelines depend on your team, and we don't estimate other providers' timelines.
Does a studio guarantee a listing?
No one can. Each platform decides whether and where to list a connector. We build to the published rules, and a Sprint includes two rejection cycles.
Can we start on a platform and move in-house later?
It is easier if the domain and the OAuth setup are yours from the start. A Claude connector is submitted as a server URL, so keeping that URL on your own domain keeps your options open.
What does each option cost?
Our Audit is $750 and a Sprint starts at $9,500, both fixed in writing. We don't publish other providers' prices, and in-house cost is your team's time.
Sources
- Model Context Protocol: Architecture overview (accessed )
- Anthropic Engineering, September 11, 2025: Writing effective tools for AI agents (accessed )
- Claude docs: Publish to the directory (accessed )
- OpenAI Developers: Submit plugins (accessed )