Product integration
Ship the official MCP server for your product
A remote, OAuth-secured MCP server for your product's public API, so your customers use your SaaS inside Claude, ChatGPT and Cursor. Up to 15 curated tools, multi-tenant, listed in the registry under your brand. $4,500 to build in three weeks, $700/mo to maintain, including migration when the spec revises. The point is not endpoint coverage; it is a curated tool surface a model can actually choose from without guessing.
Who this is for, and who it is not
This is for a B2B SaaS with a public or partner API whose customers already live in an AI client. It is not for a product with no API yet, and it is not for an internal-only tool. If your users would never think to ask an assistant about your product, a server is premature, and we will say so.
Your competitor may already be in the client. Check.
The fastest way to know whether this matters yet: search the official registry and the main aggregators for your category. If a competitor is listed and you are not, their product is one sentence away inside the assistant and yours is a browser tab. If nobody in your category is listed, you have a first-mover window that closes the month someone takes it.
What gets built
- Up to 15 curated tools, named for what a user asks, not for your endpoints.
- OAuth 2.1 with PKCE, multi-tenant, per-customer authorisation.
- An audit log for every call, and dry-runs on every write.
- A server.json registry listing under your brand.
Eight tools beat eighty
How it fits together
The timeline
| Days | What happens |
|---|---|
| D0–D3 | Tool manifest written and signed |
| D4–D14 | Build against the manifest and acceptance script |
| D15–D18 | Automated checks, founder diff review |
| D19–D21 | Deploy, registry listing, handover |
What "done" means
Done is not our opinion. It is a ten-prompt acceptance script you approve up front, run against every client you support. If a prompt fails, it is not done.
1. "Find my three most recent invoices" -> invoices.search returns 3
2. "Void invoice INV-4021 (dry run)" -> dry_run response, no write
3. "Who is the billing contact for Acme?" -> customers.lookup
...Product integration questions
- How is this different from an auto-generated server?
- An auto-generator maps every endpoint to a tool, which produces a worse AI interface, not a better one. Thoughtworks placed naive API-to-MCP conversion in the Hold ring of its Technology Radar, and Cloudflare collapsed 2,500-plus endpoints behind two tools. We design a curated surface of up to fifteen tools with names and error messages a model can act on.
- What auth do you use?
- OAuth 2.1 with PKCE, multi-tenant, so each of your customers authorises against their own account. Client tokens are never passed downstream, and we validate the audience claim. The 2026-07-28 revision changed the authorization model and every server we ship negotiates the current revision.
- How long does it take?
- Three weeks from a signed manifest to a deployed server, with a ten-prompt acceptance script defining "done". The bottleneck is almost always your side: getting us API docs, a test tenant and sign-off on the tool list. We tell you the honest client-effort estimate in hours before we start.
- Do you list it in the registries?
- Yes, under your brand, in the official MCP registry via a server.json record, plus the main aggregators. Being listed is how a client discovers the server exists; it is not a ranking play. We prepare the metadata once and reuse it.
- What can it do to our data?
- By default, read only. Any write tool is scoped, dry-runnable and audited, and there are no destructive deletes ever. A server cannot fix a model that reads a poisoned input; it can bound the blast radius. The full posture is on the security page.
- What happens when the spec changes again?
- We migrate it on the retainer before the deprecation window closes. That is not hypothetical: the 2026-07-28 revision put Roots, Sampling and Logging on a twelve-month clock, and every server built earlier is already behind.