Custom MCP development
MCP servers agents actually use correctly.
We design, build, and evaluate custom MCP servers for SaaS products and enterprise systems: task-shaped tools, OAuth 2.1 built in, and eval scorecards. You own every line.
Choose a fixed price deliverable or longer-term partnership.
Escalate the Northwind outage ticket and tell Priya we’re on it.
- find_tickets(“Northwind outage”)
- escalate_ticket(TCK-4182, “P1”)
- reply_to_customer(TCK-4182, …)
“Hi Priya, we’ve escalated your outage to our on-call team. You’ll get an update within the hour.”
Done. TCK-4182 is now P1, and Priya has your reply.
The problem
Wrapping an API isn’t an MCP server.
Most MCP servers are thin wrappers, hand-built or generated, that work just well enough to demo. Here’s where they fall over:
01 / tool sprawl
Every endpoint became a tool
Forty-odd endpoint-shaped tools blow the context budget before the user says a word, and the agent picks the wrong one.
02 / auth bolted on
One API key for everyone
Shared credentials, no per-user scopes, and no audit trail. Fine for a demo, and a non-starter for your security review.
03 / inspector-green
Works in the Inspector, fails in Claude
It lists tools and answers a test call, then stalls on real tasks in real clients, where nobody's measuring.
Tool design
Fewer, task-shaped tools. Measurably better agents.
Agents don’t read your API docs. They read your tool list every turn, and they have to choose. We redesign around the jobs people actually ask for, then prove the difference with evals.
before · 47 tools mirroring REST endpoints
- get_tickets
- get_ticket_by_id
- post_ticket
- patch_ticket
- get_ticket_comments
- post_ticket_comment
- get_users
- get_user_by_id
- search_users
- get_orgs
- get_tags
- put_ticket_tags
- + 35 more…
after · 9 task-shaped tools
- find_tickets(query, status?, assignee?)
- get_ticket_context(ticket_id)
- create_ticket(summary, details, customer)
- update_ticket(ticket_id, changes)
- reply_to_customer(ticket_id, message) · approval
- assign_ticket(ticket_id, to)
- escalate_ticket(ticket_id, reason)
- summarize_queue(team)
- find_customer(query)
- Tool-selection accuracy
- 61% 96%
- tools/list size
- 18.4k tokens 3.1k tokens
- Calls per task
- 7.2 2.4
Who we build for
Two kinds of MCP server. One standard of craft.
For SaaS teams
Put your product inside Claude and ChatGPT.
Your customers are asking for a connector. We build the remote MCP server that makes your product usable from the assistants they already pay for.
- Remote MCP server over Streamable HTTP
- Per-user OAuth 2.1 against your existing auth
- MCP Apps UI for flows that need more than text
- Claude connector and ChatGPT app submission prep
- Multi-tenant isolation and rate limits
For enterprise
Give your agents governed access to internal systems.
Your teams are rolling out Claude, ChatGPT Enterprise, or Copilot. We connect them to the ERP, CRM, data warehouse, and internal APIs they need, inside your controls.
- SSO through your identity provider
- RBAC and least-privilege scopes per tool
- Human approval gates on consequential actions
- Audit logs your security team can query
- Deployed in your VPC or on-prem
How we’re different
Built like product, not like a demo.
01
Tool design
Tools shaped around tasks, not endpoints. Clear names, tight schemas, structured errors, and pagination an agent can follow.
02
Evals
Scenario suites that measure tool selection and task success per client. Shipped with every build, rerun in CI on every change.
03
Security
OAuth 2.1 with per-user, least-privilege scopes. Approval gates on consequential actions, audit logs, and a prompt-injection threat model.
04
Ownership
Your repo, your cloud, your keys. Documented and handed over, with no runtime lock-in. Cloudflare, Vercel, AWS, GCP, Azure, or your own Kubernetes.
Packages
Published prices. Fixed scope.
Every engagement is a fixed-price proposal after a 30-minute scoping call.
MCP Audit & Eval
from $4,500
1–2 weeks
We run your existing or generated MCP server through the same evals and security review we use on our own builds.
- Tool-selection eval scorecard across Claude, ChatGPT, and Cursor
- Context-budget analysis of your tools/list
- Auth review: OAuth 2.1 flow, scopes, and token handling
- Prompt-injection and data-exfiltration threat review
- Prioritized fix list with redesigned tool signatures
Custom MCP Build
from $18,000
Typically 3–6 weeks
A production MCP server designed around the tasks your users and their agents actually perform. Secured, evaluated, and handed over.
- Scoping and tool-design workshop
- Remote MCP server (Streamable HTTP) in your repository
- OAuth 2.1 with per-user, least-privilege scopes
- Human approval gates and audit logging for consequential actions
- MCP Apps UI where a widget beats a wall of text
Care & Evolve
from $1,500/mo
Ongoing, monthly
MCP moves fast. We keep your server current with the spec and the clients, and keep the evals green.
- Spec and client-compatibility tracking
- Eval regression runs on every change
- New tools and tool refinements
- Monitoring, alerting, and incident support
How we build
From scoping call to a scorecard you can rerun.
- 1
Scope
A 30-minute call, then a fixed-price proposal with the tool list sketched.
- 2
Design tools
We map the jobs agents need to finish and design the smallest set of tools that covers them.
- 3
Build and secure
Remote MCP server, OAuth 2.1, scopes, approval gates, and audit logging, in your repo.
- 4
Evaluate
Scenario evals across Claude, ChatGPT, and Cursor. We iterate until the scorecard holds.
- 5
Ship and hand over
Deployed to your cloud with a runbook, evals in CI, and a walkthrough with your team.
FAQ
Questions buyers ask us
Do we own the code?
Yes. Everything we build lives in your repository and runs in your cloud account from day one. There's no proprietary runtime and no license fee, and the handover includes documentation, a runbook, and the eval suite.
Which MCP clients do you support?
We build to the MCP specification and evaluate against the clients your users actually have, typically Claude (web, desktop, and Claude Code), ChatGPT, Cursor, VS Code, and your own agents. The scorecard reports results per client.
Where will the server run?
Wherever you already run software: Cloudflare, Vercel, AWS, GCP, Azure, your Kubernetes cluster, or on-prem. Remote servers use Streamable HTTP; local developer tools can use stdio where that fits.
Can you fix the MCP server we already have?
Yes. That's the MCP Audit & Eval package. We measure how your server performs today, show you where agents go wrong, and hand you a prioritized fix list. If you'd rather we make the fixes, we'll scope that too.
Why not generate an MCP server from our OpenAPI spec?
Generation is a fine starting point and a poor finish line. Endpoint-shaped tools spend the model's context and make it guess. We often start from a generated server, redesign it around tasks, and prove the difference with evals.
How long does a build take?
Most custom builds take 3–6 weeks from signed proposal to production. A focused audit takes 1–2 weeks. The scoping call is where we pin down the real number for your project.
Can you work inside our VPC or under NDA?
Yes. We sign NDAs, follow your access and review processes, and can build and deploy entirely inside environments you control.
Will you get us listed in the Claude and ChatGPT directories?
We prepare everything a submission needs (auth, policies, metadata, test accounts, and reviewer notes) and submit on your behalf. Approval is up to Anthropic and OpenAI, so we don't promise it, but we build to their published requirements from day one.
Tell us what your agents need to reach.
An engineer who would build your server reads every message. We reply within two business days with questions or a time for a scoping call.
Start an MCP project