SearchTrophy Engine · for developers
Build a company MCP server, fast. On a framework already proven in production.
Read-only MCP tools that answer as the person asking, plus usage insights. Your developers build the tools for your questions. The engine handles keys, installs, updates, exports and measurement.
What's in it
The parts every build needs, already built. They come from our own production build, so your developers spend their time on tools for your questions.
The server pattern
An MCP endpoint over Streamable HTTP inside your existing backend. It's stateless, so it survives restarts and needs no session affinity. Every tool calls your own API over loopback, as the user who asked.
Personal keys
Each person creates a key in your web app, names it after their computer and copies it once. Keys are stored only as a hash, up to five per person. Revoked keys stop working within a minute. Admins see every key with its last use.
Off switches
Keys stop working when a user is deactivated. One central setting turns all AI access off in seconds, without a restart.
The installer
A setup program for Windows and a package for macOS. The person pastes only their key. It registers the server for Claude Code in the terminal and in VS Code, stores the key encrypted for that user, never shows it to Claude, and pre-approves the read-only tools.
Instructions that update themselves
The laptop holds a skill of about 2.5 KB. The full guide is served by your server and ships in the same build as the tools. At each session start a hook checks the version, downloads and verifies a new one, and swaps it in silently. A test fails the build if any tool is missing from the guide.
Exports for big data
“Give me all of it” becomes a compressed CSV: an owner-only download with the person's own key, integrity-checked, expiring after 24 hours. Claude then analyses it on the laptop with pandas or DuckDB.
Usage insights
Every tool call is stored with the person's question, the tool, its arguments, the outcome, size and time. Questions that take many steps show exactly what to build next.
Test tools
Per-role tests compare your web API, the MCP answers and real Claude Code sessions, using throwaway users for each role. A suite of your staff's real questions runs through Claude Code before and after each round of tools.
How we build tools
Six rules that made the answers good. Learned on real staff questions, measured before and after.
Shape tools like questions
“Which campaigns use the AI receptionist?” first took 93 calls and about 38,000 API requests, then timed out. A purpose-built tool answers it in 3 calls and 12 seconds.
Make “nothing” final
The AI didn't trust empty results and kept checking. Answers now say how many records were checked, so zero rows clearly means none.
Never one call per record
Totals, rankings and the same screen for many records each come back in one call.
Cache outside data, compute once
Third-party data is cached for 30 days and refreshed weekly. Each scan is reduced to a few numbers once, and those matched the vendor's own figures on tens of thousands of historical scans.
Ship instructions with the code
Because the guide comes from the server, a new tool reached an installed laptop in the middle of a test, and the same question dropped from several calls to one.
Measure with real questions
27 real staff questions, before and after one round: 646 to 300 tool calls, 17 to 5 refusals, about 45% lower cost and 20% faster. Later rounds took typical questions from 10–12 calls to 3.
Fit
What your system needs.
- An API that enforces its own access rules on the server, or a read-only database login with a clear visibility rule
- A way to tie a personal key to a user
- A paid Claude plan with Claude Code for each person who asks
- Somewhere to run the endpoint: inside your backend, on AWS or any other cloud
If your permissions live only in the user interface, talk to us first. That needs work before AI access is safe.
Our reference build
The endpoint runs inside a .NET backend. It is standard MCP, deployed and tested with Claude Code in the terminal and in VS Code.
Our own build
First commit to production in a day. The speed comes from reuse. Nothing about access control was re-implemented, so there was nothing new to get wrong.
- Day 1 · 16 tools
Personal keys, the access page and the MCP server
Verified in production with real Claude Code sessions.
- Day 2 · 23 tools
Ownership checks and deeper answers
Per-record ownership checks in the platform, deactivated accounts locked out everywhere, payments, paging for any screen, client overview, statistics and rank history.
- Day 3
Self-updating instructions and usage insights
The server-served guide, a 30-day cache for outside data, exports, the question log, an access audit of every read endpoint, and analytics across campaigns.
- Day 4 · about 37 tools
Language, hours and expansion questions
Answers in the language of the question, hours from the time tracker with its own visibility rule, public Census data for expansion questions, and installers that ask only for the key.
Tested with about 900 unit tests, protocol-level smoke tests of every tool, real Claude Code sessions for six roles, and browser tests for every screen change.
Get started
$200 a month. Your developers, our engine.
Prefer we build it with you? That's $3,000 setup, then $450 a month.