- ai
- mcp servers & agent-ready apis
your systems, open to your ai.
We build MCP servers that let Claude, ChatGPT, Codex and your own agents read, write and act in the software you run. Every assistant signs in as one of your people, works within that person's permissions, and asks before anything that sends, deletes or costs money.
for teams that already ask ai for help
Your staff already paste work into ChatGPT or Claude. An MCP server lets the assistant open the real record itself, signed in as them.
- Law firms on Clio
- Trades and field-service businesses with job and estimating tools
- Software companies whose customers expect an MCP server
- Any business with a custom app or an API
from first task to daily use
-
01
map the work
We sit with your team and list what they'd hand an assistant. Each task gets a rating for time saved and risk, and a point where we'd stop.
-
02
reads first
The first tools look things up: matters, jobs, invoices, documents. Your team builds trust before anything can change.
-
03
writes that ask
Every write tool gets a risk level. Routine ones run. Anything that sends, deletes or costs money returns a summary and waits for a yes.
-
04
roll out
Each person connects their own assistant with one sign-in, and we write a setup guide for every assistant your team uses.
-
05
keep it current
New assistants, new spec versions and new business rules arrive over time. We keep the server working through all of them.
a server your team can trust with real work
- The MCP server, deployed and monitored
- A tool catalogue with a risk level and a plain description for every tool
- Sign-in for every person: OAuth 2.1 with dynamic client registration, Client ID Metadata Documents and PKCE, or scoped tokens that expire
- Confirmations on risky actions, with idempotency so a retried request never runs twice
- An audit log of every call, with sensitive fields redacted
- Automated tests covering the tools and the full sign-in flow
- If you have a REST API, a check that keeps every API action and its AI tool in step
- Agent-ready docs: llms.txt, OpenAPI, and setup guides for Claude, ChatGPT, Codex and Cursor
- A sandbox or test mode wherever your system supports one
built to say no
The safety rules live in the server, where the assistant can't talk its way around them.
- credentials stay on the server The AI client gets its own token and never sees your system's keys.
- one real person per assistant Every assistant acts as one of your people, with that person's role checked on every call.
- admins decide who connects Choose who may connect, and revoke any connection.
- risky actions wait for a yes Anything that sends, deletes or costs money comes back for approval.
- text is data, not instructions Document and record text never steers the assistant, which guards against prompt injection.
- keys never sit in an agent Files move through one-time upload links.
- nothing runs twice Retries are idempotent and edits check revisions, so nothing is overwritten silently.
- every call is logged Each user is rate-limited, and every call lands in the audit log.
running today
-
01
our product · legal
attorneys run clio by asking
Attorneys at a personal-injury firm ask Claude, ChatGPT or Codex about a matter and get the answer straight from Clio. In the same conversation they log time, add notes and open new matters, so work that meant clicking through screens now takes a sentence.
Read the case study -
02
our product · e-signature
e-signatures, sent by asking
Every eSignStack plan connects to Claude, ChatGPT and Codex. Your team drafts, sends, reminds and voids envelopes by asking, and the assistant checks with them before anything goes out.
Visit (opens in a new tab) -
03 trades · ai assistant
a landscaping business run by conversation
The owner of a landscaping and turf company builds estimates and runs jobs by talking to an AI assistant, and every estimate and job lands in the app with the numbers right.
In production -
04
our product · billing
billing your ai can run
In eBillStack, an AI assistant can do anything the app does, from logging time to sending invoices, with the same permissions as the person using it, and the books always reconcile.
Visit (opens in a new tab) -
05 finance · ai assistant
research by chat, changes by agent
A private trading client researches strategies from ChatGPT or Claude and asks for changes the same way. Each request becomes an issue that coding agents pick up and build.
In production
mcp, in plain words
What is an MCP server?
Which assistants can connect?
Can we start read-only?
Who looks after it once it's live?
plan your first server.
Bring one system and one task. We'll map the first tools with you.