ForthClover
Agent-Ready Integration · For SaaS product teams

Let AI agents use your product, safely.

Your customers are starting their work in ChatGPT, Claude and their own agents. In two to three weeks we build a production MCP server for your product: a short list of well-designed tools, secure sign-in, rate limits, audit logs and evals. Those assistants can then act for your customers, inside the permissions your product already enforces.

See tools for your productHow it plugs in
Price
From $6,500
Fixed. No day rates.
Timeline
2–3 weeks
Set by tool count and auth
Scope
~10 tools
Read-first; writes quoted
Ownership
Your code
In your repository
Pick your product type

What an agent could do with your product.

Typical tool sets for five kinds of SaaS products, with the access each tool needs and what a real request turns into. Yours will be shaped by what your customers actually ask for.

CRM & sales tools

Reps ask their AI assistant about deals and accounts, and log what happened, without opening your app.

Typical tool set

Example
ToolWhat it doesAccess
search_contactsFind people and companies by name, domain or ownerread
get_deal_summaryStage, value, contacts and next step for one dealread
find_stalled_dealsDeals with no activity for a given number of daysread
get_activity_timelineRecent calls, emails and notes on a deal or contactread
log_noteAdd a note to a deal or contactwrite · confirm
update_deal_stageMove a deal to another stagewrite · confirm

Agent conversation → tool calls

Illustrative

  1. Which of my deals haven't moved in two weeks, and what did we last discuss?

  2. find_stalled_deals{ owner: "me", idle_days: 14 }read

    → 3 deals

  3. get_activity_timeline{ deal_id: "D-1042", limit: 3 }read

    → last call 16 days ago, pricing question open

  4. Three deals are idle. Acme is the biggest: the last call ended on an open pricing question.

  5. Log a note on Acme that I'll call them Thursday.

  6. Add this note to Acme Corp: “Call planned Thursday.” The user confirms before anything is written.

  7. log_note{ deal_id: "D-1042", text: "…" }write · confirm

    → note saved, logged to the audit trail

Scopes the user approves

contacts:readdeals:readnotes:write

Design note

One find_stalled_deals tool instead of making the model page through every deal and compare dates itself.

Different kind of product? If it has an API and customer accounts, it can usually be made agent-ready. Write tools are a quoted extra and always ask the user to confirm.

What you get

A server your customers can connect to, and the proof it behaves.

Everything ships into your repository with tests, so your team can read it, run it and change it.

find_stalled_dealsread

get_deal_summaryread

log_notewrite

End of week 1

Tool spec

The jobs your customers want done from an AI assistant, turned into a short list of tools with names, inputs, descriptions and access levels. You sign it off before we build.

POST /mcp · tools/list

200 OK · 9 tools

rate_limit: per user

audit_log: on

tracing: otel

Weeks 1–2

Production MCP server

A remote MCP server in your repository, calling your existing API. Rate limits per user, structured audit logs and usage tracing with OpenTelemetry.

An AI assistant wants access

deals:readnotes:write
AllowCancel

Week 2

Secure sign-in

OAuth-based authorization as the MCP spec describes it. Your customer signs in to your product and approves narrow scopes; tokens are checked on every call.

Tool selection

Arguments

Refusals

Week 3

Eval report

A test set of real customer requests, scored for picking the right tool, passing the right arguments and refusing what it shouldn't do. It runs in CI on every change.

Setup guide

Tool reference

Listing copy

Reviewer test account

Week 3

Docs and listing kit

A setup guide and tool reference for your customers. If you add a directory listing, we also prepare the listing copy, screenshots and reviewer test account.

How it plugs in

A thin, secure layer in front of the API you already have.

Your API and data stay where they are. The MCP server translates agent requests into calls your product already understands, under the signed-in customer's permissions.

  1. 01

    Agent clients

    Where your customers already work

    ChatGPTClaudeCursorVS Code + CopilotCustomers' own agents

  2. 02

    Your MCP server

    What we build

    Task-shaped toolsOAuth sign-inPer-user rate limitsAudit logUsage tracing (OpenTelemetry)

  3. 03

    Your API

    Same rules your app enforces

    Existing endpointsYour permission modelTenant isolation

  4. 04

    Your data

    Stays where it is

    Your databaseCustomer workspacesFiles

Across every layer

Security review

Prompt injection, scope creep and data leakage tested before launch.

Across every layer

Evals in CI

Tool selection and correctness re-scored on every change.

Tool design

Good tools matter more than many tools.

Agents struggle with tools that are too many, vaguely described or too powerful. Four rules we design by.

Fewer tools, shaped like tasks

Agents choose better from a short list. We build around the jobs customers do, not one tool per API endpoint.

find_stalled_deals, not list_deals + list_activities

Descriptions written for the model

Each tool says what it does, when to use it, and what it returns. Errors tell the agent how to recover.

“Use when the user asks about idle or stuck deals.”

Least privilege

Narrow OAuth scopes, the user's own permissions, and read-only by default. Nothing an agent can reach that the user couldn't.

deals:read granted; deals:delete never exposed

Confirm before writing

Write tools ask the user to confirm, and money-moving or customer-facing actions end as drafts in your product.

create_draft_order, not place_order

The three weeks

Spec first, then build, then prove it.

Shown on business days for a standard ten-tool scope. Smaller tool sets with an existing OAuth setup usually finish in about two weeks.
  1. D1–3Jobs & tool specCustomer jobs, API review, tool list, auth plan
  2. D4–9Build the serverTools, OAuth sign-in, rate limits, audit log
  3. D7–11Evals & tracingTest set from real requests, scoring, OpenTelemetry
  4. D9–13Test in real clientsChatGPT, Claude and the clients your customers use
  5. D13–15Security pass & handoverInjection and scope checks, docs, walkthrough
Scope

What the price covers, and what we quote on top.

From $6,500 covers the standard scope. Extras are priced on the fit call, before you sign anything.

Standard scope

  • Up to ~10 well-designed tools, read-first
  • Secure sign-in with OAuth-based MCP authorization
  • Per-user rate limits and audit logging
  • Evals for tool selection and correctness, running in CI
  • Usage tracing so you can see which tools get used
  • Setup guide and tool reference for your customers

Quoted extras

  • Directory listings. Submission to directories where available, such as ChatGPT's app directory and Claude's Connectors Directory. Each runs its own review.
  • Per-customer permissions. Tenant- or role-level control over which tools each customer's users can reach.
  • Write actions with confirmations. Create, update and draft tools, each with a confirmation step.
  • Hosting in your cloud. Setting up and hardening the server's hosting in your own cloud account.
Who does what

What we need from you, and what we take off your plate.

You provide

  • A product owner who can decide tool scope, about two hours a week
  • API access in a staging environment, with test accounts
  • Your identity provider or OAuth details, or a decision on which to use
  • An engineer to review code and approve the deploy

We provide

  • A senior engineer from spec to handover
  • Tool design, the MCP server and OAuth sign-in
  • Rate limits, audit logging and usage tracing
  • Evals, client testing, docs and a handover walkthrough

Not included

  • Building new API endpoints beyond small additions
  • More than ~10 tools, writes, listings or per-tenant permissions (quoted)
  • Hosting and API costs after launch
  • Ongoing maintenance (see AI Operations)
Is this you?

Built for SaaS teams whose customers already use AI assistants.

A good fit if…

  • You run a SaaS product with an API and customer accounts
  • Customers ask to use your product from ChatGPT, Claude or their own agents
  • You can name the handful of jobs customers do most in your product
  • You want it built properly once, with sign-in and evals, not as a weekend demo

Not the right start if…

  • There's no API yet, internal or public
  • You want an agent inside your own product (that's an Agent Build)
  • You need hundreds of endpoints mirrored one-to-one as tools
  • Your data can't be shown to third-party AI clients under your current terms
Where it sits

Make it agent-ready, keep it safe, keep it current.

Once agents can reach your product, test how it holds up under attack, build your own agent on the same tools, and keep it all running. How we build every system
  1. Step 1 · you are hereAgent-Ready IntegrationFrom $6,500fixed · 2–3 weeks
  2. Step 2Agent Security Review$3,500fixed · 1 week · one agentDetails
  3. Step 3Agent Build · Proof stage$9,500fixed · 3–4 weeksDetails
  4. Step 4AI OperationsFrom $990/mothree tiers · 3-month minimumDetails
FAQ

Questions CTOs and product leads ask us

What is MCP?
The Model Context Protocol is an open standard for connecting AI applications to tools and data. Your product exposes a set of tools through an MCP server; an AI client such as ChatGPT or Claude discovers them and calls them for the signed-in user. It started at Anthropic and is now an open project under the Linux Foundation's Agentic AI Foundation.
Which AI clients support it?
Claude and ChatGPT both connect to remote MCP servers, as do developer tools such as Cursor and Visual Studio Code with GitHub Copilot, and a growing list of other agents. Support for specific features (sign-in, interactive UI, write actions) varies by client and plan, so we test in the clients your customers actually use.
How is sign-in handled?
Through the MCP authorization spec, which is built on OAuth 2.1. Your customer signs in to your product and approves the scopes the AI client asks for. The client gets a token issued for your MCP server only, and the server checks it on every call. Nobody pastes passwords or API keys into a chat.
Can it write data, not just read it?
Yes, if you want it to. The standard scope is built around read tools. Write actions, such as creating a task or logging a note, are a quoted extra. Each one asks the user to confirm, is limited by scope, and lands in the audit log.
What does it cost to run?
The AI model runs on your customer's side, inside their ChatGPT, Claude or other plan, so you don't pay for their model usage. Your running costs are hosting a small service, its logs and traces, and the extra traffic on your API. We size these with you before launch.
Who maintains it when the spec changes?
You own the code, and we hand it over with tests, evals and docs your team can work with. The MCP spec is versioned by date and revised from time to time, and AI clients change how they use it. If you'd rather not track that yourself, AI Operations keeps the server current and re-runs the evals after every change.
From $6,500 · fixed · 2–3 weeks

Your customers' assistants are already asking for your product.

Bring your API docs and the jobs customers ask about most. On a free fit call we'll sketch the tool list and give you a fixed quote.

Book the fit call