# For GTM engineers who build the sales agents

**You built the agent, and Reynforce is what it checks with before it promises a price, a date or a discount. It also keeps a record of every promise, and you connect it once without changing how your agent runs.**

GTM means go-to-market: the people and tools that sell. You already know
what goes wrong. Your agent is good at talking to buyers. But it can't tell
a number someone checked last week from a number that sat in an old slide
deck eighteen months ago. The prompt can't tell it either, because the
prompt is often where the old number came from.

## What your agent gets

Your agent talks to Reynforce over MCP. MCP is a standard way for an AI
agent to call outside tools. Reynforce gives it two tools.

| tool | what it does |
|---|---|
| `search_claims` | reads your company's facts. Each one has an owner and the date it was last checked |
| `check_action` | asks *may I say this?* The answer comes from the rules your company set, and the question is recorded |

**Your agent reads facts and asks questions. It can never change your
facts.** It sends a short note of what it is about to do. It gets back one
of three answers: *allowed*, *ask first* or *not allowed*. The answer also
lists the facts it relied on and the state of each one. We write the record
on our side, from the check.

**Every question leaves a trace.** If an agent could ask *would this be
allowed?* with nothing written down, it could keep asking until it got a
yes. If your agent is only writing a draft, it says so. The check is still
recorded, but it doesn't count as a message sent.

## What happens when a fact can't be used

Some facts can't be used. Nobody owns them, or they are past their check
date, or someone has disputed them. When your agent asks about one, it
hears that the fact can't be used. It then tells the buyer it is checking,
and we ask the owner. It doesn't make up an answer, and no setting lets it.

## How to connect

Add the MCP server and your agent's key. It takes a few steps. When your
agent needs a person to approve something, the request reaches your team by
email, Slack or text, so nobody has to sit and watch a dashboard.
[Connect an agent](/docs.html) walks you through it.

## When Reynforce is the wrong choice

- **You need something to run your agent.** Reynforce never runs your
  agent. Your own framework does that. We only answer its question.
- **You need to filter wording or stop prompt injection.** Prompt injection
  is when someone hides instructions in text to trick the agent. Use a
  guardrail library for that. We don't filter text.
- **Your agent can't call a tool before it sends a message.** Then it can't
  ask us first. Tell us what it runs on and we'll tell you plainly whether
  it fits.

## What it costs

**£99 per agent a month** for your first 10 agents, and **£79** for each one
after that, up to 49. Pay yearly and you get two months free. From 50 agents
we agree a price on a call. We don't charge per check, so your agent never
has a reason to ask less. [Pricing](/pricing.html) has the detail.

[What we keep, and where](/privacy.html) explains all the data we hold.