Imagine a perfectly ordinary scenario: a customer asks in a chat box, “Where are the shoes I ordered last week?”
These questions always get stuck in the same place. If a bot has only read the help docs, it knows how returns are calculated and who pays shipping, but it does not know which sorting center this customer’s package is in right now; at that point it can only ask for an order number and hand the case to a human, or give an answer that sounds knowledgeable but was never actually checked.
Chatwith wants the bot to look in the back office first, then speak. The official homepage says it connects a website, help docs, files, YouTube videos, and the apps the team already uses to answer customers; the “training” list in the FAQ is narrower—websites, PDFs, help content, and plain text. Apps are not on that list; they take another route: connect via OAuth or an API key, and let the bot do things, not just know things. [1]
The difference is not whether the answer is accurate, but whether the bot is allowed to touch your business systems.
Image: source official image cited by the Chatwith website[2], retrieved 2026-11-11, original interface preserved.
Read an action, write an action
Chatwith’s help docs draw this line clearly: for things like checking order status, use read actions; only when the agent really needs to create, update, send, cancel, or refund should write actions be enabled, and the provider’s permissions should be limited to the necessary accounts and resources. [3]
The same doc has a very plain sentence: actions may change external data, and testing destructive or customer-facing operations requires caution. That warning does not show up on the pricing table.
Then there is the bill. Only calls that are successfully connected through MCP and actually executed by an external app count toward the monthly MCP action quota; built-in actions, custom REST API actions, and webhooks do not. The pricing page breaks out this quota as an add-on: an extra 1,000 MCP action calls for $6/month. [4]
So “check an order once” and “reply with one sentence” really do fall under two different line items on this bill. The $9 Starter plan includes 250 message credits a month, which the pricing page illustrates as “about 50 conversations,” on the premise that each conversation has 5 AI replies and each reply consumes 1 credit, with a note that actual usage varies with the model and conversation length. Ibid.[4] It is in effect a reminder: do not treat message count as conversation count, and even more, do not treat message count as action count.
On the same pricing page, the version retrieved on 2026-11-14 starts at $9 a month for Starter; annual payment is a single charge and excludes applicable taxes; the Hobby, Standard, and Business plan cards mark a seven-day free trial, while the Starter plan card does not.
Permissions Are Granted One by One, and Someone Has to Pick Up When Things Go Wrong
Before letting an agent “act,” there is an unglamorous piece of work.
Connecting an app requires OAuth or an API key; connecting it does not mean opening all of that provider’s operations—you have to choose the specific ones from Chatwith’s curated action catalog. Skills marketplace and action catalog - Help Center - Chatwith[3] In other words, permissions are granted one by one. After you hand over the refund button, the docs do not assign who should handle failures and reauthorization; this is the operational burden that granular permissions bring, and it is an operational judgment that can be drawn from the docs, not a process the company has disclosed.
If you want to connect your own interface, there is a harder option. REST API Actions connect to external APIs through an OpenAPI specification, letting the agent retrieve information or perform actions allowed by the specification in response to user requests; the official description calls this an advanced approach that usually requires API access and technical ability, and it currently supports only header-based authentication. The line in the docs—“it is impossible to make it plug-and-play”—is quite blunt. [5]
This limit draws the boundary of “how much work this product can save you”: what it saves is the few actions that can already be connected, not every back office.
The Ad Bill That Came Close to Going Unpaid
To understand how this company operates, go back to a stretch in 2024.
The author of that retrospective wrote that early in the year a payment of about $21k was about to arrive from Chatwith’s AppSumo launch, and that his personal savings runway would last until the end of 2024. He decided to put the AppSumo money aggressively into Chatwith’s Google Ads, with a very practical set of goals—learn how to run ads, validate whether it could be profitable, and run A/B tests on the landing page. [6]
Over the next few months, spending was higher than inflow; by November, the account at one point hit $0, and one Google Ads bill could not be paid. He wrote himself that for a long time he was not sure whether any of this meant anything, or whether he was simply burning money.
This is not “just hang in there and it will be fine” inspirational material. It is more like someone who builds products admitting that he used cash flow as tuition. Later he wrote that MRR was rising slowly, customer acquisition cost was falling, trial conversion was improving, churn was falling, and over the long term every $1 of ad spend brought in $3 of revenue; Chatwith reached his self-set year-end threshold of $5k MRR (Black Friday helped), so he decided to keep investing. [6]
In the same piece, what he listed for 2025 explains today’s pricing page better than “keep spending on ads”: he said he wanted to give Chatwith a new pricing scheme; the current usage-based model was headed in the right direction but not flexible enough, and the new pricing needed to manage costs more finely and open up his target customers—agencies. Ibid.[6] Whether the current price table is the “new pricing” he described is not addressed in the material; what can be seen is that messages and actions are metered separately, and the direction really is toward “more granular,” not toward “bundled.”
He also acknowledged in passing that Chatwith’s organic search traffic was “almost nonexistent,” even though the site’s DR is 45, and so he classified SEO as an underperforming channel. Someone who builds tools writing this way about his own distribution is closer to the facts than any growth methodology.
42 Subscriptions, and a Profit That Is Hard to See
Move the clock to 2026-11-14 04:43: on the TrustMRR page, Chatwith’s MRR is $3,622, with 42 active subscriptions and $116,936 in lifetime revenue; the revenue is verified through a Paddle API key, and the rank is #588. [7]
These numbers and the $5k threshold in that 2024 retrospective differ in time, source, and verification method, and the material does not say what happened in between. The only thing that can be established is: MRR is not profit. The page does not disclose ad spend, churn, or model and hosting costs, so whether this company is making money is still unclear. Apart from 42 active subscriptions, there is also no customer size, industry distribution, or plan distribution, so customer count cannot be inferred from MRR.
Put It Next to DocsBot, and Don’t Look at the Sticker Price First
Among similar products, DocsBot’s pricing page lists Personal at $49 a month, including 1 DocsBot, 5,000 monthly AI credits, 3 actions per bot, and remote MCP connections; a footnote says AI credits start at 1x per message, and larger models and AI actions may consume more. Its credit add-on is described as covering chat, skills, and API usage, but the same page does not make clear how MCP connections are counted. [8]
So to compare “how much one order lookup costs,” you cannot judge who is cheaper by list price alone. Chatwith gives successful MCP calls a separate quota, purchasable at $6/1,000 calls; DocsBot puts chat, skills, and API usage into one credit pool, but each bot still has an action cap, and how MCP connections are billed is not made clear on the same page.
The Judgment We Stop At
What Chatwith sells is not “smarter answers”; it is handing the few clicks that previously only a human could make to an agent that needs authorization and a quota to click them.
Read and write actions have to be selected separately, successful MCP calls are metered under a separate quota, and REST API and webhook actions are not counted; taken together, these are what it is actually selling.
[1] AI Customer Support Chatbot for Your Website | Chatwith
[2] official image cited by the Chatwith website
[3] Skills marketplace and action catalog - Help Center - Chatwith
[4] Chatwith Pricing: AI Agent Plans from $9/Month
[5] REST API Actions - Help Center - Chatwith