The Comsky Chat connector for Claude

What this connector is

An MCP server we run at mcp.chat.comsky.ai. It sits between Claude and the Comsky Chat API that your own dashboard uses, and it holds no data of its own: no database, no cache of your conversations, no copy of anything. Every request it makes upstream carries your credential, so it can reach exactly what you can reach and nothing else.

That is worth saying plainly, because it is the part that decides how much a compromise of this service would cost. There is no service account behind it and no key that reads every workspace. Take the whole process and you have the traffic passing through it at that moment.

Connecting it

In Claude, add a custom connector pointing at https://mcp.chat.comsky.ai/mcp. Authentication is a bearer token: either a Comsky Chat API token, which you create under Settings, or — once the listing is live — a Connect button that signs you in through account.comsky.ai and asks you to approve what the connector may do.

The endpoint answers JSON-RPC over HTTP. Opening it in a browser returns a 401 with a pointer to the metadata a client needs, which is correct behaviour and not a fault.

What it can read

• Conversations — list them across WhatsApp, email and the web widget, open one, and read its messages.
• Contacts — search them, open one, and read the history of everything that contact has ever sent or been sent.
• Tickets — list them, open one, and read its comments.
• Search — one query across messages, contacts and tickets at once.
• The shape of the workspace — inboxes, labels, teams, members and the numbers you may use, so Claude can name things the way your team does rather than guessing at an id.
• Reports — the overview, a time series, SLA performance, per-agent figures, and a breakdown by channel, label or team.

Phone numbers follow the same masking rule as the dashboard. If your role sees a redacted number on screen, the connector returns it redacted too.

What it can change

Nine tools write, and an administrator can switch all of them off for the whole deployment. They are: assign a conversation, set its status, set its priority, add a label, create a ticket, update a ticket, comment on a ticket, update a contact, and reply to a conversation.

Replying is metered separately and far more tightly than the rest. Sending to a customer is the one action here that cannot be undone and that a mistake makes expensive — a WhatsApp number that sends like a robot gets banned, and a banned number does not come back. So replies have their own hourly ceiling, well below the write ceiling, and they go out on the paced lane the product uses for automated sending rather than the fast lane a person typing gets.

What it cannot do

• Nothing is ever deleted. There is no delete tool, and the code that talks upstream cannot issue one — a build check fails the release if a delete appears anywhere on that path.
• It cannot write to somebody who has not written in. The first-contact endpoint is not on its list, so a cold message cannot originate here.
• It cannot change who can see what. No tool touches permissions, roles, teams, inboxes or number access.
• It cannot reach another workspace. The workspace is fixed from your credential before any tool runs, and a tool that tried to name its own would fail outright rather than be obeyed.

It acts as you, never above you

Every answer passes the same fences your own session does: which numbers and inboxes you may read, whether a conversation is assigned away from you, whether your role may see a ticket at all. The connector adds no access. If you cannot see something in the dashboard, asking Claude about it returns nothing.

Beyond that, every result is checked before it is handed over: a record that does not belong to your workspace fails the whole request rather than being filtered out quietly, and anything that looks like a credential is stripped from the response regardless of which tool produced it.

What happens to your data

What you ask Claude, and what the connector returns, go to Anthropic and are handled under their terms. We do not control that and do not claim to. What we can tell you is what happens on our side: the connector stores nothing. It keeps no transcript, no copy of a conversation and no record of what was asked. Its logs record which tool ran and whether it succeeded — the names of the arguments, never their values, because the values are your customers' words.

A write is recorded in the workspace's own audit trail, the same as if you had done it in the dashboard, with the fact that it came through the connector attached.

How Comhard Technologies handles personal information generally — what is collected, who processes it, how long it is kept and your rights under the DPDP Act — is in the Privacy Policy.

Rate limits

Per credential: sixty requests a minute, four at a time, and a budget on how many upstream calls a single tool may make. Per workspace, a higher shared ceiling. Writes are capped separately at ten a minute, and replies lower still.

These exist to protect your database as much as ours. A model exploring a large inbox can ask a great many questions very quickly, and an unmetered connector turns that curiosity into load somebody has to pay for.

Support

Write to [email protected]. For anything that looks like a security problem, [email protected] reaches the same people faster.