HomeFeaturesAPI and agents

API and agents

Connect your software. Or your AI.

Two different readers, deliberately kept apart: your own software, which does exactly what you programmed, and an AI agent, which does something reasonable that you did not specify. They get separate doors, separate keys, and different powers.

Two doors
API keys and MCP keys never cross
Role-scoped
every key, intersected with its maker
Bring your own
model and API key

What you get

A REST API for your own systems, an MCP server so an AI agent can work your inbox, and scoped keys that can never do more than the person who made them.

Scoped API keys

Every key carries a role. Minting one and being allowed to send are two separate permissions, deliberately not one.

Send from your own code

Post a message or a file, create and update contacts, import a list. Same rules as the panel, same delivery receipts.

An MCP server for agents

Point Claude or your own agent at Birsend and it can read conversations, message people and pause a campaign, within its role.

Your own model writes replies

Connect a model you pay for and let an auto-reply rule ask it what to say. Your key, your provider, your bill.

An agent narrows, never widens

The MCP tools let an agent pause a campaign but never resume one. Undoing a deliberate human stop is not a thing to hand to a sentence.

A run log for every rule

Each automatic reply records what it did, or why it did not — cooling down, no model connected, refused by the filter.

How it works

01

A key can only ever be narrower than you

A key carries the role its maker chose for it, and what it may do is the overlap between that role and the maker’s own permissions. Two things follow, and both matter: nobody can mint a key more capable than themselves, and a key stops growing when the person behind it is demoted.

  • An API key and an MCP key are different abilities — one leak stays one door
  • Holding a key and being able to send through it are two permissions
  • A key is shown once. Only its hash is stored, so a lost key is replaced rather than recovered
02

What an agent may and may not do

The MCP tools are chosen so an agent can act usefully and cannot escalate. It can pause a campaign that is going out; it cannot resume one. Widening is a human decision, because resuming puts thousands of messages back in motion off a single sentence.

  • Read conversations, send a reply, tag a contact, pause a campaign
  • Every send still passes the sending window, the daily cap, consent and the word filter
  • The public API has both pause and resume — a person minted that key on purpose
03

Your model, not ours

Birsend does not resell AI. You connect a provider and a model with your own key, and an automatic reply rule can ask it to write the answer, with per-rule instructions so one connection can handle a price question and a complaint differently.

  • The key is stored write-only — it is never sent back to the panel
  • A test call is one button, because a form that saved successfully proves nothing
  • If the model cannot be reached, nothing is sent. A canned line going out under its name would be worse than a visible failure

What it does not do

The honest list. We would rather you found this here than three weeks into a trial.

  • An AI reply sends on its own once the rule is on. It is not a suggestion queue — there is no drafting, rewriting or approval step in the composer.
  • There is no AI anywhere else in the product: no draft-from-a-prompt, no tone rewriting, no translation, no reply suggestions in the inbox, no send-time recommendations.
  • Model connections are configured through the API rather than a panel screen.
  • There are no outgoing webhooks. Your systems poll the API; the WhatsApp service is the only thing that posts to us.

Questions

Which model do I need?

Any provider with an OpenAI-style chat endpoint over HTTPS. You choose the model and hold the key, so the running cost is between you and them.

Is the MCP server safe to give an agent?

It is scoped by a role you pick, it cannot exceed your own permissions, and every message it sends passes the same guardrails as one you type. Pausing is available to it; resuming is not.

Cookies

The site runs what it needs to work, including the Crisp chat bubble. Google Analytics is the one thing we will not load unless you agree to it.