Claude API (Anthropic) ↔ Gmail integration

An AI agent on Claude reads the whole Gmail thread with its PDF attachments, drafts the next reply in context and leaves it for a person to send

Some letters cannot be answered without reading everything before them: a deal discussed over thirty messages, a contract sent back with edits, a specification in a PDF. The manager spends half an hour restoring the context before writing three lines. The Claude ↔ Gmail pair gives that work to an AI agent: it takes the whole thread in the order it was written, reads the PDF attachments including tables, and prepares the next reply as a draft in the same thread, with a short note on what changed since the last letter. A person reads it, corrects it and sends it. Below: what moves, which scenarios we cover and what we rely on in both systems.

Request an integration System page: Claude API (Anthropic) →System page: Gmail →

What moves between the systems

Gmail → Claude API (Anthropic)When a new letter arrives in a thread marked for the agent, the whole thread goes to Claude in order: every reply, quote and signature, so nothing is read out of context.
Gmail → Claude API (Anthropic)PDF attachments in the thread, such as contracts, specifications and price lists with tables, are passed along with the text.
Claude API (Anthropic) → GmailClaude prepares the next reply in the language of the correspondence and it is saved as a draft in the same Gmail thread.
Claude API (Anthropic) → GmailNext to the draft the manager gets a short note: what the other side asked, what changed in the attached document compared with the previous version and what is still open.

What we connect

We agree how a thread is handed to the agent, usually by a Gmail label the manager sets, so the agent reads only the threads it was given.
The Gmail connection asks for the narrowest set of permissions: read the labelled threads and create drafts, with no right to send.
The tone, signature and standard terms (prices, delivery, payment) are written into the agent's instructions together with your manager, from their real letters.
The Claude usage is billed to your own Anthropic account, and we test the drafts on ten closed deals before the agent sees live ones.

Automation scenarios

A contract comes back with edits

Before the lawyer or the manager compares two PDF versions line by line to see what the client changed.

After the note next to the draft lists the changed clauses, and the draft reply already answers them.

A long deal changes hands

Before a manager on holiday leaves a thread of forty letters, and a colleague reads all of them before answering.

After the colleague gets a summary of what was agreed and a draft of the next letter.

A foreign client writes in their language

Before letters in Polish or German go through a translator and wait.

After the draft is written in the client's language and the note next to it is in yours, so the manager knows what they are sending.

How it works technically

For those who want to check: each point links to the official documentation page it comes from.

Replies and the original letter are grouped into one conversation thread, so the whole exchange on a deal can be read in order.
Source: developers.google.com ↗ checked 2026-09-27
A draft in Gmail is an unsent letter with the system label DRAFT; when it is sent, the draft disappears and a letter with the label SENT appears — this is how the AI draft waits for the manager.
Source: developers.google.com ↗ checked 2026-09-27
Google recommends requesting the narrowest set of permissions possible and not asking for access the application does not need.
Source: developers.google.com ↗ checked 2026-09-27
The model's working memory for one request, the context window, holds the conversation history plus the new answer; its size depends on the model and reaches up to a million units of text.
Source: platform.claude.com ↗ checked 2026-09-27
Claude answers questions about any text, pictures, charts and tables in the PDF documents you provide; all active models support PDF processing.
Source: platform.claude.com ↗ checked 2026-09-27
Claude reads and writes in most world languages that use standard Unicode characters; performance varies by language and is strongest in widely spoken ones.
Source: platform.claude.com ↗ checked 2026-09-27

Limits

The limits come from both systems. The model can be wrong: it may miss a changed figure or misread a clause, so the draft and the note are proposals that a person checks before anything is sent. A very long thread with many attachments has to fit into the model's working memory for one request; for the longest deals the agent reads the latest letters and attachments in full and the early part as a summary, and says so in the note. Claude reads most languages, but quality is highest in widely spoken ones, so a draft in a rare language gets a closer look.

How to start

Three steps from a first message to work on your project.

01
Leave a request
Describe the task in the form — in text or by voice. Budget and timing are optional.
02
We review and estimate
A developer studies the brief, asks about anything unclear and sends an estimate.
03
We start
Once scope and price are agreed, the work begins.

Leave a request

or write to info@coreviaflow.space

Live project We built and run Gemalli — a B2B export platform with producer pages, a CRM and a deal chat. Read the case →

Questions

Does the agent read the whole mailbox?

No. Only threads a person has marked with the agreed label, and the Gmail connection is set up with the narrowest permissions it needs.

Will our letters be used to train Claude?

By default, no: Anthropic states that inputs and outputs of its commercial products are not used to train its models.

What if the manager edits the draft?

That is the normal case. The manager edits the draft in Gmail and sends it; the sent letter becomes part of the thread, and the agent reads it next time as the latest word.

Tell us what needs to be connected

Describe the task in text or a voice message. We look at your setup and reply with a plan.

Request an integration

What you can order

Four services people come to us for most often. Each page explains what we do, how the work goes and where to start.

CRM implementation

Choose the system, set up the pipeline, migrate data, train the team.

Learn more →

CRM integration

Accounting, telephony, website, payments and messengers feed the CRM on their own.

Learn more →

Telegram chatbots for business

Orders, requests and answers around the clock, straight into your CRM.

Learn more →

n8n and Make automation

Flows between CRM, email, spreadsheets, payments and AI, on your own server.

Learn more →

What you can order from CoreviaFlow →

Each system on its own

AI modelsUS

Claude API (Anthropic)

Claude inside your CRM, phone system and chats: it reads whole documents and long client histories, drafts replies and proposals, fills in cards, and a person has the final say.

See scenarios →

Automation & workspaceUS